Studies in Computational Intelligence 758

László T. Kóczy
Jesús Medina Editors

Interactions
Between
Computational
Intelligence and
Mathematics
Studies in Computational Intelligence

Volume 758

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
The series “Studies in Computational Intelligence” (SCI) publishes new develop-
ments and advances in the various areas of computational intelligence—quickly and
with a high quality. The intent is to cover the theory, applications, and design
methods of computational intelligence, as embedded in the fields of engineering,
computer science, physics and life sciences, as well as the methodologies behind
them. The series contains monographs, lecture notes and edited volumes in
computational intelligence spanning the areas of neural networks, connectionist
systems, genetic algorithms, evolutionary computation, artificial intelligence,
cellular automata, self-organizing systems, soft computing, fuzzy systems, and
hybrid intelligent systems. Of particular value to both the contributors and the
readership are the short publication timeframe and the world-wide distribution,
which enable both wide and rapid dissemination of research output.

More information about this series at http://www.springer.com/series/7092

László T. Kóczy Jesús Medina
•

Editors

Interactions Between
Computational Intelligence
and Mathematics

123
Editors
László T. Kóczy Jesús Medina
Faculty of Engineering Sciences Departamento de Matemáticas,
Széchenyi István University Facultad de Ciencias
Gyõr Universidad de Cádiz
Hungary Puerto Real, Cádiz
Spain
and

Budapest University of Technology

and Economics
Budapest
Hungary

ISSN 1860-949X ISSN 1860-9503 (electronic)

Studies in Computational Intelligence
ISBN 978-3-319-74680-7 ISBN 978-3-319-74681-4 (eBook)
https://doi.org/10.1007/978-3-319-74681-4
Library of Congress Control Number: 2018930367

© Springer International Publishing AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

Computational intelligence and traditional mathematics are two interwoven and

extremely important scientific areas nowadays. For example, the treatment of big
amount of data is an eminent goal in numerous topics which have been taken into
consideration in important funding programs, such as the current European research
and innovation program, Horizon 2020, which includes this problem in many of its
subtopics. Hence, the interactions between computational intelligence and mathe-
matics are very important in order to tackle a variety of important challenges. This
volume is a new step in this direction including interesting contributions of engi-
neers and mathematicians focused on the resolution of significant and actual
computer science problems.
The first paper deals with a social network-related problem: the investigation
of the two most up-to-date ranking algorithms (PageRank and Katz) used for
identifying the most influential participants in the discussion on a particular topic.
The research is based on a case study, namely, the “London Riots”, using the
Twitter database. In this study, most influential participating persons in the dis-
cussion are considered as trendsetters, i.e., having the ability to change the opinion
of many others. The frequency of mentioning a participant’s name and forwarding
his/her tweets is used as the main measurable feature of a participant. After
introducing the influence measures used by the two ranking algorithms mentioned
above, the authors described their own experiments and results. They point out that
for most users, the two methods agree but sometimes there are some surprising
differences in the rank. While Katz uses simple frequency proportional mention
weights, PageRank penalizes those mentions which come from less important
participants. As the result of a deeper analysis, the authors conclude that PageRank
has proved to be more robust a solution to identify influential users than the other
top approaches.
The second paper deals with the problem of the Synthetic Aperture Radar
(SAR) imagery filtering. The problem in SAR images is the presence of atmo-
spheric conditions introducing noise and partial covering by clouds. In addition, a
large amount of speckle, a multiplicative non-Gaussian noise degrades the quality
of the SAR images. The study compares two classes of aggregation operators: the

v
vi Preface

Weighted Mean (WM) and the Weighted Ordered Average (OWA). A new family,
merging the advantages of both previously mentioned classes, the Weighted OWA
Operators (WOWA), offers a combination of advantages of the abovementioned
two methods. The paper describes in detail the procedure of filtering the SAR
images, with all three approaches. For learning the parameters, the authors use a
Genetic Algorithm (GA), the classical evolutionary approach. As the GA is not one
of the most efficient evolutionary methods, it is no surprise that the authors con-
clude that there is a strong dependence of the quality of the results on the number of
generations. The final result points toward all proposed filters being useful, or even
at the best—depending on the type of image and the circumstances of application.
The directions of future research are indicated at the end.
The third paper proposes a novel combination of wavelet transform and fuzzy
rule interpolation for evaluating very big data obtained from the telecommunication
field. The engineering problem is the pre-evaluation of copper wire pairs from the
point of view of transmission speed in SHDSL transmission networks. In the
present best practice, the insertion loss must be measured at a large number of
different frequencies for every wire pair. Based on several 100,000s of real mea-
surement results, it is proposed that the combination of the Haar and Daubechies-4
wavelets combined with the Stabilized KH rule interpolation leads to a much better
prediction of the transmission quality than any method known so far in the liter-
ature. The conclusion suggests that this new method might be useful in a much
wider field of big data analysis, as well.
The fourth paper introduces a logic called the Molecular Interaction Logic,
which semantically characterizes the Molecular Interaction Maps (MIM) and,
moreover, makes it possible to apply deductive and abductive reasoning on MIMs
in order to find inconsistencies, answer queries, and infer important properties about
those networks. This logic can be applied to different metabolic networks, as the
one given by the cancer, which can sometimes appear in a cell as a result of some
pathology in a metabolic pathway.
The fifth paper is based on another logic. In this case in the general fuzzy logic
programming called multi-adjoint logic programming. This framework was intro-
duced by Medina et al. in 2001 as a general framework in which the minimal
mathematical requirements are only considered in order to ensure the main prop-
erties given in the diverse usual logic programming frameworks, such as in
possibilistic logic programming, monotonic and residuated logic programming,
fuzzy logic programming, etc. Since its introduction, this framework has been
widely developed from the theoretical and applied aspects by diverse authors. In
this paper, the authors perform experiments which have shown the benefits of using
a new c-unfolding transformation which reuses some variants of program
transformation techniques based on unfolding, which have been largely exploited
in the pure functional—not fuzzy—setting. Specifically, this new c-unfolding
transformation has been applied to fuzzy connectives, and the authors have shown
how to improve the efficiency of the proper unfolding process by reusing the very
well-known concept of dependency graph. Furthermore, the paper includes a cost
analysis and discussions on practical aspects.
Preface vii

The last two papers are in the area of the Formal Concept Analysis (FCA). This
framework arose as a mathematical theory for qualitative data analysis and has
become an interesting research topic both on its mathematical foundations and on
its multiple applications. Specifically, the sixth paper provides an overview of
different generalizations of formal concept analysis based on fuzzy sets. First of all,
a common platform for early fuzzy approaches is included. Then, different recently
fuzzy extensions have been recalled, such as the generalized extension given by
Krajči, the multi-adjoint concept lattices introduced by Medina et al., heterogeneous
extension studied by diverse research groups exemplified by those of Krajči,
Medina et al., Pócs, Popescu, using heterogeneous one-sided extension given by
Butka and Pócs, and the higher order extension presented by Krídlo et al. The paper
also discusses connections between the related approaches.
The seventh paper is related to knowledge extraction from databases using rules,
which is a compact tool in the representation of the knowledge. Based on concepts
from lattice theory, this paper has introduced a new kind of attribute implications
considering the fuzzy notions of support and confidence. The authors have also
studied different properties of the particular case in which the set of attributes are
intensions and, finally, they have also included an application to clustering for size
reduction of concept lattices.
We would like to express our gratitude to all the authors for their interesting,
novel, and inspiring contributions. Peer-reviewers also deserve our deep appreci-
ation because their deep and valuable remarks and suggestions have considerably
improved the contributions.
And last but not least, we wish to thank Dr. Tom Ditzinger, Dr. Leontina di
Cecco, and Mr. Holger Schaepe for their dedication and help to implement and
finish this large publication project on time maintaining the highest publication
standards.

Gyõr, Budapest, Hungary László T. Kóczy

Puerto Real, Cádiz, Spain Jesús Medina
Contents

Page Rank Versus Katz: Is the Centrality Algorithm Choice Relevant

to Measure User Influence in Twitter? . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Hugo Rosa, Joao P. Carvalho, Ramon Astudillo and Fernando Batista
Weighted Means Based Filters for SAR Imagery . . . . . . . . . . . . . . . . . . 11
L. Torres, J. C. Becceneri, C. C. Freitas, S. J. S. Sant’Anna
and S. Sandri
On Combination of Wavelet Transformation and Stabilized KH
Interpolation for Fuzzy Inferences Based on High Dimensional
Sampled Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Ferenc Lilik, Szilvia Nagy and László T. Kóczy
Abductive Reasoning on Molecular Interaction Maps . . . . . . . . . . . . . . 43
Jean-Marc Alliot, Robert Demolombe, Luis Fariñas del Cerro,
Martín Diéguez and Naji Obeid
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic
Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Pedro J. Morcillo and Ginés Moreno
On Fuzzy Generalizations of Concept Lattices . . . . . . . . . . . . . . . . . . . . 79
Lubomir Antoni, Stanislav Krajči and Ondrej Krídlo
Generating Fuzzy Attribute Rules Via Fuzzy Formal
Concept Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Valentín Liñeiro-Barea, Jesús Medina and Inmaculada Medina-Bulo

ix
Page Rank Versus Katz: Is the Centrality
Algorithm Choice Relevant to Measure
User Influence in Twitter?

Hugo Rosa, Joao P. Carvalho, Ramon Astudillo and Fernando Batista

Abstract Microblogs, such as Twitter, have become an important socio-political

analysis tool. One of the most important tasks in such analysis is the detection of
relevant actors within a given topic through data mining, i.e., identifying who are the
most influential participants discussing the topic. Even if there is no gold standard
for such task, the adequacy of graph based centrality tools such as PageRank and
Katz is well documented. In this paper, we present a case study based on a “London
Riots” Twitter database, where we show that Katz is not as adequate for the task
of important actors detection since it fails to detect what we refer to as “indirect
gloating”, the situation where an actor capitalizes on other actors referring to him.

Keywords Page Rank · Katz · User Influence · Twitter · Data Mining

1 Introduction

Nowadays, there are 288 million active users on Twitter and more than 500 million
tweets are produced per day [17]. Through short messages, users can post about their

This work was supported by national funds through Fundação para a Ciência e a
Tecnologia (FCT) under project PTDC/IVC-ESCT/4919/2012 and funds with reference
UID/CEC/50021/2013.

H. Rosa (B) · J. P. Carvalho · R. Astudillo · F. Batista

INESC-ID, Lisboa, Portugal
e-mail: [email protected]
R. Astudillo
e-mail: [email protected]
J. P. Carvalho
Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal
e-mail: [email protected]
F. Batista
ISCTE-IUL - Instituto Universitário de Lisboa, Lisbon, Portugal
e-mail: [email protected]

© Springer International Publishing AG 2018 1

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_1
2 H. Rosa et al.

feelings, important events and talk amongst each other. Twitter has become so much
of a force to be reckoned with, that anybody from major brands and institutions, to
celebrities and political figures use it to further assert their position and make their
voice heard. The impact of Twitter on the Arab Spring [6] and how it beat the all news
media to the announcement of Michael Jackson’s death [15], are just a few examples
of Twitter’s role in society. When big events occur, it is common for users to post
about it in such fashion, that it becomes a trending topic, all the while being unaware
from where it stemmed or who made it relevant. The question we wish to answer is:
“Which users were important in disseminating and discussing a given topic?”
Much like real life, some users carry more influence and authority than others.
Determining user relevance is vital to help determine trend setters [16]. The user’s
relevance must take into account not only global metrics that include the user’s level
of activity within the social network, but also his impact in a given topic [18]. Empir-
ically speaking, an influential person can be described as someone with the ability
to change the opinion of many, in order to reflect his own. While [13] supports this
statement, claiming that “a minority of users, called influentials, excel in persuading
others”, more modern approaches [4] seem to emphasize the importance of inter-
personal relationships amongst ordinary users, reinforcing that people make choices
based on the opinions of their peers.
In [2], three measures of influence were taken into account: “in-degree is the
number of people who follow a user; re-tweets mean the number of times others
forward a user’s tweet; and mentions mean the number of times others mention a
user’s name”. It concluded that while in-degree measure is useful to identify users
who get a lot of attention, it “is not related to other important notions of influence such
as engaging audience”. Instead “it is more influential to have an active audience who
re-tweets or mentions the user”. In [8], the conclusion was made that within Twitter,
“news outlets, regardless of follower count, influence large amounts of followers to
republish their content to other users”, while “celebrities with higher follower totals
foster more conversation than provide retweetable content”. The authors in [12]
created a framework named “InfluenceTracker”, that rates the impact of a Twitter
account taking into consideration an Influence Metric, based on the ratio between
the number of followers of a user and the users it follows, and the amount of recent
activity of a given account. Much like [2], it also shows “that the number of followers
a user has, is not sufficient to guarantee the maximum diffusion of information (…)
because, these followers should not only be active Twitter users, but also have impact
on the network”.
In this paper, we analyze how two well known network analysis algorithms,
PageRank and Katz, affect the computation of mention-based user influence in Twit-
ter. Although these two methods have previously been compared [11] and found
to have been equivalent, we show that the same conclusion does not apply in the
context of social networks, and that PageRank is indeed more adequate. We base
our conclusions on a real world case study of the 2011 London Riots, since it was
an important social event where Twitter users were said to have played a role in its
origin and dissemination.
Page Rank Versus Katz: Is the Centrality Algorithm Choice Relevant to … 3

2 User Influence Representation

We propose a graph representation of user’s influence based on “mentions”. When-

ever a user is mentioned in a tweet’s text, using the @user tag, a link is made from
the creator of the tweet, to the mentioned user. For example, the tweet “Do you think
we can we get out of this financial crisis, @user B?”, from @user A, creates the link:
@user A −→ @user B. This is also true for re-tweets, e.g. the tweet “RT @userC
The crisis is everywhere!” from @user A, creates the link: @user A −→ @userC.
This representation not only is an exact structural replica of the communication
web between users, but it also provides dynamism to how influence can be given and
taken across the graph.
In graph theory and network analysis, the concept of centrality refers to the iden-
tification of the most important vertices’s within a graph, i.e., most important users.
We therefore define a graph G(V, E) where V is the set of users and E is the set of
directed links between them.

3 Network Analysis Algorithms

The computation of user influence is done by applying a centrality based algorithm

to the graph presented in Sect. 2. Here we present two of the most well-known and
used centrality algorithms, Page Rank and Katz.

3.1 PageRank

Arguably the most well known centrality algorithm is PageRank [9]. It is one of
Google’s methods to its search engine and it was created as way for computing a
ranking for every web page based on the graph of the web uses. In this algorithm,
web pages are nodes, while back-links form the edges of the graph (Fig. 1). It is
defined by Eq. 1 as P R(vi ) of a page vi .

1−d  P R(v j )
P Rvi = +d (1)
N v ∈M(v )
L(v j )
j i

It can be intuitively said about Eq. 1, that a page has high rank if the sum of
the ranks of its back-links is high. In it, v j is the sum ranges over all pages that
has a link to vi , L(v j ) is the number of outgoing links from v j , N is the number
of documents/nodes in the collection and d is the damping factor. The PageRank
is considered to be a random walk model, because the weight of a page vi is “the
probability that a random walker (which continues to follow arbitrary links to move
from page to page) will be at vi at any given time. The damping factor corresponds to
4 H. Rosa et al.

Fig. 1 A and B are

back-links of C

the probability of the random walk to jump to an arbitrary page, rather than to follow
a link, on the Web. It is required to reduce the effects on the PageRank computation
of loops and dangling links in the Web” [11]. Dangling links are “simply links that
point to any page with no outgoing links (…) they affect the model because it is
not clear where their weight should be distributed” [9]. The true value that Google
uses for damping factor is unknown, but it has become common to use d = 0.85 in
the literature. A lower value of d implies that the graph’s structure is less respected,
therefore making the “walker” more random and less strict.

3.2 Katz

Another well known method is the Katz algorithm [7]. It is a generalization of a

back-link counting method where the weight of each node is “determined by the
number of directed paths that ends in the page, where the influence of longer paths is
attenuated by a decay factor” and “the length of a path is defined to be the number of
edges it contains” [11]. It is defined by Eq. 2 “where N (vi , k) is the number of paths
of length k that starts at any page and ends at vi and α is the decay factor. Solutions
for all the pages are guaranteed to exist as long as α is smaller than λ > 1, where
1/λ is the maximum in-degree of any page” [11].
∞

Ivi = [α k N (vi , k)] (2)
k=0

It was shown in [11] that “Katz status index may be considered a more general
form of PageRank because in can be modified, within a reasonable range, to be
equivalent to PageRank” and that under a “relaxed definition of equivalence (…)
PageRank and Katz status index is practically equivalent to each other” as long as
the number of outgoing links from any vertex is the same throughout the graph,
which is very unlikely for graph modeled from a social network. On the other hand,
“it is also possible to modify PageRank to become completely equivalent to Katz
status index”, however, in that case, “the modified PageRank is no long a random
work model because it can no longer be modeled from a probabilistic standpoint”
[11].
Page Rank Versus Katz: Is the Centrality Algorithm Choice Relevant to … 5

4 Dataset

In order to test the network analysis methods presented above, a database from the
London Riots in 2011 [3] was used. The London Riots of 2011 was an event that
took place between the 6th and 11th August 2011, where thousands of people rioted
in several boroughs of London with the resulting chaos generated looting, arson, and
mass deployment of police. Although Twitter was said to be a communication tool
for rioting groups to organize themselves, there is little evidence that it was used
to promote illegal activities at the time, though it was useful for spreading word
about subsequent events. According to [5], Twitter played a big role spreading the
news about what was happening and “was a valuable tool for mobilizing support for
the post-riot clean-up and for organizing specific clean-up activities”. Therefore it
constitutes a prime data sample to study how users exert influence in social networks,
when confronted with such a high stakes event.
The Guardian Newspaper made public a list of tweets from 200 influential twitter
users, which contains 17,795 riot related tweets and an overall dataset of 1,132,938
tweets. Using a Topic Detection algorithm [1], we obtained an additional 25,757
unhastagged tweets about the London Riots. It consists of a Twitter Topic Fuzzy
Fingerprint algorithm [14] that provides a weighted rank of keywords for each topic
in order to identify a smaller subset of tweets within scope. This method has proven
to achieve better results than other well known classifiers in the context of detecting
Topics within Twitter, while also being faster in execution. The sum of posting and
mentioned users is 13,765 (vertices) and it has 19,993 different user mentions (edges),
edges
achieving a network connectivity ratio of ver tices
= 1.46.

5 Experiments and Results

In this section, we compare the results of ranking the most influential users using Page
Rank, Katz and a mentions based baseline. We proceed by performing an empirical
analysis of the users in order to ascertain their degree of influence and their position
in the ranks. The graphs and ranking were calculated using Graph-Tool [10].
Table 1 shows how both network analysis algorithms behave while highlighting
the rank differences (shown by the arrows in the last column). A “Mentions rank”
is used as a base line. Figure 2 provides a visual tool to the graph, as provided by
PageRank.
There is an obvious relation between the number of mentions and the ranking pro-
vided by the application of both algorithms: the highest ranked users in either Katz and
PageRank, are some of the most mentioned users in our dataset. In fact, the relation is
more clear between Katz and the baseline Mentions based ranking: Table 1 shows that
the rank in both approaches is always either identical (@guardian, @skynewsbreak,
@gmpolice, etc…) or at most separated by two positions (@richardpbacon is ranked
27th based on mentions, and 29th based on Katz). In order to determine the relation
 6

Table 1 London riots top 20 most influential users according to page rank, and comparison with Katz
User Mentions PageRank Katz
# Rank Score Rank Score Rank
@guardian 160 2 0.0002854 1 0.022157 2
@skynewsbreak 178 1 0.0002512 2 0.023479 1
@gmpolice 122 4 0.0002128 3 0.019009 4
@riotcleanup 107 6 0.0001767 4 0.017992 6 
@prodnose 67 14 0.0001761 5 0.014022 15 
@metpoliceuk 116 5 0.0001494 6 0.018709 5
@marcreeves 69 11 0.0001476 7 0.014195 12 
@piersmorgan 78 8 0.0001465 8 0.014959 9
@scdsoundsystem 69 12 0.0001442 9 0.014190 13 
@subedited 70 10 0.0001337 10 0.014278 11
@youtube 48 20 0.0001257 11 0.012424 20 
@bbcnews 94 7 0.0001256 12 0.016426 8 
@mattkmoore 62 15 0.0001237 13 0.013614 16 
@richardpbacon 40 27 0.0001218 14 0.011771 29 
@lbc973 34 35 0.0001150 15 0.011432 34 
@skynews 74 9 0.0001113 16 0.014638 10 
@bengoldacre 61 17 0.0001055 17 0.013526 17
@bbcnewsnight 68 13 0.0000988 18 0.014123 14 
@tom_watson 44 21 0.0000968 19 0.012107 22 
@paullewis 129 3 0.0000954 20 0.019602 3 
...
@juliangbell 61 16 0.0000275 188 0.0166597 7 
The arrows indicate most relevant rank differences
H. Rosa et al.
Page Rank Versus Katz: Is the Centrality Algorithm Choice Relevant to … 7

Fig. 2 User influence page rank graph - larger circles indicate larger user influence

between PageRank and “Mentions Rank”, the Spearman correlation was calculated
having achieved a value of ρ = 0.9372, which means they are heavily correlated.
However, when limiting this calculation to the top 20, it changed to ρ = 0.5535,
which implies that for the top users, just looking at the number of mentions, is not
enough to determine influence.
An empirical analysis also shows that both Page Rank and Katz largely agree
upon the ranking of most users, namely on the top two users: (i) @guardian, Twitter
account of the world famous newspaper “The Guardian”; (ii) @skynewsbreak, Twit-
ter account of the news team at Sky News TV channel. This outcome agrees with [8]
previous statement, that, “news outlets, regardless of follower count, influence large
amounts of followers to republish their content to other users” and can be justified by
the incredibly high London Riots news coverage. Other users seem to fit the profile,
namely @gmpoliceq, @bbcnews and @skynews. Most of the other users are either
political figures, political commentators or jornalists (@marcreeves, @piersmorgan,
and @mattkmoore).
However, when looking more closely at Page Rank versus Katz rankings, it is
also possible to realize some notorious differences: Katz’s third and seventh top
ranked users are not in PageRank’s top users. The reasons behind these differences
in the ranking positions should be thoroughly analyzed since they could highlight
the strengths and weaknesses of each algorithm in what concerns their capability to
8 H. Rosa et al.

express user influence in social networks. The two cases end up being different and
should be treated separately: (i) @paullewis, ranked 3rd by Katz shows up at 20th
according to PageRank; (ii) @juliangbell, ranked 7th by Katz shows up at 188th
according to PageRank.
The reason behind @paullewis high placement in the Katz rank is the number
of mentions. As said previously, Katz is a generalization of a back-link counting
method, which means the more back-links/mentions a user has, the higher it will be
on the ranking. This user has 129 mentions, but PageRank penalizes it, because it is
mentioned by least important users, which means a less sum weight is being trans-
fered to it in the iterative process. This logic also applies to users @bbcnewsnight,
@skynews and @bbcnews. Additionally, @paullewis is also an active mentioning
user, having mentioned other users a total of 14 tweets, while @skynewsbreak and
@guardian have mentioned none. As a consequence, Paul Lewis transfers its influ-
ence across the network while the others simply harvest it. There are several users
that drop in ranking from PageRank to Katz for the very same reason. Users such as
@prodnose, @marcreeves and @youtube do not have enough mentions for Katz to
rank them higher.
User @juliangbell, despite mentioned often (61 times), is down on the PageRank
because of indirect gloating, i.e., he retweets tweets that are mentioning himself:
“@LabourLocalGov #Ealing Riot Mtg: @juliangbell speech http://t.co/3BNW0q6”
was posted by @juliangbell himself. The user is posting somebody else’s re-tweet
of one of his tweets. As a consequence a link/edge was created from @juliangbell
to @LabourLocalGov, but also from @juliangbell to himself, since his username is
mentioned in his own tweet. Julian Bell is a political figure, making it acceptable
that he would have a role in discussing the London Riots, but the self congratulatory
behavior of re-tweeting other people’s mentions of himself, is contradictory with
the idea of disseminating the topic across the network. While Katz is not able to
detect this effect, PageRank automatically corrects it, which is why, contrary to
what is mentioned in previous works [11], it is our comprehension that Katz is not
equivalent to PageRank in the task of detecting user relevance in social networks
such as Twitter.

6 Conclusions

With this study, we have shown that in the context of user influence in Twitter,
PageRank and Katz are not equal in performance, thus disproving previous claims.
PageRank has proved a more robust solution to identify influential users in discussing
and spreading a given relevant topic, specially when considering how it deals with
indirect gloating, an item Katz fails to penalize.
Page Rank Versus Katz: Is the Centrality Algorithm Choice Relevant to … 9

References

1. Carvalho, J.P., Pedro, V., Batista, F.: Towards intelligent mining of public social networks’
influence in society. In: IFSA World Congress and NAFIPS Annual Meeting (IFSA/NAFIPS),
pp. 478 – 483. Edmonton, Canada (June 2013)
2. Cha, M., Haddadi, H., Benevenuto, F., Gummadi, K.P.: Measuring user influence in twitter: the
million follower fallacy. In: In ICWSM ’10: Proceedings of International AAAI Conference
on Weblogs and Social (2010)
3. Crockett, K.S.R.: Twitter riot dataset (tw-short) (2011)
4. Domingos, P., Richardson, M.: Mining the network value of customers. In: Proceedings of the
Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Min-
ing, pp. 57–66. KDD ’01, ACM, New York, USA (2001). http://doi.acm.org/10.1145/502512.
502525
5. Guardian, T.: http://www.theguardian.com/uk/2011/dec/07/twitter-riots-how-news-spread
6. Huang, C.: Facebook and twitter key to arab spring uprisings: report. http://www.thenational.
ae/news/uae-news/facebook-and-twitter-key-to-arab-spring-uprisings-report (June 2011).
Accessed 02 May 2014
7. Katz, L.: A new status index derived from sociometric analysis. Psychometrika 18(1), 39–43
(March 1953). http://ideas.repec.org/a/spr/psycho/v18y1953i1p39-43.html
8. Leavitt, A., Burchard, E., Fisher, D., Gilbert, S.: The Influentials: New Approaches for Ana-
lyzing Influence on Twitter (2009)
9. Page, L., Brin, S., Motwani, R., Winograd, T.: The Pagerank Citation Ranking: Bringing Order
to the Web (1999)
10. Peixoto, T.: https://about.twitter.com/company
11. Phuoc, N.Q., Kim, S.R., Lee, H.K., Kim, H.: Pagerank vs. katz status index, a theoretical
approach. In: Proceedings of the 2009 Fourth International Conference on Computer Sciences
and Convergence Information Technology, pp. 1276–1279. ICCIT ’09, IEEE Computer Society,
Washington, DC, USA (2009). http://dx.doi.org/10.1109/ICCIT.2009.272
12. Razis, G., Anagnostopoulos, I.: Influencetracker: Rating the Impact of a Twitter Account.
CoRR (2014). arXiv:1404.5239
13. Rogers, E.M.: Diffusion of Innovations (1962)
14. Rosa, H., Batista, F., Carvalho, J.P.: Twitter topic fuzzy fingerprints. In: WCCI2014, FUZZ-
IEEE, 2014 IEEE World Congress on Computational Intelligence, International Conference
on Fuzzy Systems, pp. 776–783. IEEE Xplorer, Beijing, China (July 2014)
15. Sankaranarayanan, J., Samet, H., Teitler, B.E., Lieberman, M.D., Sperling, J.: Twitterstand:
News in tweets. In: Proceedings of the 17th ACM SIGSPATIAL International Conference on
Advances in Geographic Information Systems, pp. 42–51. GIS ’09, ACM, New York, NY, USA
(2009). http://doi.acm.org/10.1145/1653771.1653781
16. Tinati, R., Carr, L., Hall, W., Bentwood, J.: Identifying communicator roles in twitter. In:
Proceedings of the 21st International Conference Companion on World Wide Web, pp. 1161–
1168. WWW ’12 Companion, ACM, New York, NY, USA (2012). http://doi.acm.org/10.1145/
2187980.2188256
17. Twitter: https://about.twitter.com/company
18. Weng, J., Lim, E.P., Jiang, J., He, Q.: Twitterrank: finding topic-sensitive influential twitterers.
In: Proceedings of the Third ACM International Conference on Web Search and Data Mining,
pp. 261–270. WSDM ’10, ACM, New York, NY, USA (2010). http://doi.acm.org/10.1145/
1718487.1718520
Weighted Means Based Filters for SAR
Imagery

L. Torres, J. C. Becceneri, C. C. Freitas, S. J. S. Sant’Anna and S. Sandri

Abstract We address parameter learning for three families of filters for SAR
imagery, based on WM, OWA and WOWA families of aggregation operators. The
values in the weight vector associated to a WM filter correspond to the same posi-
tions in the input, whereas those in OWA filters consider the ordered positions of
the input. WOWA operators make use of both vectors to weight an input data vector.
Here we use Genetic Algorithms to learn the weight vectors for OWA, WM and
WOWA filters and assess their use in reducing speckle in SAR imagery. We present
an application using simulated images derived from a real-world scene and compare
our results with those issued by a set of filters from the literature.

1 Introduction

Synthetic Aperture Radar (SAR) sensors are not as adversely affected by atmospheric
conditions and the presence of clouds as optical sensors [13]. Moreover, they can
be used at any time of day or night. The visual quality of SAR images is, however,
degraded by sudden variations in image intensity with a salt and pepper pattern, due
to the existence of a great amount of speckle, a multiplicative non-Gaussian noise,
proportional to the intensity of the received signal [17]. Speckle in SAR image
hampers the interpretation and analysis of this kind of image, as well as reduces the
effectiveness of some image processing tasks such as segmentation and classification.
For this reason, in order to process a SAR images, it is usually recommended to apply
a filter before the segmentation or classification processes, even though the image
quality with respect to some features may decrease, such as signal to noise ratio,
spatial resolution, among others.
Filters for SAR imagery can be classified according to whether they take into
account a model for speckle. A class of important representative of model-independent
filters is the Ordered Statistical Filters (OSF) [2], based on order statistics [18].

L. Torres · J. C. Becceneri · C. C. Freitas · S. J. S. Sant’Anna · S. Sandri (B)

Instituto Nacional de Pesquisas Espaciais (LAC/INPE), São José dos Campos, São
Paulo12227-010, Brazil
e-mail: [email protected]

© Springer International Publishing AG 2018 11

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_2
12 L. Torres et al.

Model-dependent filters are significantly more complex than the model-independent

ones. Well-known representatives of this class of filters are the so-called Lee filter [10,
11] and its variations, such as the Refined Lee filter [12]. A more recent approach on
model-dependent filters, called SDNLM (Stochastic Distances and Nonlocal Means)
[27], is itself based on Buades et al.’s Nonlocal Means methodology [3].
Two important families of weighted aggregation operators are the Weighted Mean
(WM) and the Weighted Ordered Average (OWA) [36]. In WM, each value in a weight
vector p measures the importance of an information source with independence of the
value that the source has captured, whereas in OWA, each value in a weight vector w
measures the importance of a value (in relation to other values) with independence
of the information source that has captured it [24].
OWA and WM filters are model-independent filters based on the use of OWA
and WM operators, respectively. OWA filters are a particular representation of OSF,
where, by definition, all coefficients are non-negative and sum up to 1. In [28, 29], we
investigated the use of OWA filters to reduce speckle in SAR imagery and proposed
strategies to learn vector w using Genetic Algorithms (GA) [6, 9]. In both works, we
only dealt with intensity images; in [28] we addressed a single polarization (HH) and
3 × 3 windows, whereas in [29] we addressed filters for images in three polarizations
(HH, HV and VV) using 5 × 5 windows.
The good results obtained by learning WM filters, that use vector p, and OWA
filters, that use vector w, led us to investigate filters that would use both vectors p and
w. Experiments that consisted in applying OWA and WM filters consecutively gave
very poor results, so we turned our attention to families of operators that generalized
both OWA and WM operators.
Weighted OWA operators (WOWA), proposed by Torra in 1997 [24], aims at
taking advantage of both OWA and WM operators (see also [25, 26]). They have
however an extra requirement, the choice of a particular type of function (called φ)
to generate a vector weight ω, from p and w. We have first introduced the concept
of WOWA filters in [30], and studied how the vectors p and w should be learned
(sequentially or concomitantly), considering intensity images in a single polarization
(HH). We also addressed learning WOWA filters for images obtained using the
arithmetic mean of intensity images in polarizations HH, HV and VV [31].
In the present paper, we compare WM, OWA and WOWA filters, considering
intensity images in three polarizations (HH, HV and VV), and study the influence of
the number of generations in the GA on the performance of each filter. We report a
series of experiments based on a fragment of a phantom described in [20]. To learn
the WOWA weight vectors p and w, we use synthetic images in polarizations HH,
HV and VV, that have been simulated using the parameters for Wishart distributions
from a real-world scene, estimated in [23]. In order to assess the quality of the results
produced by the GA, we use the Normalized Mean Square Error (NMSE) as fitness
measure (see [1]). The results produced by these filters are compared to those issued
by two model-dependent filters proposed in [12, 27].
This work is organized as follows. Section 2 discusses some basic aspects of SAR
imagery. Section 3 describes WM, OWA and WOWA filters and Sect. 4 explains how
Weighted Means Based Filters for SAR Imagery 13

WOWA filters for SAR imagery are learned with the adopted GA. Section 5 presents
an experiment in SAR imagery and Sect. 6 finally brings the conclusion.

2 Basic Concepts on SAR Imagery

SAR sensors operate in microwaves frequencies and generate images by measuring

the energy reflected by targets on the earth’s surface. SAR systems present side-
looking geometry and are mounted on a platform which moves along a predefined
trajectory, transmitting pulses at certain time intervals and receiving the backscattered
signal. The backscattered signal is mainly affected by the target’s electromagnetic
characteristics and imaging geometry, as well as the particular features of the sensor
system, such as frequency, polarization, incident angle, spatial resolution, etc. (see
[19] for details on SAR systems).
Conventional imaging radars operate with a single, fixed-polarization antenna for
both transmission and reception of radio frequencies signals [32]. In these systems,
a single measure (scattering coefficient) is taken and is referred to as SAR data.
Polarimetric radars usually operate employing horizontal (H) and vertical (V) linearly
polarized signals for transmitting and/or receiving the energy.
The polarization combination for transmitted and received signals can lead to
complex images, called polarimetric SAR data (PolSAR), which are also affected by
speckle. These images are formed by a 2 × 2 scattering matrix, where the components
are denoted by S H H and SV V (co-polarized data), and S H V and SV H (cross-polarized
data). It is important to note that for reciprocal medium and monostatic systems, the
components S H V and SV H are equal. Therefore, the four elements of the scattering
matrix can be represented by a three-place vector as [S H H S H V SV V ]. Multiplying
∗ ∗ ∗ t
the vector [SHH SHV SVV ] by its transposed conjugated vector [SHH SHV SVV ],
we obtain a 3 × 3 covariance matrix. The main diagonal elements, denoted by IHH ,
IHV , and IVV , represent the intensity values of polarizations HH, HV and VV, respec-
tively.

2.1 Filters for SAR Imagery

Given a window in an image, a filter simply substitutes the value of its central pixel by
a function of the values of the pixels in the window. Two of the most basic filters use
the arithmetic mean and the median as filtering function. In SAR imagery, the mean
filter tends to reduce the speckle but it also tends to indiscriminately blur the image
[14]. The median filter, on the other hand, reduces erratic variations by eliminating
the lowest and highest pixel values [21].
Most filters employ the convolution operation. Given an image I , whose pixels
take values in R, a m × m window around the central pixel (x, y) in I , and a matrix
of coefficients γ : {−m, ..., 0, ..., m}2 → R, the result of convolution for (x, y) in
14 L. Torres et al.

the filtered image Iγ is calculated as

 
Iγ (x, y) = γ (i, j) × I (x + i, y + j).
i=−m,m j=−m,m

In Order Statistics Filters [2], the result of filtering for a given pixel is the linear
combination of the ordered values of the pixels in the window around that pixel.
They belong to the larger class of non-linear filters based on order statistics [18],
being an application of L-estimators. An OSF is thus obtained when a convolution
filter is applied on the ordered statistic of the pixel values in a window.
Some examples of model-independent filters are the directional means filters, in
which only pixels in one of the twelve regions of the six orthogonal directions are
considered (diagonals, rows and columns) [21] and the local region filters (see [22]),
in which the window is divided in eight regions based on angular position, and the
central pixel is replaced by the mean value of the subregion presenting the lowest
variance.
The adoption of a model to the noise leads to more complex filters. One of such
filters is the so-called Lee filter, in which speckle reduction is based on multiplicative
noise model using the minimum mean-square error (MMSE) criterion [10, 11]. The
Refined Lee filter [12], here called R-Lee filter, is an improved version of the Lee
filter, and uses a methodology for selecting neighboring pixels with similar scattering
characteristics.
Buades et al. [3] proposed a model-dependentt methodology that is well-suited for
decreasing additive Gaussian noise, called Nonlocal Means (NL-means), which uses
similarities between patches as the weights of a mean filter. The SDNLM (Stochastic
Distances and Nonlocal Means) filter [27] is an adaptive nonlinear extension of the
NL-means algorithm filter, in which overlapping samples are compared based on
stochastic distances between distributions, and the p-values resulting from such
comparisons are used to build the weights of an adaptive linear filter.

2.2 Image Quality Assessment for SAR Imagery

Assessing the performance of image filters is very hard [34]. Two important indices
to measure the quality of filtered images are NMSE and SSIM, described in the
following. Index NMSE (Normalized Mean Square Error) is a general purpose error
measure, widely used in image processing (see [1]). Let r be the perfect information
data and s an approximation of r ; NMSE is calculated as:
n
j=1 (r j − s j )
2
NMSE = n 2
, (1)
j=1 r j
Weighted Means Based Filters for SAR Imagery 15

where r j and s j refer to values in r and s at the same coordinates (the position of a
given pixel in the case of images). NMSE always yield positive values, and the lower
its value, the better is the approximation considered to be.
Index SSIM (Structural SIMilarity) measures the similarity between two scalar-
valued images and can be viewed as a quality measure of one of them, when the other
image is regarded as of perfect quality [35]. It is an improved version of the universal
image quality index proposed proposed by [33]. This index takes into account three
factors: (i) correlation between edges; (ii) brightness distortion; and (iii) distortion
contrast. Let r and s be the perfect information and its approximation, respectively;
SSIM is calculated as
Cov(r, s)+α1 2r s +α2 σr 
2 σs +α3
SSIM(r, s) = × 2 2 × 2 , (2)
σr 
σs +α1 r +s +α2  σr +σs2 +α3

where r and s are sample means,  σr2 and 

σs2 are the sample variances, Cov(r, s) is the
sample covariance between r and s, and constants α1 , α2 and α3 are used the index
stabilization. SSIM ranges in the [−1, 1] interval, and the higher its value, the better
is the approximation considered to be.
Some other measures that can be used to evaluate the quality of SAR imagery
are, for instance, the equivalent number of looks (ENL), usually applied to intensity
images in homogeneous areas, and index βρ , a correlation measure between the edges
of two images (see [16]).

3 Filtering Based on WM, OWA and WOWA Operators

In the following, we first give the definitions of WM, OWA and WOWA families of
operators. We then describe the filters obtained using these operators.

3.1 WM, OWA and WOWA Operators

Let p be a weighting vector of dimension n (p = [ p1 p2 ... pn ]), such that:

– (i) pi ∈ [0, 1];
– (ii) Σi pi = 1.
A mapping f pwm : R n → R is a Weighted Mean Operator (WM) of dimension n,
associated to p , if:
f pwm (a1 , ..., an ) = Σi pi × ai . (3)

Let w be a weighting vector of dimension n (w= [w1 w2 ... wn ]), such that:
– (i) wi ∈ [0, 1];
– (ii) Σi wi = 1.
16 L. Torres et al.

A mapping f wowa : R n → R is an Ordered Weighted Average Operator (OWA) of

dimension n, associated to w, if [36]:

f wowa (a1 , ..., an ) = Σi wi × aσ (i) , (4)

where {σ (1), ..., σ (n)} is a permutation of {1, ..., n}, such that aσ (i−1) ≥ aσ (i) for
all i = {2,..., n} (i.e., aσ (i) is the i-th largest element in {a1 , ..., an }).
Some well-known OWA operators are the mean, min, max and median, which are
obtained with OWA vectors wmean , wmin , wmax , and wmed , respectively. For instance,
taking n = 3, we have: wmean = [1/3, 1/3, 1/3], wmin = [0, 0, 1], wmax = [1, 0, 0],
and wmed = [0, 1, 0].
The measures of or ness, and andness [36], associated with a given vector w are
defined as follows:
or ness(w) = n−1
1
Σi (n − i)wi
andness(w) = 1 − or ness(w)
Functions or ness and andness describe an adjustment of the levels of “or” and
“and”, respectively, in the aggregation of a set of values.
wowa
Let p and w be weighting vectors as given above. A mapping fw, p : R n → R is a
Weighted Ordered Weighted Average (WOWA) operator of dimension n, associated
to p and w, if [24]:
wowa
f w, p,φ (a1 , ..., an ) = Σi ωi × aσ (i) , (5)

where {σ (1), ..., σ (n)} is a permutation of {1, ..., n}, for all i = {2,..., n}, such that
aσ (i−1) ≥ aσ (i) . Weight ωi is defined as

ωi = φ(Pσ (i)) − φ(Pσ (i − 1)), (6)

Pσ (i) = Σ j≤i pσ ( j) , (7)

 φ is a monotone increasing function that interpolates points (0, 0) and (i/n,

and
j≤i w j ), i = 1, n. Function φ is required to be a straight line when the points can
be interpolated in a linear way. Torra [24] proves that the ωi ’s compose a weighting
vector of dimension n (ω = [ω1 ... ωn ]), such that:
– (i) ωi ∈ [0, 1];
– (ii) Σi ωi = 1.
In [24], we can also find examples of non-linear functions to implement φ and the
proof that OWA and WM are particular cases of WOWA operators.
WOWA operators, as well as OWA and WM operators, are particular cases of
Choquet integrals [25, 26]. Even though more general, Choquet integrals have the
inconvenient of requiring a large number of parameters, making them not practical for
experiments involving learning. Indeed, if n is the number of data to be aggregated,
OWA and the weighted mean require only n additional values and WOWA requires
2 × n, whereas Choquet integrals require 2n values.
Weighted Means Based Filters for SAR Imagery 17

We now present a simple example to illustrate the use of OWA, WM and WOWA
operators. Let a1 = 10, a2 = 20, and a3 = 0. We thus have σ (1) = 2, σ (2) = 1, and
σ (3) = 3. Therefore aσ (1) = 20, aσ (2) = 10, and aσ (3) = 0.
Let w = [0.5 0.3 0.2] and p = [1/6 2/3 1/6]. We thus obtain
– f wowa (10, 20, 0) = 0.5 × 20 + 0.3 × 10 + 0.2 × 0 = 13,
– f pwm (10, 20, 0) = 1/6 × 10 + 4/6 × 20 + 1/6 × 0 = 15.
Let φl be a linear by parts function, given by

(x − xk−1 )(yk − yk−1 )

∀x ∈ (xk−1 , xk ], k ∈ {1, n}, φl (x) = yk−1 + , (8)
(xk − xk−1 )

where {(xi , yi )i=1,n } is a set of predetermined points. Using w to obtain the points
{(xi , yi )i=1,n } as required in Eq. 6, and then applying them in Eq. 8, we obtain φl (0) =
0, φl (1/3) = 0.5, φl (2/3) = 0.8 and φl (1) = 1.
Taking σ and p above, we have pσ (1) = 2/3, pσ (2) = 1/6, and pσ (3) = 1/6. We
then obtain Pσ (1) = 2/3, Pσ (2) = 5/6, and Pσ (3) = 1, as well as φl (Pσ (1)) = 0.8,
φl (Pσ (2)) = 0.9, and φl (Pσ (3)) = 1. Therefore ω1 = 0.8, ω2 = 0.1, and ω3 = 0.1.
Thus
wowa
– f w, p,φ (10, 20, 0) = 0.8 × 20 + 0.1 × 10 + 0.1 × 0 = 17.

3.2 WM, OWA and WOWA Filters

OWA (respec. WM) filters (see [28, 29]) are obtained by applying OWA (respec.
WM) weight vectors in the values inside a sliding window over a given image. Both
WM and OWA filters are convolution filters, where coefficients γ are positive and
add up to 1, with the coefficients in OWA being applied in the order statistic of the
data, which makes them OSFs [7].
In the present paper, we introduce WOWA filters, based on a combination of WM
wowa
and OWA filters. Given an image I , the application of a WOWA filter Fw, p,φ derives
wowa
a filtered image Iw, p,φ , as described below. OWA and WM filters Fw and F pwm can
owa

be described similarly, but in a simplified manner.

wowa
Procedure Fw, p,φ (I )

1. Transform a weight matrix M associated to a predefined neighbourhood, into a

vector with n positions p.
2. For each pixel in position (x, y) in the original image I , transform a window
I  around (x, y), according to the predefined neighbourhood, into a vector of n
positions a.
3. Using a, derive σ and aσ .
4. Using w, p and φ, derive weight vector ω.
wowa
5. Calculate f w, p,φ (a1 , ..., an ).
18 L. Torres et al.

6. Make the result become the value for position (x, y) in the filtered image:
wowa wowa
Iw, p,φ (x, y) = f w, p,φ (a1 , ..., an ).

Given an image I and a position (x, y) in I , and considering a 3 × 3 window,

in step 2 we derive vector a, with 9 positions, as (I (x − 1, y − 1), I (x − 1, y), I (x −
1, y + 1), I (x, y − 1), I (x, y), I (x, y + 1), I (x + 1, y − 1), I (x + 1, y), I (x + 1,
y + 1)). Vector p is derived in a similar way from matrix M in step 1.

4 Learning Weight-Based Filters for SAR Imagery

In the following, we describe our framework to learn the weights for OWA, WM
and WOWA filters for SAR imagery. We first describe Genetic Algorithms, the
optimization algorithm adopted in this work, and then how it is adapted to SAR
imagery specificities.

4.1 Genetic Algorithms

Genetic Algorithms (GA), first proposed in [9] (see also [6]), combine Mendel’s
ideas about the codification of life in genes, with Darwin’s ideas on the survival of
the fittest (natural selection). They are search algorithms that evolve populations of
candidate solutions, according to a fitness function that assesses the quality of these
solutions to solve the problem at hand.
A candidate solution c ∈ C = {c1 , . . . , ck } consists of a set of parameters to a
function sol, that models the problem at hand. Each c can be thought of as a genotype
(chromosome) and sol(c) as its corresponding phenotype. A fitness function f it
evaluates the candidate solutions; f it (sol(c)) should be proportional to the capacity
of c ∈ C in solving the problem at hand.
At each GA iteration, three processes (selection, crossover and mutation) take
place, generating a new population C  . The selection process is such that the fittest
candidates in C have a higher probability of being selected for reproduction. This
process is usually performed by means of a roulette (the larger the fitness of an
individual, the larger its share in the roulette wheel) or a set of tournaments (at each
tournament, a set of individuals are chosen at random from the population and the
winner is selected for reproduction). Different forms of elitism can also be used,
by forcing the best candidates to be remain in the new population and/or to have a
stronger influence on the creation of C  . The reproduction process, called crossover,
creates two new candidate solutions by mixing the genotypes of two selected parent
candidate solutions. In the mutation process, all new candidate solutions can suffer
changes, according to a (usually small) probability, called the mutation rate, here
denoted as ρ.
Weighted Means Based Filters for SAR Imagery 19

The first initial population is usually obtained at random, but in many applications,
the use of a selected set of chromosomes may lead to better results. The stop criterion
is usually a fixed number of iterations. The combination of selection, crossover and
mutation provide GAs with a good equilibrium between exploration and exploitation
of the search space.
In our work, each chromosome consists of weight vectors and each position in
the chromosome contains a real number (a weight), that sum up to 1 altogether.
The crossover operator adopted here consists in a linear combination of the parent
chromosomes values. Given two parents c1 and c2 , and random number α ∈ [0, 1],
we generate two sons c12 and c21 , where ∀i ∈ {1, ..., n}, c12 [i] = α × c1 [i] + (1 −
α) × c2 [i] and c21 [i] = (1 − α) × c1 [i] + α × c2 [i]. The values in c12 and c21 are
then normalized to guarantee that the weights sum up to 1.
In previous works, we tested a few strategies for mutation (see e.g. [28, 29]),
and in the following we describe the ones that gave the best results, called A and
B. In these strategies, the mutation rate is not applied on each position but on the
chromosome as a whole. If a chromosome is selected for mutation, we randomly
select the main position 1 ≤ q ≤ n in the chromosome to be changed, considering a
uniform distribution. Then these mutation strategies differ as follows:
– A: The value in position q is multiplied by mutation rate ρ. The difference is
divided by n − 1 and added to each of the remaining positions. Note that the
larger is ρ, the larger is the change of the value in position q and the smaller is the
change in the other positions in the vector.
– B: The value in q is increased with the value of its neighbour, considering the
chromosome as a ring, and the neighbour receives value 0. When the neighbour
to the right (respec. left) is considered, the strategy is named Br (respec. Bl).
In our previous works with OWA filters in SAR imagery, we verified that mutation
operator A usually outperformed its B counterparts.

4.2 Learning SAR Filters with GAs

Given a SAR image on a given area, we need to have the means to assess the quality
of filters applied on it. As done in previous works (see [28, 29]), here we adopt the
following framework:
– samples from classes of interest are taken from a SAR image,
– for each class, the parameters of its associated distribution are estimated,
– using a phantom image, in which the regions are associated to the classes in the
original image, a set of simulated images is created at random using the class
distributions,
– the set of simulated images is partitioned in two sets, one for training and one for
testing, and
20 L. Torres et al.

– the best weight vectors found by the GA on the training set is used on the test set
for evaluation.
Considering a m × m standard window, WM and OWA filters require chromo-
somes with n = m 2 positions to encode weight vectors p and w, respectively. As for
WOWA filters, when p and w are learned at the same time as here, each chromosome
has n = 2m 2 positions.
In [29], we investigated the use of two strategies to learn OWA weight vectors for
k polarizations, k > 1:
– S1: k GAs are run independently, one for each polarization, with weight vectors
containing n positions.
– S2: a single GA is run, with the weight vector containing k × n; when the weight
vector is learned, it is split in k parts, resulting in a weight vector of n positions
for each polarization.
The second strategy requires a more complex dealing with the GAs than the first one:
even though selection, mutation and crossover remain basically the same, we have
to ensure consistency, with each of the k parts of the chromosome adding up to 1.
In our works, we verified that S2 produced better results for OWA filters in terms of
quality of results and computational efficiency.

5 Experiments

We conducted a series of experiments based on images from an area in the Brazilian

Amazon region, using L-band with wavelengths of [30 cm, 1 m] and frequencies
of [1MHz, 2GHz] (see Fig. 1a). We took the parameters for Wishart distributions
estimated in [23] for these images, with samples from 8 classes (see Fig. 1b), and
used them on a fragment of a phantom described in [20] (see Fig. 1c), deriving 50
synthetic intensity images. Each simulated image has 240 × 240 pixels and was
generated with 1-look.
Note that, when using a 3 × 3 window, weight vectors p and w have 9 positions
each, and 25 positions with a 5 × 5 window. Here, we adopted strategy S2, i.e., weight
vectors for the 3 polarizations are learned at the same time. Therefore, WM and OWA
chromosomes have 27 and 75 positions for 3 × 3 and 5 × 5 windows, respectively.
Here, vectors p and w are learned at the same time for WOWA filters (called strategy
p+w in [30]). Therefore, the WOWA vectors have 54 and 150 positions for 3 × 3 and
5 × 5 windows, respectively.
We performed several experiments with different parametrizations. In particular,
we verified that populations with 18, 36 and 72 elements generate quite similar
results, and that the computational effort of using 72 elements is too large compared
to the small gains it produces over its counterparts in some experiments. The use
of different seeds also did not entail large differences in performance. However, we
verified that a large number of generations does have an impact on performance for
Weighted Means Based Filters for SAR Imagery 21

(a) SAR image fragment [23] (b) Samples (ROI) (c) Phantom fragment [20]

Fig. 1 SAR L-band false color composition, using HH (red), HV (green) and VV (blue) polariza-
tions of the area of interest, ROIs and phantom used in the experiments

all filters. Here we report the use of OWA, WM and WOWA filters, considering the
following parametrizations:
– selection type: roulette,
– number of generations: 30 and 120,
– population size: 36 elements,
– mutation rates: 0.2 and 0.8,
– seeds for random numbers: 2 and 271.
For each experiment, we performed a 5-fold cross-validation, using 40 images
for training and 10 for testing in each fold. The elements in the initial population in
each experiment were chosen at random. As fitness function for each fold in each
parametrization, we took the means of the quality of the resulting filtered images,
according to index NMSE (see Sect. 2.2).
Tables 1, 2 and 3 bring the results obtained for WM, OWA and WOWA filters,
respectively, with parameters by GAs, according to NMSE, considering 5 × 5 win-
dows. In the tables, we denoted the experiments as E/ξ /ψ, where ξ ∈ {30, 120} and
ψ ∈ {0.2, 0.8}. Therefore, an experiment run with 30 generations and mutation rate
0.2 is denoted by E/30/0.2. We mark in bold the best results obtained in each seed.
The best aggregated results between the two seeds are marked with an asterisk (“∗ ”).

Table 1 NMSE mean (truncated) using WM on 50 images, with GA filters learned using 5 fold
cross-validation, 5 × 5 windows, and mean NMSE of training images as fitness function
Seed 2 Seed 271
HH HV VV Agg HH HV VV Agg
E/30/.2 0.0547 0.0561 0.0570 0.0559 0.0540 0.0541 0.0555 0.0545
E/30/.8 0.0527 0.0535 0.0545 0.0536 0.0522 0.0532 0.0544 0.0533
E/120/.2 0.0499 0.0508 0.0518 0.0508∗ 0.0500 0.0508 0.0518 0.0509
E/120/.8 0.0504 0.0511 0.0523 0.0513 0.0503 0.0512 0.0520 0.0512
22 L. Torres et al.

Table 2 NMSE mean (truncated) using OWA on 50 images, with GA filters learned using 5 fold
cross-validation, 5 × 5 windows, 36 elements in each population and mean NMSE of training
images as fitness function
Seed 2 Seed 271
HH HV VV Agg HH HV VV Agg
E/30/.2 0.0515 0.0523 0.0536 0.0525 0.0518 0.0523 0.0537 0.0526
E/30/.8 0.0516 0.0523 0.0536 0.0525 0.0517 0.0524 0.0536 0.0526
E/120/.2 0.0486 0.0494 0.0505 0.0495∗ 0.0486 0.0494 0.0506 0.0495∗
E/120/.8 0.0486 0.0495 0.0505 0.0495∗ 0.0486 0.0494 0.0506 0.0495∗

Table 3 NMSE mean (truncated) using WOWA on 50 images, with GA filters learned using 5 fold
cross-validation, 5 × 5 windows, and mean NMSE of training images as fitness function
Seed 2 Seed 271
HH HV VV Agg HH HV VV Agg
E/30/.2 0.0515 0.0524 0.0538 0.0526 0.0515 0.0522 0.0536 0.0524
E/30/.8 0.0515 0.0523 0.0536 0.0525 0.0515 0.0523 0.0535 0.0524
E/120/.2 0.0486 0.0511 0.0512 0.0503 0.0493 0.0503 0.0506 0.0501∗
E/120/.8 0.0499 0.0497 0.0511 0.0502 0.0494 0.0502 0.0513 0.0503

In Table 1, we see that the mean NMSE of all results range between 0.0499 and
0.0570. The best aggregated result was obtained using 120 generations, seed 2 and
mutation rate 0.2. We see in this table is that a large number of generations indeed
increases the performance of WM filters.
In Table 2, we see that the mean NMSE of all results are more similar in the various
parametrizations than in the case of WM, ranging between 0.0486 and 0.0537. The
best aggregated result was obtained using 120 generations, seeds 2 and 271 and
mutation rates 0.2 and 0.8. We see that also here, the most important factor on the
results is the number of generations, with the seeds and mutation rates producing
basically the same results, considering the individual polarizations as well as the
aggregated results.
In Table 3, we see that the mean NMSE of all results are very close to those
obtained by OWA filters, ranging between 0.0486 and 0.0538. The best aggregated
result was obtained using 120 generations, seed 271 and mutation rate 0.2. Again
here, the large number of generations produced the best results.
Comparing together the results of Tables 1, 2 and 3, we see that WM filters have
been outperformed by both OWA and WOWA filters in all experiments. The results of
OWA and WOWA are very similar; WOWA filters produced better results than OWA
with a smaller number of generations, however, with a large number of generations,
OWA produces the best overall results. In a nutshell, WOWA outperformed WM in
our experiments but did not outperform OWA with our choice for function φ.
Weighted Means Based Filters for SAR Imagery 23

Table 4 brings the results obtained by the best aggregated GA-learned filters,
considering the same type of window and the same number of folds. In Table 4, we
also report the results for SDNLM and R-Lee filters, with the best parametrizations
chosen after a few experiments, using 5 × 5 filtering window for both filters, with
3 × 3 patches, and significance level of 5% for for SDNLM, and with ENL = 1 for
R-Lee. The best WM, OWA and WOWA filters for 5 × 5 windows are taken from
Tables 1, 2 and 3, with ties solved considering non displayed decimal cases. For both
5 × 5 and 3 × 3 windows, the number of elements in the population is 36, and the
best number of generations and best mutation rate are 120 and 0.2, respectively.
Considering both types of windows, the seed value for OWA and WM is 2, whereas
for WOWA the seed value is 271.
In Table 4, we see that the WOWA, OWA and WM filters, with weight vectors
learned with a GA with 5 × 5 windows, outperformed all other filters considered
here with respect to NMSE (used by the fitness function). In particular, OWA filters
produced the best results, both in the individual polarizations as well as when they
are aggregated. In what regards SSIM, the best performances were obtained with
SDNLM and the mean filter. In the remaining of this section, we only discuss NMSE
results.

Table 4 NMSE and SSIM mean (truncated) on 50 images, with GA filters learned using 5 fold
cross-validation, 5 × 5 windows, and mean NMSE of training images as fitness function
NMSE SSIM
HH HV VV Agg HH HV VV Agg
Unfiltered 0.997 0.074 1.025 1.008 0.068 0.084 0.053 0.068
SDNLM 0.078 0.074 0.119 0.091 0.166 0.177 0.126 0.157
(5×5)
R-Lee (5×5) 0.079 0.075 0.125 0.093 0.155 0.167 0.124 0.149
OWA (5×5) 0.048 0.049 0.050 0.049 0.147 0.158 0.147 0.151
WOWA 0.049 0.050 0.050 0.050 0.146 0.157 0.147 0.150
(5×5)
WM (5×5) 0.049 0.050 0.051 0.050 0.148 0.159 0.148 0.152
Mean (5×5) 0.064 0.063 0.124 0.084 0.166 0.178 0.124 0.156
Median (5×5) 0.153 0.153 0.203 0.170 0.123 0.131 0.099 0.118
Min (5×5) 0.927 0.928 0.934 0.930 0.001 0.001 0.000 0.000
Max (5×5) 9.754 9.868 9.590 9.737 0.024 0.026 0.019 0.023
OWA (3×3) 0.106 0.107 0.108 0.107 0.133 0.145 0.130 0.136
WOWA 0.111 0.110 0.110 0.110 0.132 0.145 0.130 0.135
(3×3)
WM (3×3) 0.117 0.118 0.118 0.117 0.133 0.146 0.130 0.136
Mean (3×3) 0.115 0.116 0.116 0.116 0.137 0.150 0.133 0.140
Median (3×3) 0.189 0.190 0.190 0.190 0.107 0.116 0.104 0.109
Min (3×3) 0.807 0.807 0.808 0.807 0.007 0.007 0.007 0.007
Max (3×3) 4.969 5.025 4.971 4.988 0.039 0.046 0.037 0.040
24 L. Torres et al.

In [28], we have verified that filters learned with 3 × 3 windows did not fare so
well when compared with the model-based filters SDNLM and R-Lee, which use
5 × 5 windows, considering a GA configuration of 30 generations at most. That
indicated that 5 × 5 windows are more appropriate 3 × 3 ones for our filters. Here,
we see that with 120 generations, the model-based filters still outperform our filters
with 3 × 3 windows but are outperformed when 5 × 5 windows are used. We see
that even a very large increase in the number of generations in the GA is not capable
of overcoming the negative impact that small windows cause on our filters.
Figure 2 brings an unfiltered synthetic image and the filtered images obtained from
it using some of the methods considered here. We note that WM, OWA and WOWA

(a) Unfiltered (0.9991, 0.0699) (b) OWA (0.0525, 0.1626) (c) WOWA (0.0526, 0.1622)

(d) WM (0.0533, 0.1637) (e) Mean (0.0863, 0.1504) (f) SDNLM (0.0912, 0.1579)

(g) R-Lee (0.0953, 0.1509) (h) Min (0.9270, 0.0010) (i) Max (9.6841, 0.0234)

Fig. 2 False color composition of results from the same simulated images, considering methods
using 5 × 5 windows, for HH (red), HV (green) and VV (blue) polarizations, with mean NMSE
and SSIM from the polarizations inside parentheses
Weighted Means Based Filters for SAR Imagery 25

in Fig. 2 are visually superior to those obtained with both complex filters; SDNLM
produces a blurred image and the Lee filter yields a pixelated image.
Figure 3 brings a square fragment of the false composition of the original image,
with the ROIs from which the samples for each class were extracted (Fig. 3a),1 the
images obtained with OWA, WOWA and WM filters (Fig. 3b–d), all using as weights
the best parametrization learned in fold 3, and the model-independent SDNLM and
R-Lee filters (Fig. 3e, f, resp.).
Figure 3b, c, d show the effect of the OWA, WOWA and WM on windows of size
5 × 5 over the whole image. Albeit the noise reduction is evident, it is also clear that
the blurring introduced eliminates useful information as, for instance, curvilinear
details in the river area.
Figure 3f is the result of applying the R-Lee filter; although presenting a good
performance, some details in the edges are eliminated, being worse than all the
previous filters in this aspect. Figure 3e present the result of smoothing the original
data set with the SDNLM filter at the level significance α = 10%. The noise effect is
alleviated, e.g. the graininess is reduced specially in the planting and over the river
areas, but fine details are more preserved than when the R-Lee filter is employed.
Another important point is that land structures within the river are enhanced, and
their appearance is maintained.
Figures 4 and 5 respectively illustrate the weight vectors w and p found in our
best experiments. We see that the weights found are not distributed homogeneously
(specially in what regards VV), but the the highest values for HH and HV are around
the median. The orness found for vectors w for HH, HV and VV are respectively
0.494, 0.496 and 0.511, which is coherent with the visual inspection. The weights
in vectors p, all polarizations considered, varied between 0.009 and 0.067. We see
that HH and HV produce a more homogeneous distribution of weights than VV.
Although there are large weight values in the borders, we see that the central pixel in
all polarizations have a high weight value. We also conducted a set of experiments
(not shown here) in which we learned a smaller number of weights that were repeated,
according to the relative positions to the central pixel for p, and to the median for w,
but the results were inferior to those obtained here.
The GA was run on a machine with the following specifications: Intel i7, CPU 2.60
GHz, RAM with 16 GB, Windows 10, Fortran with Force 2.0 compiler. Considering 5
folds, with 10 images in each fold, 36 elements in the population, and 30 generations,
the GA processing for all our filters (WM, OWA and WOWA) took approximately
30 min for each polarization for 3 × 3 windows and approximately 1 h for 5 × 5
windows. When the number of generations is multiplied by 4 (120), the computation
times also increase 4 times, in all cases.

1 The available image with ROIs use the Pauli composition, that takes the complex part of the images

into account. For this reason, the colours are different from the filtered images, that use the same
composition used in Fig. 1a, which only takes the intensity images into account.
26 L. Torres et al.

(a)Unfiltered(ROI) (b)OWA

(c)WOWA (d)WM

(e)SDNLM (f)R-Lee
Fig. 3 False color composition of the original SAR image unfiltered, with ROIs, and images
obtained with a set of filters
Weighted Means Based Filters for SAR Imagery 27

Fig. 4 Weight vectors w found in the best experiment

(a) HH (b) HV (c) VV

Fig. 5 Weight vectors p found in the best experiment (values multiplied by 102 with a single digit
displayed)

6 Conclusions and Future Work

In this paper, we addressed OWA, WM and WOWA filters, based on OWA, WM and
WOWA operators (see [25, 36]), respectively. We compared their ability to reduce
speckle in SAR imagery, considering simulated images from three polarizations (HH,
HV and VV) and two types of windows (3 × 3 and 5 × 5). The weight vectors of
the three filters were learned with Genetic Algorithms, for both a small and a large
number of generations.
Our experiments have shown that the number of generations used in the G.A. has
a strong impact on the quality of the results. They have also shown that the size of
the windows is however the most crucial factor for obtaining good results.
We compared the results against a set of filters from the literature, including
model-independent filters WM and OWA (which also had their parameters learned)
and two model-dependent ones, SDNLM [27], a recent parametrized family of filters
(here with, parameters chosen after trial-and-error), and the well-known Refined Lee
filter [12]. OWA filters, followed by WOWA and WM filters, with parameters learned
using 5 × 5 windows outperformed all other filters, according to indice NMSE, used
in the learning process.
The quality of filters depend to what one wants to extract from the filtered images.
Visual inspection indicates that the sole use of one index to assess the quality of
filters may not be enough for some applications. We are currently investigating the
use multi-objective optimization to produce filters that take into account not only the
28 L. Torres et al.

error, measured by NMSE, but the preservation of the structures in the image, using
a measure based on edge-detection.
This work and previous ones show that automatic learning of parameters can be
crucial for parametrized families of filters. For instance, the filters obtained using the
min, max, median and arithmetic mean operators are all particular cases of the OWA
and WOWA operators and fared significantly worse than their learned counterparts.
To be fair, the results obtained by OWA and WOWA filters should also be compared
to model-dependent filters whose parametrizations are also learned, but this is out of
the scope of this work.
As future research, we intend to use other alternatives to model function φ, used
by the WOWA operators, instead of the linear by parts function adopted here. We
also intend to study the influence of initial populations in the GA, and the use of
other types of windows, such as the ones based on non-standard neighbourhoods
(see, e.g. [5]). Last but not least, we intend to use the whole information contained in
a POLSAR image, taking into account also the complex components, and not only
the real components, as we have done so far.

Acknowledgements The authors are indebted to Vicenç Torra and to Lluis Godo for discussions
on WOWA operators, and to Benicio Carvalho for providing computational resources.

References

1. Baxter, R., Seibert, M.: Synthetic aperture radar image coding. MIT Lincoln Lab. J. 11(2),
121–158 (1998)
2. Bovik, A.C., Huang, T.S., Munson, D.C.: A generalization of median filtering using linear
combinations of order statistics. IEEE Trans. ASSP 31(6), 1342–1349 (2005)
3. Buades, A., Coll, B., Morel, J.M.: A review of image denoising algorithms, with a new one.
Multiscale Model. Simul. 4(2), 490–530 (2005)
4. Carlsson, C., Fullér, R.: Fuzzy reasoning in decision making and optimization. In: Studies in
Fuzziness and Soft Computing Series. Springer (2002)
5. Fu, X., You, H., Fu, K.: A statistical approach to detect edges in SAR images based on square
successive difference of averages. IEEE GRSL 9(6), 1094–1098 (2012)
6. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-
Wesley (1989)
7. Grabisch, M., Schmitt, M.: Mathematical morphology, order filters and fuzzy logic. In: Pro-
ceedings of FuzzIEEE 1995, Yokoham, pp. 2103–2108 vol. 4 (1995)
8. Herrera, F.: Genetic fuzzy systems: status, critical considerations and future directions. Int. J.
Comput. Intell. Res. 1(1), 59–67 (2005)
9. Holland, J.H.: Adaptation in Natural and Artificial Systems. University of Michigan Press,
USA (1975)
10. Lee, J.-S., Grunes, M.R., de Grandi, G.: Polarimetric SAR speckle filtering and its implication
for classification. IEEE Trans. GRS 37(5), 2363–2373 (1999)
11. Lee, J.-S., Grunes, M.R., Mango, S.A.: Speckle reduction in multipolarization, multifrequency
SAR imagery. IEEE Trans. GRS 29(4), 535–544 (1991)
12. Lee, J.-S., Grunes, M.R., Schuler, D.L., Pottier, E., Ferro-Famil, L.: Scattering-model-based
speckle filtering of polarimetric SAR data. IEEE Trans. GRS 44(1), 176–187 (2006)
13. Lee, J.-S., Pottier, E.: Polarimetric radar imaging: from basics to applications. In: Optical
Science and Engineering. Taylor and Francis, UK (2009)
Weighted Means Based Filters for SAR Imagery 29

14. Mascarenhas, N.: An overview of speckle noise filtering in sar images. In: 1st Latin-American
Seminar on Radar Remote Sensing—Image Processing Techniques, pp. 71–79 (1997)
15. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial
domain. IEEE Trans. Image Process. 21(12), 4695–4708 (2012)
16. Moschetti, E., Palacio, M.G., Picco, M., Bustos, O.H., Frery, A.C.: On the use of Lee’s protocol
for speckle-reducing techniques. Lat. Am. Appl. Res. 36(2), 115–121 (2006)
17. Mott, H.: Remote Sensing with Polarimetric Radar. Wiley, USA (2006)
18. Pitas, I., Venetsanopoulos, A.N.: Nonlinear Digital Filters: Principles and Applications.
Springer, Berlin (2013)
19. Richards, J.A.: Remote Sens. Imaging Radar. Springer, Signals and Communication Technol-
ogy Series (2009)
20. Saldanha, M.F.S.: Um segmentador multinível para imagens SAR polarimétricas baseado na
distribuição Wishart. Ph.D. Thesis, INPE, Brazil (2013)
21. SantAnna, S.J.S., Mascarenhas, N.: Comparação do desempenho de filtros redutores de
“speckle” . In: VIII SBSR, pp. 871–877 (1996)
22. Sheng, Y., Xia, Z.G.: A comprehensive evaluation of filters for radar speckle suppression. In:
IGARSS ’96, vol. 3, pp. 1559 – 1561 (1996)
23. Silva, W.B., Freitas, C.C., Sant’Anna, S.J.S., Frery, A.C.: Classification of segments in PolSAR
imagery by minimum stochastic distances between Wishart distributions. IEEE J-STARS 6(3),
1263–1273 (2013)
24. Torra, V.: The weighted OWA operator. Int. J. Intell. Syst. 12(2), 153–166 (1997)
25. Torra, V.: On some relationships between the WOWA operator and the Choquet integral. In:
Proceedings of IPMU’98, Paris, France (1998)
26. Torra, V., Narukawa, Y.: Modeling Decisions: Information Fusion and Aggregation Operators.
Springer, Berlin (2007)
27. Torres, L., Sant’Anna, S.J.S., Freitas, C.C., Frery, A.C.: Speckle reduction in polarimetric SAR
imagery with stochastic distances and nonlocal means. Pattern Recogn. 47(1), 141–157 (2014)
28. Torres, L., Becceneri, J.C., Freitas, C.C., Sant’Anna, S.J.S., Sandri, S.: OWA filters for SAR
imagery. In: Proceedings of LA-CCI’15, Curitiba, Brazil, pp. 1–6 (2015)
29. Torres, L., Becceneri, J.C., Freitas, C.C., Sant’Anna, S.J.S., Sandri, S.: Learning OWA filters
parameters for SAR imagery with multiple polarizations. In: Yang, X.-S., Papa, J.P. (eds.) Bio-
Inspired Computation and Applications in Image Processing, pp. 269–284. Elsevier, Nether-
lands (2016)
30. Torres, L., Becceneri, J.C., Freitas, C.C., Sant’Anna, S.J.S., Sandri, S.: WOWA image filters.
In: Proceedings of CBSF’16, Campinas, Brazil (2016)
31. Torres, L., Becceneri, J.C., Freitas, C.C., Sant’Anna, S.J.S., Sandri, S.: Comparing OWA and
WOWA filters in mean SAR images. In: Proceedings of SBSR’17, S.J.Campos, Brazil (2017)
32. Ulaby, F.T., Elachi, C.: Radar Polarimetry for Geoscience Applications. Artech House, USA
(1990)
33. Wang, Z., Bovik, A.C.: A universal image quality index. IEEE SPL 9(3), 81–84 (2002)
34. Wang, Z., Bovik, A.C., Lu, L.: Why is image quality assessment so difficult? In: IEEE ICASSP,
vol. 4, pp. 3313–3316, Orlando (2002)
35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error
visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600–612 (2004)
36. Yager, R.R.: On ordered weighted averaging aggregation operators in multi-criteria decision
making. IEEE Trans. Syst. Man Cybern. 18, 183–190 (1988)
On Combination of Wavelet
Transformation and Stabilized KH
Interpolation for Fuzzy Inferences Based
on High Dimensional Sampled Functions

Ferenc Lilik, Szilvia Nagy and László T. Kóczy

Abstract A new approach for inference based on treating sampled functions is

presented. Sampled functions can be transformed into only a few points by wavelet
analysis, thus the complete function is represented by these several discrete points.
The finiteness of the teaching samples and the resulting sparse rule bases can be
handled by fuzzy rule interpolation methods, like KH interpolation. Using SHDSL
transmission performance prediction as an example, the simplification of inference
problems based on large, sampled vectors by wavelet transformation and fuzzy rule
interpolation applied on these vectors are introduced in this paper.

Keywords Fuzzy inference · Performance prediction · Fuzzy rule interpolation

Wavelet analysis

1 Introduction

Due to the great number of input values, making inference on phenomena which can
be described by large-sized vectors is difficult and expensive. In order to construct
efficient inference systems, simplification of the input space is needed. This sim-
plification makes the process of the inference easier, however, it unavoidably rises
the system’s level of uncertainty and inaccuracy. During our previous research on
performance prediction of physical links of telecommunications access networks,
we had to encounter such problems in two ways.
Due to the limited calculation capacity not all the measured values can be used as
the bases of the inference. Drastically lowering the number of the measured frequency
dependent input values caused an inaccuracy in the final results. Later, this type of
sparseness will be referred to as vertical sparseness.

F. Lilik (B) · S. Nagy · L. T. Kóczy

Széchenyi István University, Győr 9026, Hungary
e-mail: [email protected]
L. T. Kóczy
Budapest University of Technology and Economics, Budapest 1117, Hungary

© Springer International Publishing AG 2018 31

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_3
32 F. Lilik et al.

As it is not possible to measure all possible data, the finite teaching sample set
will naturally result in sparse rule bases: there will be points, which will be outside
of all the supports of the antecedent sets of the corresponding dimensions. Later, this
behaviour will be referred to as horizontal sparseness.
In Sect. 2 the primary technical problem underlying the research on performance
prediction is briefly reviewed, the first version of the inference method, its test results
and the horizontal and vertical sparseness derived from the simplification of the input
space are also shown. In Sect. 3 wavelet transformation and fuzzy rule interpolation
as the algorithmic techniques applied in a combined way for handling the problems
of simplification are described, and in Sect. 4 we present the test results of the new
approach based on these techniques.

2 Fuzzy Performance Prediction of Physical Links

in Telecommunications Access Networks of SHDSL
Transmission

In certain cases, the technological demands of modern telecommunications services

can be hardly, or can not be fulfilled by the existing telecommunication network-
s. This is especially true for mobile or copper wire pairs based access networks.
Moreover, the performance possibilities of the individual links of these networks can
be totally different from one another. Therefore, lately, telecommunications service
providers apply various, usually rather resource demanding methods for predict-
ing the performance of the individual physical links. The aim of our research was
constructing a simple, but acceptably precise performance predicting method.
In this work, as an example, we have used symmetrical copper wire based access
networks with SHDSL connections. SHDSL [1] is a symmetrical digital telecom-
munications transmission method, which establishes high speed data transmission
over symmetrical wire pairs. Although the newest members of the SHDSL systems
can reach higher data transmission speed (bit rate) than the one studied in the present
contribution, our approach can be successfully applied also for those systems.
In this context, performance is a property of the physical link, therefore it has to be
predicted from the measurable physical parameters of the link. Based on the related
ITU-T recommendations [1, 2] and on our own measurements and investigations,
SHDSL performance can be predicted from the frequency dependent insertion loss
of the studied line [3]. Insertion loss is theoretically a continuous function in the
lowest 1.5 MHz wide frequency domain, where the SHDSL transmission is realized.
For technical reasons, only discrete values were measured from 10 kHz in 10 kHz
wide steps, thus resulting in insertion loss values of the studied wire pairs at 150
discrete frequency points. The available maximal data transfer rate of the lines were
also measured by installing SHDSL equipments to the bare lines. The measured
On Combination of Wavelet Transformation and Stabilized … 33

insertion loss series were clustered into 5 groups according to the measured bit rates.
This grouping is similar to the practice of telecommunication service providers when
offering packages of DSL services.
Based on insertion loss values at 6 well-selected characteristic frequencies, fuzzy
rule bases were created to predict the maximal available SHDSL bit rate of sym-
metrical wire pairs. The result of this prediction is the label of one of the bit rate
groups. Two types of rule bases were created. One of them was constructed directly
from the measured values. It consists of five six dimensional rules, in which trian-
gular antecedent and consequent fuzzy sets are used. The input dimensions are the
insertion loss values that can be measured at the 6 characteristic frequencies. Each
rule can be unambiguously assigned to one of the output states. Using the measured
and clustered data as teaching samples, another type of rule bases were created by
bacterial evolutionary algorithm [4], resulting in a rule base with trapezoidal fuzzy
sets and ten rules. This rule base has the same 6 dimensions: the 6 characteristic
frequencies [5]. Examples of the rule antecedents from each type of rule base can be
seen in Fig. 1. In the figure, the upper diagram belongs to the triangular rule base, and
the lower one to the rule base constructed by bacterial evolutionary algorithm. The
upper rule is obviously sparse. Even though seemingly there are no gaps between the
trapezoidal fuzzy sets in the rule base made by the bacterial evolutionary algorithm,
this rule base can be also considered as a sparse or incomplete one, because insertion
loss values can be measured outside of the supports of all fuzzy sets, as well.
The above two rule bases were tested by the measurements of more than 60 wire
pairs in operating access networks and there were no relevant differences between
their respective results. In most of the cases, where all measured values belonged
to insertion loss areas covered by antecedent sets, the predictions were successful.
Only 13 lines out of 65 could be evaluated, and the predictions were correct in case
of 12 lines form this 13.

Fig. 1 Examples of rule 1

D1
antecedents from our D2
D3
previous predicting methods D4
0,5 D5


[6] D6

0
0 20 40 60 80
Insertion Loss [dB]
1

0,5


0
0 20 40 60 80
Insertion Loss [dB]
34 F. Lilik et al.

Fig. 2 Success rate of the

correct predictions
rule bases failed predictions
invalid predictions
18,46%

1,538%

80%

However correct the results of the successful predictions were, the high proportion
(80%) of the lines (Fig. 2) where no results were produced due to the gaps in the rule
base (no overlap with any antecedent), is not acceptable in practice. This phenomenon
was caused by two reasons.
The reason for unsuccessful evaluations, where no valid results were produced,
was the insufficient knowledge of the observed physical phenomenon. Obtaining
only a limited amount of valid data during the measurement, only sparse rule bases
could be constructed. Measurement results of wire pairs to be tested contained values
also outside of the supports of the antecedent fuzzy sets, hence, in these cases there
could be no valid conclusion calculated. This deficiency of the method is considered
as a “horizontal sparseness”.
The reason for valid, but incorrect prediction was investigated as well. This was
caused by the drastically lowered dimensionality of the input space (values only at
6 points out of 150 were considered). Even though, generally, the insertion loss
measured at the selected characteristic frequencies led to correct predictions, in
several, infrequently occurring cases, measured values at these 6 frequency points
could have such deviations which could bias the result of the prediction. Using
only several variables from the whole input space can be considered as “vertical
sparseness” of the model of the reality.
In the next section a new approach will be proposed, suitable for the elimination
of both mentioned problems.

3 Methods for Handling the Vertical and Horizontal

Sparsenesses

Vertical sparseness of the rule bases was derived from the partial usage of the possible
input data. It was needed in order to decrease the dimensionality of the applied fuzzy
inference system, however, a large amount of information of the measured insertion
loss functions was wasted. Finding a method which keeps the simplicity of the fuzzy
On Combination of Wavelet Transformation and Stabilized … 35

system and the information of the used insertion loss functions was needed. As
wavelet transformation is efficient in reducing the size of any continuous or discrete
functions down to a required level, it seemed to be successfully applicable in the
problem.
Horizontal sparseness of the fuzzy system, namely the sparseness of the rule
bases, can be handled by the techniques of fuzzy rule interpolation. Stabilized KH
interpolation fits continuous and mathematically stable functions to all α-cuts of the
membership functions in the rules, which can treat the observations in the gaps and
out of the domains of the rules, too (in this way performing also extrapolation).
Basics of wavelet transformation and stabilized KH interpolation are briefly
overviewed in the followings.

3.1 On Wavelet Analysis

In data processing in general wavelet theory [7, 8] has proved to be an extremely

useful tool. The largest part of the methods using wavelets is the image [9] and video
compression [10–12]. The first large-scale application was a fingerprint compression
method [13], and also the compressors used in space equipments are mostly based
on wavelet analysis [14], but one of the pioneering application was seismic signal
analysis [15], which is a 1D data set similar to insertion loss values in many ways.
Wavelet analysis from our point of view is similar to the discretized versions of
Fourier analysis. In the next paragraphs their similarities and the differences will be
summarized.
Both continuous wavelet transform and Fourier analysis [16] of a function pro-
vides data about the function’s fine-scale and rough-scale behavior. Similarly, the
discrete transforms turn the initial data vector or distribution into a set of coeffi-
cients corresponding to higher frequencies and lower frequencies. The well-known
windowed Fourier transform of a function f can be summarized by the formula
 ∞
Fb { f }(ω) = w(x − b) f (x) e−iωx d x, (1)
−∞

with the window function w, which usually is either compactly supported or has
very quickly decaying tails towards the positive and negative infinity. This provides
a short sample of the function f to be transformed. However, if the window function
is combined with the transforming exponential, a short wave

wb,ω (x) = eiωx w(x − b),

is produced as a transforming function, and the windowed Fourier transform turns

into  ∞
Fb { f }(ω) = wb,ω (x) f (x)d x. (2)
−∞
36 F. Lilik et al.

Wavelet transform is summarized as

 ∞
Wψ { f }(ba) = |a|−1/2 ψb,a (x) f (x)d x,
−∞

with the transforming function

 
−a/2 x −b
ψba (x) = |2| ψ .
2a

It can be seen from the above formula, that all the wavelets ψb,a (x) are generated
from one mother wavelet ψ(x) by dilating and shifting, thus whereas in case of
the windowed Fourier transform, the grid distance and the size of the transforming
function remains the same at all resolutions, only its shape changes, for wavelet
transforms, the shape of the functions remains the same and both the grid and the
width of the function shrinks as the resolution (i.e., the frequency or spatial frequency)
increases. This property is one of the driving forces behind the success of wavelet
analysis, as usually the high frequency terms are localized in smaller spatial domains
than the slowly varying parts, thus changing the window size with the frequency is
usually very effective.
Wavelet analysis can be carried out by a series of filter pairs. There is a high-pass
and a low-pass filter in all of the pairs, as it can be seen in Fig. 3, the high-pass ones
(after a downsampling) giving the wavelet components and the low-pass ones being
transformed further. The downsampling steps in each of the branches in Fig. 3 take
only every second element of the result and neglect the points in between. In each of
such steps the frequency limit of the low-pass distribution is at around the half of the
highest frequency of the incoming distribution. It can be seen easily, that low pass
outputs give a coarse-grained, or averaged behaviour of the distribution, whereas the
wavelet terms provide the fine-scale details.
The total number of the elements in the resulting vectors ci and di is almost the
same as that of the original vector ci , only a slight increase might arise due to the
size of the filters. The size of the filters Ns can be different for the different wavelet
types, however it is typically less than 20—in image processing usually less than

Fig. 3 One filter pair of the discrete wavelet transform. After the high pass and low pass convolu-
tional filters and the downsamplings the transformed vectors ci and di arise, their size is about half
of the size of the original ci
On Combination of Wavelet Transformation and Stabilized … 37

8, so this increment of Ns compared to the size of the data is usually negligible.

Moreover, usually many of the vector components di can be omitted as the fine
details are localized to small fractions of the data, thus the resulting vectors are
usually significantly (sometimes by one or two orders of magnitudes) smaller than
the size of the original data, hence the data compression ability [9, 14, 17].
Mathematically, the wavelet transforms of a function can be written as
 ∞
cba = f (x)φba (x). (3)
−∞

for the low-pass coefficients of resolution level 2−a , and as

 ∞
dba = f (x)ψba (x). (4)
−∞

for the high-pass coefficients. The transforming function φba is a so called scaling
function, and ψba a wavelet, as previously. From these coefficients the function f
can be approximated at resolution level 2−A as

∞
  ∞
A−1 
f A (x) = cb0 φb0 (x) + gba φba (x), (5)
−∞ a=0 −∞

One step of this synthesis procedure of the wavelet transform is summarized in Fig. 4,
where the synthesis filters are after an upsampling step that introduces zeros between
the elements of the incoming vectors. Usually, more of these steps are following each
other with introducing finer and finer resolution details.
In data analysis—also in our case—the starting point is a sampled function and
the end result is the lowest resolution level low pass vector and the high pass vectors.
Our starting vector is a series of insertion loss values measured at consecutive fre-
quency points, and the resulting vectors will give information about the large-scale
behavior of the insertion loss vs. frequency function. In the following considerations

Fig. 4 Synthesis steps of the wavelet transform. The synthesis filters are plotted with yellow
rectangles and the upsampling step with a rectangle including an upright arrow
38 F. Lilik et al.

Daubechies’s [7] wavelet and scaling function sets are used with 2 and 4 nonzero
filter coefficients. In case of the 2-element filter, the averaging process is without
overlaps with the neighboring frequency domains, whereas, in case of the 4-element
filter, the domains of the weighted averages overlap.
Transformations of the starting sampled insertion loss functions were carried out
until only 10 and 5 vector elements remained. As the power spectral density of the
transmission is larger in the lower frequency domain, we have merged the results of
the two vectors so that the points would meet our previously selected characteristic
frequency points.

3.2 Stabilized KH Rule Interpolation

In case of sparse rule bases, KH interpolation [18, 19] is a mathematically stable

and widely applicable fuzzy rule interpolation method. Its improved version is the
stabilized KH interpolation. In our work we used this improved technique in order
to eliminate the problems originating from the sparseness of the rule bases.
The method is based on the distances between the examination vector and the an-
tecedent sets of the rule base. The closures of the α-cuts of the interpolated resolution
are given by
2n 
 k
1
inf{Biα }
dαL (A∗ , Ai )
inf{Bα∗ } =
i=1
2n  k (6)
 1
i=1
dαL (A∗ , Ai )

and
2n 
 k
1
sup{Biα }
dαU (A∗ , Ai )
sup{Bα∗ } =
i=1
2n  k , (7)
 1
i=1
dαU (A∗ , Ai )

where i denotes the number of the rules, k the number of the dimensions (variables),
A∗ the observation, Ai the antecedent sets in rule i, dαL (A∗ , Ai ) and dαU (A∗ , Ai ) the
lower and upper bounds of the distance between the α-cuts of observation and the
antecedents, and B ∗ stands for the corresponding fuzzy conclusion [20]. In practice
it is efficient to calculate the values of Bα∗ for α = 0 and 1.
On Combination of Wavelet Transformation and Stabilized … 39

4 A New Performance Prediction Method Based on the

Combination of Wavelet Transformation and Stabilized
KH Rule Interpolation

In order to avoid the problems reviewed in Sect. 1, the techniques of Sects. 3.1 and
3.2 were used.
First, the wavelet transformed version of the insertion loss values used in rule base
construction were calculated. Daubechies-2 (Haar) [21] and Daubechies-4 wavelets
were used and the transformations were performed down to 5 points resolution as
described in the previous section. Figure 5 shows the original and the Haar wavelet
transformed insertion loss values as an example. As a matter of course, wavelet
transformation results in discrete values, however, to make the corresponding points
visible, they are graphically linked in the figure.
New rule bases were created by the Wavelet transformed insertion loss series. In
this pattern, the rule base based on Daubechies wavelets did not give better results
than the old one without any wavelet transformation, moreover, several additional
errors were detected. On the contrary, in case of the rule base made by Haar wavelets,
accurate results were gained for each of the 13 line that produce valid results and
one further line could be assessed, too, as it can be seen in Fig. 6.
In order to evaluate those lines that were previously not to be assessed, this new
(Haar wavelets based) rule base was applied together with the stabilized KH rule in-
terpolation. The 65 test lines were re-processed, thus the predictions became feasible
in the case of all lines. The predictions for the 13 wire pairs which were correctly

Fig. 5 Insertion loss values and the corresponding wavelet transforms. Different performance
classes are indicated by different colors
40 F. Lilik et al.

Fig. 6 Efficiency of the rule correct predictions

base based on Haar wavelets invalid predictions

21,54%

78,46%

Fig. 7 Efficiency of the correct predictions

Haar wavelets based rule acceptable predictions
base supplemented with the
stabilized KH rule
29,23%
interpolation

70,77%

evaluated previously remained valid, moreover, results of the predictions of 33 from

the other 52 were correct, and 19 acceptable (in this contribution, results with a devi-
ation of −1 from the correct values are considered as acceptable ones, all the others
as incorrect) and there were no incorrect results see Fig. 7.
The simplified “algorithm” of the construction of the predicting system is as
follows.

– Collection of insertion loss and bit rate data of wire pairs.

– Dividing the whole bit rate domain into groups (the more the number of the mea-
sured lines, the finer is the possible resolution) and clustering the measured values
into these groups.
– Generation of several discrete values (6 in this case, however, other resolutions are
examined by our ongoing investigations) from measured insertion loss functions
by wavelet transformation (Haar wavelets are now recommended, though investi-
gating other types of wavelets with other resolution levels are being in progress).
– Construction of fuzzy rule bases by clustered and wavelet transformed values.
– Wavelet transformation of the insertion loss function of the wire pair to be pre-
dicted.
– Prediction making by stabilized KH interpolation (can be made even if the input
values can be found within the areas covered by antecedent fuzzy sets).
On Combination of Wavelet Transformation and Stabilized … 41

5 Conclusions

A novel performance prediction method based on interpolated fuzzy inference for

telecommunications transmission lines and wavelet transformation of the values of
the physical parameters influencing the performance was presented. The combination
of the fuzzy rule interpolation and wavelet transformation was proposed in this paper
in the first time. Wavelet transform was used for generating a coarse-grained view of
the measured data, whereas the interpolation was applied for treating the sparseness
of the rule bases. The method performed very well for the model system of the
SHDSL connections, 52 predictions from 65 test cases were correct, and the other
19 were acceptable.

References

1. ITU.: Single-pair high-speed digital subscriber line (SHDSL) transceivers, Technical Recom-
mendation G.991.2, ITU Std., Dec 2003
2. ITU.: Single-pair high-speed digital subscriber line (SHDSL) transceivers, amendment 2, Tech-
nical Recommendation G.991.2 amendment 2, ITU Std., Feb 2005
3. Lilik, F., Nagy, Sz., Kóczy, L.T.: Wavelet Based Fuzzy Rule Bases in Pre-Qualification of
Access Networks Wire Pairs. IEEE Africon 2015, Addis Ababa, Ethiopia, 14–17 Sept 2015
paper P-52, 5 p
4. Balázs, K., Kóczy, L.T.: Constructing dense, sparse and hierarchical fuzzy systems by applying
evolutionary optimization techniques. Appl. Comput. Math. 11(1), 81–101 (2012)
5. Lilik, F., Botzheim, J.: Fuzzy based prequalification methods for EoSHDSL technology. Acta
Tech. Jauriensis Ser. Intelligentia Computatorica 4(1), 135–145 (2011)
6. Lilik, F., Kóczy, L.T.: The determination of the bitrate on twisted pairs by Mamdani infer-
ence method, issues and challenges of intelligent system and computational intelligence. Stud.
Comput. Intell. 530, 59–74 (2014). https://doi.org/10.1007/978-3-319-03206-1_5
7. Daubechies, I.: Ten Lectures on Wavelets. CBMS-NSF Regional Conference Series in Applied
Mathematics 61. SIAM, Philadelphia (1992)
8. Chui, C.K.: An Introduction to Wavelets. Academic Press, San Diego (1992)
9. Christopoulos, Ch., Skodras, A., Ebrahimi, T.: The JPEG2000 still image coding system: an
overview. IEEE Trans. Consum. Electron. 46, 1103–1127 (2000). https://doi.org/10.1109/30.
920468
10. MJ2 format standard, ISO/IEC 15444-3:2002/Amd 2:2003
11. Martucci, S.A., Sodagar, I., Chiang, T., Zhang, Y.-Q.: A zerotree wavelet video coder. IEEE
Trans. Circ. Sys. Video Technol. 7, 109–118 (1997)
12. Chai, B.-B., Vass, J., Zhuang, X.: Significance-linked connected component analysis for
wavelet image coding. IEEE Trans. Image Proc. 8, 774–784 (1999)
13. Montoya Zegarra, J.A., Leiteb, N.J., da Silva, Torres R.: Wavelet-based fingerprint image
retrieval. J. Comput. Appl. Math. 227, 297–307 (2008). https://doi.org/10.1016/j.cam.2008.
03.017
14. Kiely, A., Klimesh, M.: The ICER Progressive Wavelet Image Compressor. IPN Progress
Report 42-155, 15 Nov 2003. http://ipnpr.jpl.nasa.gov/tmo/progressreport/42-155/155J.pdf
15. Grossmann, A., Morlet, J.: Decomposition of hardy functions into square integrable wavelets
of constant shape. SIAM J. Math. Anal. 15, 723–736 (1984)
16. Fourier, J.B.J.: Theorie Analitique de la Chaleur. Didot, Paris (1822)
42 F. Lilik et al.

17. Nagy, Sz., Pipek, J.: On an economic prediction of the finer resolution level wavelet coefficients
in electron structure calculations. Phys. Chem. Chem. Phys. 17, 31558–31565 (2015). https://
doi.org/10.1039/C5CP01214G
18. Kóczy, L.T., Hirota, K.: Approximate reasoning by linear rule interpolation and general ap-
proximation. Int. J. Approximate Reasoning 9, 197–225 (1993). https://doi.org/10.1016/0888-
613X(93)90010-B
19. Kóczy, L.T., Hirota, K.: Interpolative reasoning with insufficient evidence in sparse fuzzy rule
bases. Inf. Sci. 71, 169–201 (1993). https://doi.org/10.1016/0020-0255(93)90070-3
20. Tikk, D., Joó, I., Kóczy, L.T., Várlaki, P., Moser, B., Gedeon, T.D.: Stability of interpolative
fuzzy KH-controllers. Fuzzy Sets Syst. 125, 105–119 (2002). https://doi.org/10.1016/S0165-
0114(00)00104-4
21. Haar, A.: Zur Theorie der Orthogonalen Funktionensysteme. Math. Ann. 69, 331–371 (1910)
Abductive Reasoning on Molecular
Interaction Maps

Jean-Marc Alliot, Robert Demolombe, Luis Fariñas del Cerro,

Martín Diéguez and Naji Obeid

Abstract Metabolic networks, formed by a series of metabolic pathways, are made

of intra-cellular and extracellular reactions that determine the biochemical proper-
ties of a cell, and by a set of interactions that guide and regulate the activity of these
reactions. Cancer, for example, can sometimes appear in a cell as a result of some
pathology in a metabolic pathway. Most of these pathways are formed by an intri-
cate and complex network of chain reactions, and are often represented in Molecular
Interaction Maps (MIM), a graphical, human readable form of the cell cycle check-
point pathways. In this paper, we present a logic, called Molecular Interaction Logic,
which semantically characterizes MIMs and, moreover, allows us to apply deductive
and abductive reasoning on MIMs in order to find inconsistencies, answer queries
and infer important properties about those networks.

1 Introduction

Metabolic networks, formed by series of metabolic pathways, are made of intra-

cellular and extracellular reactions that determine the biochemical properties of a cell,
and by a set of interactions that guide and regulate the activity of these reactions.
These reactions, which can be positive (production of a new protein) or negative
(inhibition of a protein in the cell), are at the center of a cell’s existence and they can

This research was partially supported by the French Spanish Laboratory for Advanced Studies
in Information, Representation and Processing (LEA-IREP). Martín Diéguez was supported by
the Centre international de mathématiques et d’informatique (contract ANR-11-LABX-0040-
CIMI).

J.-M. Alliot · R. Demolombe · L. Fariñas del Cerro · M. Diéguez (B) · N. Obeid

University of Toulouse, CNRS, IRIT, Toulouse, France
e-mail: [email protected]

© Springer International Publishing AG 2018 43

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_4
44 J.-M. Alliot et al.

be modulated by other proteins, which can either enable these reactions or, on the
opposite, inhibit them.
Medical and pharmaceutical researches [15, 19] showed that the break of the
double strand of DNA sometimes appears in a cell as a result of some pathology in
a metabolic pathway, and double strand break (dsb) is a major cause of cancer.
These pathways are used to investigate the molecular determinants of tumor
response in cancers. The molecular parameters include the cell cycle checkpoint,
DNA repair and apoptosis1 pathways [15, 19, 21, 25, 26]. When DNA damage
occurs, cell cycle checkpoints are activated and can rapidly kill the cell by apoptosis
or arrest the cell cycle progression to allow DNA repair before cellular reproduction
or division (see, for instance, the atm-chk2 and atr -chk2 pathways in [26]).
Most of these pathways are formed by an intricate and complex network of chain
reactions, and are often represented in Molecular Interaction Maps (MIM), a human
readable form of the cell cycle checkpoint pathways. MIMs become increasingly
larger and their density is constantly enriched with new information (references, date,
authors, etc.). Although essential for knowledge capitalization and formalization,
MIMs are difficult to use:

– Reading is complex due of the very large number of elements, and reasoning about
the map is even more difficult.
– Using a map to communicate goals is only partially suitable because the represen-
tation formalism requires expertise.
– Maps can contain implicit knowledge, that is taken for granted by one expert, but
is missed by another one.
– Maps can be inconsistent, and these inconsistencies are difficult to detect just by
looking at the map itself.

These problems have been faced from the point of view of nonmonotonic rea-
soning, specially Answer Set Programming (ASP) [13, 14] or action languages [2],
taking advantage of non-monotonictiy and the efficient ASP solvers that are avail-
able. Our approach is based on a reduction to classical logic which allows applying
classical reasoning on such kind of networks.
In this paper we present a method that automatically transforms a MIM into a set
of logical formulas, taking as input the XML files generated by a graphical editor
for biological pathways such as pathvisio [31]. Along this paper we will use, as
examples, subgraphs of the pathway of Fig. 1, which represents the modelling of the
atm-chk2 pathway leading to apoptosis.
The rest of this paper is organized as follows: Sect. 2 introduces the concept of
Molecular Interaction Maps and how they can be translated into a set of logical for-
mulas. Section 3 describes Molecular Interaction Logic, a logic which is capable of
describing and reasoning about general pathways. Section 4 investigates the appli-
cation of deductive as well as abductive reasoning on MIMs. Section 5 presents the
tools used and/or implemented and Sect. 6 investigates the future lines of work.

1 Apoptosis is the process of programmed cell death.

Abductive Reasoning on Molecular Interaction Maps 45

Fig. 1 atm − chk2 pathway

2 Molecular Interaction Maps

A Molecular Interaction Map [20] (MIM) is a diagram convention which repre-

sents the interaction networks of multi-protein complexes, protein modifications and
enzymes that are substrates of other enzymes. Although interactions between ele-
ments of a MIM can be complex, they can be represented using only three basic
connectors: production ( A ), activation () and inhibition (). Figure 1 presents
the atm-chk2 pathway, using only the aforementioned connectors.
A production relation means that a new substance is created as a result of a reaction
on several primary components. For instance, the protein atm can be dimerized
to become the atm_atm protein or phosphorylated at serine 1981 resulting in the
production of atm_ ps1981. These reactions can be triggered or blocked by other
46 J.-M. Alliot et al.

proteins or conditions. For example, in Fig. 1, atm_ ps1981 blocks the dimerization
of atm into atm_atm, while the double strand break (dsb) of DNA triggers the
production of atm_ ps1981 by atm.
These interactions can be “stacked”: for example, protein p53 can be phosphory-
lated at serine 15 to become p53_ ps15 (see Fig. 1). This reaction is triggered by atm,
but the triggering itself has to be activated by dsb and can be blocked by atm_atm.
Thus, the two main actions (production of a protein or inhibition of a protein) can be
triggered or blocked by a stack of preconditions.

2.1 Translation of MIMs into Formulas

Our first goal is to translate any MIM into a set of logical expressions in order to
perform several automated reasoning tasks such as deduction or abduction. First,
focusing on the diagram of Fig. 2 (which corresponds to a sub-diagram of Fig. 1)
will help getting an intuitive idea of how translation is performed.
Here apoptosis arises when protein p53 is phosphorylated at serine 20 or 15
(instances p53_ ps20 and p53_ ps20 respectively). However, apoptosis would not
happen if the dimer p53_mdm2 is present. Thus the context would be if p53 and
either p53_ ps20 or p53_ ps15 are present and p53_mdm2 is absent then apoptosis
is produced (this example should of course be completed with the rules for producing
the rest of objects in the diagram).
The general form of production relations is displayed in Fig. 3.
Each arrow can be either an activation or an inhibition of the relation it applies to,
and these activations/inhibitions can be stacked on any number of levels. The above
examples give the idea behind the translation: it is a recursive process starting from
the production relation and climbing up the tree. In order to formally describe these
graphs, we define below the concepts of pathway context and pathway formula.

Definition 1 (Pathway context) Given a set of entities, a pathway context is formed

by expressions defined by the following grammar:

α ::= α P , α Q |P , Q 

Fig. 2 Apoptosis by p53

p53_ ps20 and p53_ ps15
mediation
p53 ps15 p53 ps20

p53, mdm2 p53 mdm2

apoptosis
Abductive Reasoning on Molecular Interaction Maps 47

Fig. 3 The general form of a c1 ,c2 ,...,cn

basic production

e1 ,e2 ,...,en f1 ,f2 ,...,fn

a1 ,a2 ,...,an b

h1 ,h2 ,...,hn i1 ,i2 ,...,in

g1 ,g2 ,...,gn

where P and Q are sets (possibly empty) of propositional variables representing the
conditions of activation () or inhibition () of the reaction. The first part of the
pair is the activation context, the second part is the inhibition context. One, or both
sets can be empty. 

For example, the p53 A apoptosis reaction of Fig. 2 would lead to the following
two pathway contexts:

 p53_ ps20 , p53_mdm2  (1)

 p53_ ps15 , p53_mdm2  (2)

Definition 2 (Causal pathway formulas) A causal pathway formula is defined by

the following grammar:

F ::= [α]( p1 ∧ · · · ∧ pn → Pr q) |
[α]( p1 ∧ · · · ∧ pn → In q) |
F∧F

where α is a pathway context, p1 , . . . , pn , q are propositional variables while Pr

and In are modal concepts that qualify the process of activation or inhibition of
proteins. 

Applied to the example of Fig. 2, the causal pathway formula associated with the
production rule p53 A apoptosis is
48 J.-M. Alliot et al.

[(1)]( p53 → Pr apoptosis) ∧

[(2)]( p53 → Pr apoptosis) . (3)

Observation 1 Each MIM can now be represented as a causal pathway

formula. 

3 Semantics for Causal Pathway Formulas

In this section the semantics of causal pathway formulas is formally introduced. The
resulting logic, denoted by MIL (Molecular Interaction Logic), extends a previous
work [4, 5] where the MIMs were formalized via first order logic with equality, in
which the pathway contexts were limited to one level of depth. From now on, p
means protein p is present and ¬ p means protein p is absent. Before going into
details, we first provide a method to translate a pathway context into a classical
Boolean expression, which will be used in the definition of the satisfaction relation.

Definition 3 (Activation and inhibition expressions) Given a pathway context α =

α  P , β  Q , the activation and the inhibition expressions associated with the
context α (denoted by A(α) and I (α)) are defined recursively as:
 
A(α) = p ∧ A(α  ) ∧ ( ¬q ∨ I (β  ))
p∈P q∈Q
 
I (α) = ¬ p ∨ I (α  ) ∨ ( q ∧ A(β  ))
p∈P q∈Q

The above expressions define the general forms of A(α) and I (α). If one part of
the context α is empty, then the corresponding part is of course absent in A(α)
and I (α). 

Following such definition, formulas associated with (1) are:

A((1)) = p53_ ps20 ∧ ¬ p53_mdm2

I ((1)) = ¬ p53_ ps20 ∨ p53_mdm2

Definition 4 (MIL interpretation) A MIL interpretation consists of a pair (V1 , V2 ) of

classical evaluations i.e. V : P → {T r ue, False} where P is the set of propositional
variables. 
Abductive Reasoning on Molecular Interaction Maps 49

The intuitive meaning behind these two evaluations correspond for V1 to the
protein present or absent, and for V2 to the state of the protein resulting from the
chemical reactions in the cell.2

Definition 5 (Satisfaction relation) Given a MIL interpretation (V1 , V2 ), the satis-

faction relation is defined as:
(1) (V1 , V2 )  p iff V1 ( p) = T r ue for p ∈ P;
(2) ∧, ∨ and → are satisfied as in classical logic;
(3) (V1 , V2 )  Pr p iff V1 ( p) = V2 ( p) = T r ue, with p ∈ P;
(4) (V1 , V2 )  In p iff V1 ( p) = V2 ( p) = False, with p ∈ P;
(5) (V1 , V2 )  [α]F iff (V1 , V2 )  A(α) or (V1 , V2 )  F, for any pathway formula
[α]F.


While the first four satisfaction relations are simple to understand, the fifth one
is a little bit trickier according to Definition 3; its meaning is that if the conditions
of activation of α are satisfied, then the reaction represented by F holds. As usual, a
formula F is satisfiable if there is a model (V1 , V2 ) such that (V1 , V2 )  F.

Observation 2 MIL can be characterized by the axioms of classical logic, plus the
axioms:
1. [α]F ↔ (A(α) → F)
2. Pr p → p, if p is produced then p is present
3. In p → ¬ p, if p is inhibited then p is absent


As a result of MIL semantics, the causal pathway formula (3) is logically equivalent
to the conjunction of the following implications:

( p53 ∧ p53_20 ∧ ¬ p53_mdm2) → Pr apoptosis (4)

( p53 ∧ p53_20 ∧ ¬ p53_mdm2) → Pr apoptosis (5)

Observation 3 Any MIM can be transformed into a causal pathway formula, and
every causal pathway formula is equivalent to a boolean composition of:
– propositional variables or their negation
– propositional variables qualified by Pr or In or their negation


Axioms (2) and (3) of Observation 2 have as consequence:

2 If the semantics of the modal logic S5 is restricted to have at most two worlds then a strong normal

form in which conjunctions and disjunctions are not in the scope of a modal operator can be found
for this new logic [11]: the pathway causal formulas of MIL verify this condition.
50 J.-M. Alliot et al.

Observation 4 Given a MIM formula F, adding Pr p → p and In p → ¬ p

for each propositional variable p in F, enables us to embbed MIL into classical
logic. 
The notions of completion [3] and production axioms, which are both important
and implicit in MIMs, are presented first.

Definition 6 (Completion Axiom from [3]) Let p be an object that can be produced
through different pathways:

C1 → Pr p
···
Cn → Pr p

Every Ci formula represents one pathway leading to p. The completion axiom for
Pr p is the implication Pr p → (C1 ∨ · · · ∨ Cn ). 

This axiom means that Pr p is a notion local to the current map and, therefore, p has
to be produced by at least one of the Ci possible pathways. For instance, pathway
rules related to apoptosis in Fig. 2 would lead to the following completion axiom:

Pr apoptosis → (( p53 ∧ p53_ ps20 ∧ p53_mdm2)∨( p53 ∧ p53_ ps15 ∧ p53_mdm2)) ,

which corresponds to all the pathways producing apoptosis in Fig. 2.

Before presenting the second axiom, the term endogenous entity has to be intro-
duced. Endogenous entities are the ones which have to be produced in the map in
order to be present.
It is worth to mention that the concept of endogenous entity is a biological notion
and is moreover local to a given MIM; it is up to the biologists to mark in an interaction
map which objects they consider as endogenous. This is an information which is not
currently present in MIMs, as it is implicit knowledge for biologists.3

Definition 7 (Production axiom) Given an endogenous entity p the production

axiom associated with p is defined by p → Pr p. 

For instance, the logical representation of the diagram of Fig. 2 would require the
addition of such axioms for apoptosis, p53_mdm2, p53_ ps20 and p53_ ps15, the
endogenous entities occurring in this map.

3 Ina broadly simplified way, endogenous entities can be considered as “internal variables” or
“output variables” of the model, while exogenous entities are “command variables”.
Abductive Reasoning on Molecular Interaction Maps 51

4 Deductive and Abductive Reasoning on MIMs

In this section, deductive and abductive reasoning are used on MIMs in order to find
inconsistencies in the representations and to answer questions about production or
inhibition of proteins in a MIM.

4.1 Deductive Reasoning

Figure 4 represents the dimerization of atm into atm_atm. When translated into its
logical representation, a SAT-checker finds that this representation is inconsistent.
Why so? When atm and not dsb are present then atm_atm is produced. However,
with dsb, dsb_atm_atm enables the phosphorylation of atm into atm_ ps1981
which then blocks atm_atm, which is inconsistent with the fact that atm_atm is
necessary to produce dsb_atm_atm.
This inconsistency arises because biologists (at least some of them) have some
implicit temporal knowledge about the way these reactions take place, but they do
not represent this knowledge explicitly in MIMs. This shows that if MIMs are to
become a consistent medium for representing proteins interaction, they have to be
enriched with temporal informations.

4.2 Abductive Reasoning

In this section abductive reasoning is used to answer queries formulated on MIMs.

The diagram of Fig. 2 will be used as an example. Its logical encoding consists of
the conjunction of the implications given below:

Fig. 4 The dsb_atm_atm atm atm atm

loop
dsb atm atm

dsb,atm atm atm ps1981

52 J.-M. Alliot et al.

p53 ∧ p53_ ps20 ∧ ¬ p53_mdm2 → Pr apoptosis (6)

p53 ∧ p53_ ps15 ∧ ¬ p53_mdm2 → Pr apoptosis (7)
mdm2 ∧ ¬ p53_ ps20 ∧ p53∧
¬ p53_ ps15 → Pr p53_mdm2 (8)
p53 → Pr p53_ ps20 (9)
p53 → Pr p53_ ps15. (10)

These formulas correspond to 5 different pathways, two describing the production

of apoptosis, two defining the phosphorylation of the protein p53 into p53_ ps20
and p53_ ps15 respectively and, finally, one representing the production of the dimer
p53_mdm2 from mdm2 and p53. The corresponding completion axioms are:

Pr apoptosis → ( p53 ∧ p53_ ps20 ∧ ¬ p53_mdm2) ∨

( p53 ∧ p53_ ps15 ∧ ¬ p53_mdm2) (11)
Pr p53_mdm2 → mdm2 ∧ ¬ p53_ ps20 ∧ p53 ∧ ¬ p53_ ps15 (12)
Pr p53_ ps20 → p53 (13)
Pr p53_ ps15 → p53. (14)

The production axioms must now be added for all endogenous proteins: p53_ ps15
→ Pr p53_ ps15, p53_ ps20 → Pr p53_ ps20, p53_mdm2 → Pr p53_mdm2, and
apoptosis → Pr apoptosis.
The database is now complete and the system can be queried using abductive
methods (see Sect. 5); for example, the question ?apoptosis gives the answer:
apoptosis ∨ p53_ ps20 ∨ p53_ ps15 ∨ p53. These results are coherent from a bio-
logical point of view. If p_53 is present, then apoptosis occurs, but the same con-
clusion can be obtained from either p53_ ps15 or p53_ ps20, since any of them is
present if p53 is.
The diagram of Fig. 2 is now modified by adding a new protein, chk2, which is
required to activate the production of p53_ ps20.
This new representation (Fig. 5) implies that rule (9) becomes p53 ∧ chk2 →
Pr p53_ ps20 and the corresponding completion axiom (13) becomes: Pr p53_ ps20
→ p53 ∧ chk2.
When applying abductive reasoning on the new diagram, the answer is pretty
unexpected:

(apoptosis) ∨ (¬ p53_mdm2 ∧ p53_ ps15)∨

( p53 ∧ ¬ p53_mdm2) ∨ ( p53_ ps15 ∧ ¬mdm2)∨
( p53_ ps15 ∧ chk2) ∨ ( p53 ∧ ¬mdm2)∨
( p53 ∧ chk2) ∨ ( p53_ ps20).
Abductive Reasoning on Molecular Interaction Maps 53

Fig. 5 Modification of the pathway 2 chk2 pathway 1

diagram of Fig. 2

p53

p53 ps15 p53 ps20

p53, mdm2 p53 mdm2

apoptosis

Now p53 alone is not enough for apoptosis anymore; ( p53 and chk2), or ( p53 and
not mdm2) are needed for apoptosis. It is not immediately clear why the simple
introduction of chk2 in the pathway labelled as pathway 2 changes the answer. This
comes from the fact that, in this diagram, p53_ ps15 does not block the production
of p53_mdm2 explicitly. Therefore, if both p53 and mdm2 are present but chk2 is
not, then p53_mdm2 is produced in pathway 2 blocking apoptosis in pathway 1.
As a last, complicated test, the complete diagram in Fig. 1 which represents the whole
atm − chk2 pathway was used. The logical representation of this diagram requires
around 80 clauses. The abductive query (?apoptosis) was performed to find all
conditions on exogenous variables that lead to apoptosis. The answer (( p53 ∧ dsb ∧
atm) ∨ ( pml ∧ chk2 ∧ dsb ∧ atm)) are both correct and a little bit surprising. They
clearly correspond to the paths leading to apoptosis1 and apoptosis3. The reason
why the conditions leading to apoptosis2 do not appear becomes clear when solving
?apoptosis2 alone. The answer to this question (( p53 ∧ chk2 ∧ dsb ∧ atm)) is
subsumed by the one leading to apoptosis1.
Abductive queries give also the states of all the endogenous variables in the graphs
which lead to apoptosis. They can be an invaluable tool for biologists on complex
MIMs, where establishing such results by hand is both tedious and error-prone.

5 Implementation

All the examples presented here were written using Pathvisio, a public-domain MIM
editing software.
A parser was developed to automatically translate the MIM from the XML gener-
ated by Pathvisio, add the formulas generated by the completion axioms and trans-
form the formulas into Conjunctive Normal Form.
As demonstrated above, using Observation 4 we transform each set of clauses of
our language in a set of classical clauses, allowing the use of both deduction and
abduction algorithms for classical logic. Regarding deduction and SAT-checking,
several efficient tools are available, such as the minisat solver [9].
54 J.-M. Alliot et al.

However, regarding abductive reasoning, finding efficient systems able to solve

quickly complex propositional problems is much more difficult. Different algorithms
(see the survey [8] to get an overview of the state of the art) and systems [1, 24,
27] are present in literature. All those tools are implemented in pseudo-interpreted
or interpreted languages such as Prolog, and they are slow when computing the
whole set of prime implicates. The best solver found was SOLAR [23], which is
a general purpose solver for abduction on first order theories with equality. When
applied to our case, SOLAR could not take advantage of the fact that our theories
are always propositional and, due to the complexity of general MIM’s provided by
the biologist, we developed a solver for computing prime implicates, which is based
on the algorithm presented by Kean, Tsiknis [18] and Jackson [17].
This solver has been designed with only efficiency in mind. It uses bit repre-
sentations for clauses and performs subsumptions and resolutions using machine
instructions such as and, or and xor , with a parallel implementation using shared
memory. Using this solver, the same example of Fig. 1 is solved using an ordinary
laptop PC, with 185162519 subsumptions and 1519820 resolutions computed in
0.5 s. The detailed results of this work will be presented in another paper.

6 Conclusions and Future Work

We have presented a method to automatically translate MIMs into logical formulas:

inconsistencies and missing knowledge in existing MIMs can be found, and abduc-
tive problems on quite complex MIMs can be automatically solved. There remain,
however, different kind of problems:
– Currently the parser is only able to translate a subset of MIM relations; while this
subset is enough to express all relations, many “shortcuts” relations in real MIMs
have to be implemented to have an operational tool.
– While our implementation is able to solve abductive queries for an already quite
complex MIM in a very short amount of time, it remains to be seen how it will
behave on more realistic and thus much larger MIMs.
– MIMs have to be extended to represent reaction times, which are currently implicit
and our logic will thus have to be extended to express reaction times.
– In order to reduce the research in a MIM, we plan to enrich the language with
concepts like “aboutness” able to qualify, for example, proteins. It should allow
us to isolate the subgraph of a given MIM, regarding the qualified proteins.

References

1. Alberti, M., Gavanelli, M., Lamma, E., Mello, P., Torroni, P.: The SCIFF abductive proof-
procedure. In: AI*IA’05, p. 135147 (2005)
2. Baral, C., Chancellor, K., Tran, N., Tran, N., Joy, A., Berens, M.: A knowledge based approach
for representing and reasoning about signaling networks. Bioinformatics 20 (suppl 1), i15–i22
(2004)
Abductive Reasoning on Molecular Interaction Maps 55

3. Clark, K.L.: Negation as failure. In: Logic and Databases, pp. 293–322. Plenum Press (1978)
4. Demolombe, R., Fariñas del Cerro, L., Obeid, N.: Logical model for molecular interactions
maps. In: Logical Modeling of Biological Systems, pp. 93–123. Wiley, New York (2014)
5. Demolombe, R., Fariñas del Cerro, L., Obeid, N.: Translation of first order formulas into ground
formulas via a completion theory. J. Appl. Logic (), to appear
6. Demolombe, R., Fariñas del Cerro, L.: An inference rule for hypothesis generation. In: IJCAI’91
(1991)
7. Demolombe, R., Fariñas del Cerro, L.: Information about a given entity: from semantics towards
automated deduction. J. Logic Comput. 20(6), 1231–1250 (2010)
8. Denecker, M., Kakas, A.: Abduction in logic programming. In: Computational Logic: Logic
Programming and Beyond, pp. 402–436 (2002)
9. Een, N., Srensson, N.: An extensible sat-solver. In: SAT’03, pp. 502–518 (2003)
10. Erwig, M., Walkingshaw, E.: Causal reasoning with neuron diagrams. In: VLHCC ’10, pp.
101–108 (2010)
11. Fariñas del Cerro, L., Herzig, A.: Contingency-based equilibrium logic. In: LPNMR’11, pp.
223–228 (2011)
12. Fariñas del Cerro, L., Inoue, K. (eds.): Logical Modeling of Biological Systems. Wiley, New
York (2014)
13. Gebser, M., Guziolowski, C., Ivanchev, M., Schaub, T., Siegel, A., Thiele, S., Veber, P.: Repair
and prediction (under inconsistency) in large biological networks with answer set programming.
In: KR’10 (2010)
14. Gebser, M., Schaub, T., Thiele, S., Veber, P.: Detecting inconsistencies in large biological net-
works with answer set programming. Theory Pract. Logic Program. 11(2–3), 323–360 (2011)
15. Glorian, V., Maillot, G., Poles, S., Iacovoni, J.S., Favre, G., Vagner, S.: Hur-dependent loading
of mirna risc to the mrna encoding the ras-related small gtpase rhob controls its translation
during uv-induced apoptosis. Cell Death Differ. 18(11), 1692–1701 (2011)
16. Inoue, K.: Linear resolution for consequence finding. Artif. Intell. 56(2–3), 301–353 (1992)
17. Jackson, P.: Computing prime implicates incrementally. In: CADE’92, pp. 253–267 (1992)
18. Kean, A., Tsiknis, G.: An incremental method for generating prime implicants/implicates. J.
Symbolic Comput. 9, 185–206 (1990)
19. Kohn, K.W., Pommier, Y.: Molecular interaction map of the p53 and mdm2 logic elements,
which control the off-on swith of p53 response to dna damage. Biochem. Biophys. Res. Com-
mun. 331(3), 816–27 (2005)
20. Kohn, K.W., Aladjem, M.I., Weinstein, J.N., Pommier, Y.: Molecular interaction maps of
bioregulatory networks: a general rubric for systems biology. Mol. Biol. Cell. 17(1), 1–13
(2006)
21. Lee, W., Kim, D., Lee, M., Choi, K.: Identification of proteins interacting with the catalytic
subunit of pp2a by proteomics. Proteomics 7(2), 206–214 (2007)
22. Muggleton, S., Bryant, C.H.: Theory completion using inverse entailment. In: ILP’00, pp.
130–146 (2000)
23. Nabeshima, H., Iwanuma, K., Inoue, K., Ray, O.: SOLAR: an automated deduction system for
consequence finding. AI Commun. 23(2–3), 183–203 (2010)
24. Nuffelen, B.V.: A-system: problem solving through abduction. BNAIC01 Sponsors 1, 591–596
(2001)
25. Pei, H., Zhang, L., Luo, K., Qin, Y., Chesi, M., Fei, F., Bergsagel, P.L., Wang, L., You, Z., Lou,
Z.: MMSET regulates histone H4K20 methylation and 53BP1 accumulation at DNA damage
sites. Nature 470(7332), 124–128 (2011)
26. Pommier, Y., Sordet, O., Rao, V.A., Zhang, H., Kohn, K.W.: Targeting chk2 kinase: molecular
interaction maps and therapeutic rationale. Curr. Pharm. Des. 11(22), 2855–72 (2005)
27. Ray, O., Kakas, A.: ProLogICA: a practical system for abductive logic programming. In:
Proceedings of the 11th International Workshop on Non-monotonic Reasoning, pp. 304–312
(2006)
28. Ray, O., Whelan, K., King, R.: Logic-based steady-state analysis and revision of metabolic
networks with inhibition. In: CISIS’10, pp. 661–666 (2010)
56 J.-M. Alliot et al.

29. Reiser, P.G., King, R.D., Kell, D.B., Muggleton, S., Bryant, C.H., Oliver, S.G.: Developing a
logical model of yeast metabolism. Electron. Trans. Artif. Intell. 5, 233–244 (2001)
30. Rougny, A., Froidevaux, C., Yamamoto, Y., Inoue, K.: Analyzing SBGN-AF Networks Using
Normal Logic Programs. In: Logical Modeling of Biological Systems, pp. 44–55. Wiley, New
York (2014)
31. van Iersel, M.P., Kelder, T., Pico, A.R., Hanspers, K., Coort, S., Conklin, B.R., Evelo, C.:
Presenting and exploring biological pathways with pathvisio. BMC Bioinform., p. 399 (2008)
Efficient Unfolding of Fuzzy Connectives
for Multi-adjoint Logic Programs

Pedro J. Morcillo and Ginés Moreno

Abstract During the last decade we have designed several tools for assisting the
development of flexible software applications coded with a promising language in
the fuzzy logic programming area. In the so-called multi-adjoint logic programming
approach, a set of logic rules are assembled with a set of fuzzy connective definitions
(whose truth functions are defined as functional rules) for manipulating truth degrees
beyond the simpler case of {true,false}. Moreover, we have recently provided opti-
mization techniques by reusing some variants of program transformation techniques
based on unfolding which have been largely exploited in the pure functional -not
fuzzy- setting for enhancing the behavior of such operators. In this paper we exper-
imentally show the benefits of using the new c-unfolding transformation applied on
fuzzy connectives and how to improve the efficiency of the proper unfolding pro-
cess by reusing the very well-known concept of dependency graph. Moreover, we
accompany our technique with cost analysis and discussions on practical aspects.

Keywords Fuzzy Logic Programming · Connectives · Unfolding

1 Introduction

Although logic programming [24] has been widely used as a formal method for prob-
lem solving and knowledge representation, fuzzy logic programming has emerged as a
growing research area for incorporating techniques or constructs based on fuzzy logic

This work has been partially supported by the EU (FEDER), the State Research Agency (AEI)
and the Spanish Ministerio de Economía y Competitividad under grant TIN2016-76843-C4-2-R
(AEI/FEDER, UE).

P. J. Morcillo (B) · G. Moreno

Department Computing System,
University of Castilla-La Mancha, 02071 Albacete, Spain
e-mail: [email protected]
G. Moreno
e-mail: [email protected]
© Springer International Publishing AG 2018 57
L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_5
58 P. J. Morcillo and G. Moreno

to explicitly deal with uncertainty and approximated reasoning. Most fuzzy logic
languages developed during the last decades implement (extended versions of) the
resolution principle introduced by Lee [21], such as Elf-Prolog [14], F-Prolog [23],
generalized annotated logic programming [19], Fril [7], MALP [27], FASILL [15],
the QLP scheme of [37] and the many-valued logic programming language of [38].
In this paper we focus on the so-called multi-adjoint logic programming approach
MALP [25–27], a powerful and promising proposal in the area of fuzzy logic pro-
gramming for which we have developed the FLOPER system (see [31, 36] and visit
the Web site http://dectau.uclm.es/floper/) as well as several techniques for develop-
ing, optimizing and tuning fuzzy applications [10, 17, 33]. Intuitively speaking, logic
programming is extended with a multi-adjoint lattice L of truth values (typically, a
real number between 0 and 1), equipped with a collection of adjoint pairs &i , ←i 
and connectives: implications, conjunctions, disjunctions, and other operators called
aggregators, which are interpreted on this lattice.
In Sect. 2 we explain both the syntax and operational semantics of the MALP
language where, in essence, to solve a MALP goal, i.e., a query to the system plus
a substitution (initially the empty substitution, denoted by id), a generalization of
the classical modus ponens inference rule called admissible steps are systematically
applied on atoms in a similar way to classical resolution steps in pure logic program-
ming, thus returning a state composed by a computed substitution together with an
expression where all atoms have been exploited. Next, this expression is interpreted
under a given lattice, hence returning a pair truth degree; substitution which is
the fuzzy counterpart of the classical notion of computed answer used in pure logic
programming.
In Sect. 3 we collect from [29] our technique for reducing the complexity of con-
nectives (also alleviating the computational cost of derivations) by safely removing all
the intermediate calls performed on the equations defining the behavior of such con-
nectives. We show that this process can be easily described in terms of “unfolding”,
a well-known, widely used, semantics-preserving program transformation operation
which in most declarative paradigms is usually based on the application of computa-
tion steps on the body of program rules (in [10, 16, 17] we describe our experiences
regarding the unfolding of fuzzy logic programs). The novelty of our approach is that
it is the first time that unfolding is not applied to program rules, but to connective
definitions, maintaining the same final goal, i.e., generating more efficient code. The
main three goals addressed in this paper are:
1. The advantages of the unfolding transformation adapted to connective definitions
in our fuzzy setting are experimentally checked at the end of Sect. 3.
2. Moreover, we reuse in Sect. 4 some techniques based on dependency graphs for
improving the proper transformation process as much as possible, also including
computational cost analysis for our resulting algorithm.
3. In Sect. 5 we discuss some practical issues by connecting our techniques with
several tools implemented in our research group.
Finally, before concluding in Sect. 7, a few hints on related work—which also help
to motivate our approach—are presented in Sect. 6.
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 59

2 Multi-adjoint Logic Programming

This section summarizes the main features of multi-adjoint logic programming (see
[25–27] for a complete formulation of this framework). We work with a first order
language, L, containing variables, constants, function symbols, predicate symbols,
and several (arbitrary) connectives to increase language expressiveness: implication
connectives (←1 , ←2 , . . .); conjunctive operators (denoted by &1 , &2 , . . .), disjunc-
tive operators (|1 , |2 , . . .), and hybrid operators (usually denoted by @1 , @2 , . . .),
all of them are grouped under the name of “aggregators” or directly “connectives”.
Aggregation operators are useful to describe/specify user preferences. An aggrega-
tion operator, when interpreted as a truth function, may be an arithmetic mean, a
weighted sum or in general any monotone application whose arguments are values
of a complete bounded lattice L. For example, if an aggregator @ is interpreted
as [[@]](x, y, z) = (3x + 2y + z)/6, we are giving the highest preference to the
first argument, then to the second, being the third argument the least significant.
Although these connectives are binary operators, we usually generalize them as
functions with an arbitrary number of arguments. So, we often write @(x 1 , . . . , xn )
instead of @(x1 , . . . , @(xn−1 , xn ), . . .). By definition, the truth function for an n-
ary aggregation operator [[@]] : L n → L is required to be monotonous and fulfills
[[@]](, . . . , ) = , [[@]](⊥, . . . , ⊥) = ⊥.
Additionally, our language L contains the values of a multi-adjoint lattice,
L , , ←1 , &1 , . . . , ←n , &n , equipped with a collection of adjoint pairs ←i , &i ,
where each &i is a conjunctor which is intended to the evaluation of modus ponens
[27]. In general, L may be the carrier of any complete bounded lattice but, for read-
ability reasons, in the examples we shall select L as the set of real numbers in the
interval [0, 1]. A L-expression is a well-formed expression composed by values and
connectives of L, as well as variable symbols and primitive operators (i.e., arithmetic
symbols such as ∗, +, min, etc.). In what follows, we assume that the truth func-
tion of any connective @ in L is given by its corresponding connective definition,
that is, an equation or rewriting rule of the form @(x 1 , . . . , xn ) = E, where E is a
L-expression not containing variable symbols apart from x1 , . . . , xn .
A rule is a formula H ←i B, where H is an atomic formula or atom (usually
called the head) and B (which is called the body) is a formula built from atomic
formulas B1 , . . . , Bn —n ≥ 0—, truth values of L, conjunctions, disjunctions and
aggregations. A goal is a body submitted as a query to the system. Roughly speaking,
a multi-adjoint logic program is a set of pairs R; α (we often write R with α),
where R is a rule and α is a truth degree (a value of L) expressing the confidence of
a programmer in the truth of the rule R. By abuse of language, we sometimes refer
a tuple R; α as a “rule”.
The procedural semantics of the multi-adjoint logic language L can be thought as
an operational phase (based on admissible steps) followed by an interpretive one. In
the following, C[A] denotes a formula where A is a sub-expression which occurs in
the—possibly empty—context C[]. Moreover, C[ A/A ] means the replacement of A
by A in context C[], whereas Var (s) refers to the set of distinct variables occurring
60 P. J. Morcillo and G. Moreno

in the syntactic object s, and θ [Var (s)] denotes the substitution obtained from θ by
restricting its domain to Var (s).

Definition 1 (Admissible Step) Let Q be a goal and let σ be a substitution. The

pair Q; σ  is a state and we denote by E the set of states. Given a program P, an
admissible computation is formalized as a state transition system, whose transition
AS
relation  ⊆ (E × E) is the smallest relation satisfying the following admissible
rules (where we always consider that A is the selected atom in Q and mgu(E)
denotes the most general unifier of an equation set E ):
AS
(1) Q[ A]; σ (Q[A/v&i B])θ ; σ θ 
if θ = mgu({A = A}), A ←i B; v in P and B is not empty.
AS
(2) Q[ A]; σ (Q[A/v])θ ; σ θ 
if θ = mgu({A = A}) and A ←i ; v in P.
AS1 AS2
As usual, rules are taken renamed apart. We shall use the symbols  and  to
distinguish between computation steps performed by applying one of the specific
admissible rules. Also, the application of a rule on a step will be annotated as a
AS
superscript of the  symbol.

Definition 2 Let P be a program and let Q be a goal. An admissible derivation is

AS
a sequence Q; id ∗ Q ; θ . When Q is a formula not containing atoms (i.e.,
a L-expression), the pair Q ; σ , where σ = θ [Var (Q)], is called an admissible
computed answer (a.c.a.) for that derivation.

Example 1 Let P be the following multi-adjoint logic program:

R1 : p(X ) ←godel &prod (|luka (q(X ), 0.6), r (X )) with 0.9

R2 : q(a) ← with 0.8
R3 : r (X ) ← with 0.7

where the labels luka, godel and prod mean respectively for Łukasiewicz logic,
Gödel logic and product logic, that is,

|luka (x1 , x2 ) = min(1, x1 + x2 )

&prod (x1 , x2 ) = x1 ∗ x2
←godel (x1 , x2 ) = i f (x1 > x2 ) then x2 else 1
&godel (x1 , x2 ) = min(x1 , x2 )

Now, we can generate the following admissible derivation (we underline the selected
atoms in each step):
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 61

AS1R1
 p(X ); id 
AS2R2
&godel (0.9, &prod (|luka (q(X 1 ), 0.6), r (X 1 ))); {X/ X 1 } 
AS2R3
&godel (0.9, &prod (|luka (0.8, 0.6), r (a))); {X/a, X 1 /a} 
&godel (0.9, &prod (|luka (0.8, 0.6), 0.7)); {X/a, X 1 /a, X 2 /a}

Here, the admissible computed answer (a.c.a.) is the pair: &godel (0.9,&prod
(|luka (0.8, 0.6), 0.7)); {X/a}.

If we exploit all atoms of a goal, by applying admissible steps as much as needed dur-
ing the operational phase, then it becomes a formula with no atoms (a L-expression)
which can be then directly interpreted w.r.t. lattice L. Although in [27] this last pro-
cess is implicitly included in a definition similar to the previous one for describing
the intended notion of fuzzy computed answer, here we prefer to model it as a new
computational process (transition system) by applying the following definition we
initially presented in [18] (as we will see in further sections, the cost measures pro-
posed in this paper and previous ones [30, 32], are strongly related with the behaviour
and detailed definition of interpretive step):

Definition 3 (Interpretive Step) Let P be a program, Q a goal and σ a substitu-

tion. We formalize the notion of interpretive computation as a state transition sys-
IS
tem, whose transition relation  ⊆ (E × E) is defined as the least one satisfying:
IS
Q[@(r1 , . . . , rn )]; σ   Q[@(r1 , .., rn )/[[@]](r1 , .., rn )];σ , where [[@]] is the
truth function of connective @ in the lattice L ,  associated to P.

Definition 4 Let P be a program and Q; σ  an a.c.a., that is, Q is a goal not
containing atoms (i.e., a L-expression). An interpretive derivation is a sequence
IS
Q; σ  ∗ Q ; σ . When Q = r ∈ L, being L ,  the lattice associated to P, the
state r ; σ  is called a fuzzy computed answer (f.c.a.) for that derivation.

Example 2 We complete the previous derivation of Example 1 by applying 3 inter-

pretive steps in order to obtain the final f.c.a. 0.7; {X/a}, thus generating the fol-
lowing interpretive derivation D1 :

IS
&godel (0.9, &prod (|luka (0.8, 0.6), 0.7)); {X/a} 
IS
&godel (0.9, &prod (1, 0.7)); {X/a} 
IS
&godel (0.9, 0, 7); {X/a} 
0.7; {X/a}.

We have implemented the previous procedural principle into FLOPER as shown

in Fig. 1, where each state (containing its associated goal and substitution) is colored
in yellow and computational steps appear in blue circles (each admissible step is
labeled with the used program rule, whereas label “is” reflects interpretive steps).
62 P. J. Morcillo and G. Moreno

Fig. 1 The fuzzy logic programming environment F LOPER

3 Unfolding Connective Definitions

As we said in the previous section, connective definitions are equations (or rewrit-
ing rules) of the form @(x1 , . . . , xn ) = E, where E is a L-expression which might
contain variable symbols in the set {x1 , . . . , xn }, as well as values, primitive opera-
tors and connectives of a multi-adjoint lattice L. The use of connectives inside the
definition of other connectives is a powerful expressive resource useful not only
for programmers interested in describing complex aggregators, but it also plays an
important role in fuzzy transformation techniques such as the fold/unfold framework
we have described in [10, 16–18]. Consider for instance, the following connective
definition: @complex (x1 , x2 ) = &prod (|luka (x1 , 0.6), x2 ). This hybrid aggregator
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 63

was used in [30, 32] (with slight modifications) for pointing out some observed
discrepancies when measuring the interpretive cost associated to the execution of
MALP programs.

Example 3 A simplified version of rule R1 in Example 1, whose body only contains

an aggregator symbol is R4 : p(X ) ←godel @complex (q(X ), r (X )) with 0.9. Note
that R4 has exactly the same meaning (interpretation) that R1 , although different
syntax. In fact, both rules have the same sequence of atoms in their head and bodies.
The differences are the set of connectives which explicitly appear in their bodies
since in R4 we have moved &P and |luka (as well as value 0.6) from the body of the
rule (see R1 ) to the connective definition of @complex .
On the other hand, a classical, simple way for estimating the computational cost
required to built a derivation, consists in counting the number of computational steps
performed on it. So, given a derivation D, we define:
– operational cost O c (D), as the number of admissible steps performed in D, and
– interpretive cost I c (D), as the number of interpretive steps done in D.
Note that the operational and interpretive costs of derivation D1 performed in the pre-
vious section are O c (D1 ) = 3 and I c (D1 ) = 3, respectively. Intuitively, O c informs
us about the number of atoms exploited along a derivation. Similarly, I c seems to
estimate the number of connectives evaluated in a derivation. However, this last state-
ment is not completely true: I c only takes into account those connectives appearing
in the bodies of program rules which are replicated on states of the derivation, but no
those connectives recursively nested in the definition of other connectives, as we are
going to see. Let us use rule R4 instead of R1 for generating the following derivation
D1∗ which returns the same f.c.a than D1 :

AS1R4
 p(X ); id 
AS2R2
&godel (0.9, @complex (q(X 1 ), r (X 1 )); {X/ X 1 } 
AS2R3
&godel (0.9, @complex (0.8, r (a))); {X/a, X 1 /a} 
IS
&godel (0.9, @complex (0.8, 0.7)); {X/a, X 1 /a, X 2 /a} 
IS
&godel (0.9, 0.7); {X/a, X 1 /a, X 2 /a} 
0.7; {X/a, X 1 /a, X 2 /a}

Note that, since we have exploited the same atoms with the same rules (except for
the first steps performed with the equivalent rules R1 and R4 , respectively) in both
derivations, then O c (D1 ) = O c (D1∗ ) = 3. However, although connectives &prod and
|luka have been evaluated in both derivations, in D1∗ such evaluations have not been
explicitly counted as interpretive steps, and consequently they have not been added
to increase the interpretive cost measure I c . This unrealistic situation is reflected by
the abnormal result: I c (D1 ) = 3 > 2 = I c (D1∗ ). In [30, 32] we have described two
different techniques (based respectively on a redefinition of the notion of interpretive
step and in the introduction of the concept of “weight of a connective”) evidencing
64 P. J. Morcillo and G. Moreno

that the interpretive cost of derivation D1∗ is not only lower, but even greater than
derivation D1 . The main reason is that complex connective definitions involving calls
to other aggregators consume more computational resources than other connectives
which only evaluate primitive operators.

The previous example motivates the following definition, which in essence describes
a technique based on classical unfolding transformations for simplifying, when pos-
sible, connective definitions by “unnesting” unnecessary calls to other connectives.

Definition 5 (C-Unfolding) Let L ,  be a multi-adjoint lattice containing the con-

nective definitions @(x1 , . . . , xn ) = E and @ (x1 , . . . , xm ) = E , such that a call
to @ of the form @ (t1 , . . . , tm ) appears in E. Then, the unfolding of connec-
tive @ w.r.t. connective @ or directly, the c-unfolding of @, is the new equa-
tion: @(x1 , . . . , xn ) = E[@ (t1 , . . . , tm )/E ], where E is obtained from the L-
expression E by replacing each variable (formal parameter) xi by its corresponding
value (actual parameter) ti , 1 ≤ i ≤ m, that is E = E [x1 /t1 , . . . , xm /tm ].

We assume here that the rules (equations) describing connective definitions are taken
renamed apart (at least one of them) before applying an unfolding step, as it is also
usual with program rules in many declarative transformation tasks.

Example 4 Focusing now in the connective definition:

@complex (x1 , x2 ) = &prod (|luka (x1 , 0.6), x2 )

... and remembering that |luka (x1 , x2 ) = min(1, x1 + x2 ), then, we can unfold con-
nective @complex w.r.t. connective |luka as follows:
– Firstly, we generate the “matcher” between the call |luka (x1 , 0.6) appearing in the
“rhs” (right hand side) of the first equation rule and the “lhs” (left hand side) of
the second rule |luka (x1 , x2 ), thus producing links x1 /x1 and x2 /0.6.
– Next, we apply both bindings to the rhs of the second rule, obtaining the L-
expression min(1, x1 + 0.6).
– Then, this L-expression is used to replace the original call to |luka in the rhs of
the first rule, producing &prod (min(1, x1 + 0.6), x2 ).
– Finally, this last L-expression conforms the rhs of the new connective definition
for @complex , that is: @complex (x1 , x2 ) = &prod (min(1, x1 + 0.6), x2 ).
Following the same method, but performing now the c-unfolding of @complex w.r.t.
&prod whose connective definition is &prod (x1 , x2 ) = x1 ∗ x2 , we obtain the final
rule defining @complex with the following shape @complex (x1 , x2 ) = min(1, x1 +
0.6) ∗ x2 . Note that the new connective definition is just a simple arithmetic expres-
sions involving primitive operators but no calls to other connectives, as wanted. From
now on, this improved definition will be referred as @unfolded .

To finish this section, in Table 1 we show the benefits of using c-unfolding by means
of a experimental evaluation performed on a desktop computer equipped with an
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 65

Table 1 Evaluating original and unfolded fuzzy connectives: runtimes and speed-up
MN 100 1000 10,000
100 0.04 / 0.03 = 1.33 0.31 / 0.3 = 1.03 2.91 / 2.89 = 1.01
1000 0.11 / 0.03 = 3.67 0.38 / 0.3 = 1.27 3.03 / 2.93 = 1.03
10000 0.76 / 0.04 = 19 1.02 / 0.3 = 3.4 3.62 / 2.92 = 1.24

i3-2310M CPU @ 2.10 GHz and 4,00 GB RAM. We consider an initial com-
plex connective whose definition contains (in a direct or indirect way) N calls to
other connectives and requires the evaluation of M primitive operators like min,
∗, +, and so on. More exactly, assuming that a connective @ directly evaluates m
primitive operators and performs n direct calls to connectives @1 , . . . , @n , then,
we can compute values M and N by means of the auxiliary functions oper s and
calls as follows: M = oper s(@) = m + oper s(@1 ) + · · · + oper s(@n ) and N =
calls(@) = n + calls(@1 ) + · · · + calls(@n ). For instance, if we try to compute
the values M and N for connective @complex (x1 , x2 ) = &prod (|luka (x1 , 0.6), x2 )
in Example 4, we have that m = 0, n = 2 and since |luka (x1 , x2 ) = min(1, x1 + x2 )
and &prod (x1 , x2 ) = x1 ∗ x2 , then oper s(|luka ) = 2 and oper s(&prod ) = 1, while
calls(|luka ) = calls(&prod ) = 0, which implies that M = oper s(@complex ) =
m + oper s(|luka ) + oper s(&prod ) = 0 + 2 + 1 = 3 and N = calls(@complex ) =
n + calls(|luka ) + calls(&prod ) = 2 + 0 + 0 = 2. Remember that after applying
c-unfolding, the improved connective definition does not perform calls to any
other connective, but evaluates the same number of primitive operators M (in fact,
observe that in the improved definition @complex (x1 , x2 ) = (min(1, x1 + 0.6)) ∗ x2
of Example 4 we have that M = 3 and N = 0, as wanted).
Each row in Table 1 refers to a different value for N while each column indicates
an alternative assignment to M. Both parameters vary according to values 100, 1000
and 10,000. Each case has been executed 1000 times and the content of each cell has
the form “runtime-original-connective / runtime-unfolded-connective = speed-up”,
where execution times are expressed in milliseconds. Note that the speed up in the
cells of the first row are not significant due to the fact that the number of connective
calls N is never greater than the number of primitive operators M. On the contrary,
in the last row, since N is always greater or equal than M, we obtain good ranges
of speed up. In particular, this measure is 19 in the leftmost cell due to the fact that
c-unfolding removes all the connective calls which caused the low efficiency of the
initial connective definition.

4 C-Unfolding and Dependency Graphs

The use of “graphs” (and many different extensions/variants of this formal concept)
as an auxiliary data structure helping to analyze the behaviour of systems/programs
66 P. J. Morcillo and G. Moreno

at several levels, also taking profit in practice of its deep mathematical background.
For instance, and simply focusing on termination topics in declarative programming
(which has somehow influenced our recent research interest), the notions of depen-
dency graphs and size-change graphs have been well reported in [5, 22]. In this
section, we will use the first concept for efficiently guiding the c-unfolding process.
Our experiences in fuzzy fold/unfold transformations [10, 16, 17], reveal us that
drastic situations associated to degenerated transformation sequences might even-
tually produce highly nested definitions of connectives. For instance, assume the
following sequence of (extremely inefficient) connective definitions:

@100 (x1 , x2 ) = @99 (x1 , x2 )

@99 (x1 , x2 ) = @98 (x1 , x2 )
@98 (x1 , x2 ) = @97 (x1 , x2 )
........ .........
@1 (x1 , x2 ) = @0 (x1 , x2 )
@0 (x1 , x2 ) = x1 ∗ x2

When trying to solve two expressions of the form @100 (0.9, 0.8) and @0 (0.9, 0.8),
we obtain the same result 0.72, but the effort needed to solve the first expression is
very high (due to the 100 avoidable calls to auxiliary connectives) compared with
the second expression (which simply evaluates the arithmetic operator ∗).
Fortunately, by systematically performing c-unfolding on the previous connec-
tives, this problem is successfully solved in a simple way: after applying a c-unfolding
step on aggregator @100 we obtain @100 (x1 , x2 ) = @98 (x1 , x2 ), which admits a new
c-unfolding process to become @100 (x1 , x2 ) = @97 (x1 , x2 ), and following this trail,
after applying the final one-hundredth c-unfolding step, we reach the desired connec-
tive definition @100 (x1 , x2 ) = x1 ∗ x2 . Of course, the transformation process does
not finish here, because we also need to rearrange the shape of all the remaining
connective definitions. So, for each aggregator @i , 0 ≤ i ≤ 100, we need exactly i
c-unfolding steps to achieve the appropriate connective definition.
However, there exist a second, much more intelligent alternative to highly reduce
the number of transformation steps needed to obtain the same final set of improved
connective definitions. In our example, the idea is to proceed just in the inverse
order than previously. So, since @0 does not admit unfolding, we proceed with
@1 , whose connective definition becomes @1 (x1 , x2 ) = x1 ∗ x2 after just a single
c-unfolding step. Now, we take profit of this improved definition when unfolding
@2 , since in just a unique (not two) c-unfolding step we obtain the optimal definition
@2 (x1 , x2 ) = x1 ∗ x2 . Note that the benefits of this last process, are also inherited
when transforming @3 , @4 and so on. So, the advantages obtained after applying each
c-unfolding on a different connective, are “propagated” to the remaining connectives
being improved, which implies that we simply need one hundred transformation steps
to optimize the definitions of the whole set of connectives.
In order to identify in a systematic way the best ordering for performing c-
unfolding operations on connectives, we firstly construct the dependency graph of
a multi-adjoint lattice L associated to a given program P, i.e., a directed graph that
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 67

contains the connective symbols as nodes and an edge from connective @ to aggre-
gator @ for each connective definition in L of the form @(x1 , . . . , xn ) = E, where
the L-expression E contains a call to @ . Given an edge from node @ to node @ , we
denote it as an out-edge of @ and as an in-edge of @ . For instance, the dependency
graphs associated to all the connectives seen so far are:

∨L ←− @∗ −→ &P

@100 −→ @99 −→ . . . −→ @1 −→ @0

As we are going to see, the use of dependency graphs will largely help us to decide
when to unfold each connective in order to minimize the number of transformation
steps. Anyway, before doing this, it is important to note that the construction of such
graphs constitute a fast way to detect possibly abnormal connective definitions, that
is, those ones involved on cycles in the graph (because their further evaluation might
fall in infinite loop). Fortunately, the presence of cycles is not usual in the dependency
graphs associated to connective definitions.
As formalized in the algorithm of Fig. 2, when selecting a connective to apply
c-unfolding, we give priority to those ones without out-edges (and obviously not
belonging to cycles), as occurs in our examples with nodes labeled with ∨L , &P and
@0 , which in our particular case do not need c-unfolding because their definitions do
not perform calls to other aggregators. Once a concrete connective has been selected
and then unfolded as much as possible (and hence, its definition has been completely
improved by removing all its auxiliary calls), then the proper node as well as all
its in-edges (remember that it has not associated out-edges) are removed from the
graph. The process is iterated as much as needed until the dependency graph becomes
empty. For instance in our example, once removed nodes ∨L , &P and @0 , the new
candidates are nodes @∗ and @1 . The first one is unfolded w.r.t. ∨L and &P and
then removed, whereas the second one is dropped out after being unfolded w.r.t.
@0 . Then the process continues with @2 , next @3 and so on, being @100 the last

Fig. 2 Algorithm for efficiently unfolding connective definitions

68 P. J. Morcillo and G. Moreno

connective whose definition is optimized by applying just a single c-unfolding step,

thus accomplishing with the desired ordering which produces the benefits reported
along this section.
Focusing on cost analysis, we simply wish to indicate that, for a given dependency
graph with n nodes and m edges, it is obvious to see that the number of c-unfolding
steps performed by our algorithm is exactly m, thanks to the high priority given to
nodes without out-edges for being c-unfolded. On the contrary, if we firstly c-unfold
nodes without in-edges, then the complexity of the algorithm could be greater than
linear (even quadratic), as illustrated by our example at the beginning of this section.

5 Implementation and Practical Issues

In this section we establish the feedback and synergies among several independent
tools developed in our research group during the last decade. Whereas in [9] we
present a recent graphical tool for assisting the design of lattices of truth degrees (see
Fig. 3), the functional-logic transformation system SYN T H enables the unfolding-
based optimization of their connective definitions in order to improve the compu-
tational behaviour of those fuzzy logic programs developed with the FLOPER
environment.
The transformation system SYN T H [4, 35] was initially conceived for optimiz-
ing functional logic programs (Curry [11]), and then also used to manipulate pure
functional programs (Haskell [13]). The tool implements five basic transformation
rules (unfolding, folding, abstraction, definition introduction and definition elimina-
tion) and two automatic transformation strategies (composition and tupling), in order
to generate efficient sets of rewriting rules coded with Haskell/Curry syntax.1
For the purposes of the present work, we simply need to consider the unfolding oper-
ation of the SYN T H system for being applied on rewriting rules modeling fuzzy
connectives, as occurs with the following ones of our running example:

or_luka A B = min 1 (add A B)

and_prod A B = prod A B
agr_complex A B = and_prod (or_luka A (0,6)) B

Here, the connectives |luka , &prod and @complex are respectively denoted by
or_luka, and_prod and agr_complex; we use the names min, add and prod
for referring to the primitive arithmetic operators min, + and ∗, respectively; and
finally A and B are variable symbols. Once the previous set of connective defini-
tions is loaded into the SYN T H system, it conforms the initial program in the
transformation sequence denoted by “asc 0”. Next, we must select the third rule
and click on the unfolding button, thus obtaining the next program “asc 1”

1 In essence, both languages share the same syntax, but they have a different computational behaviour

since Curry extends with extra logic features the pure functional dimension of Haskell.
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 69

Fig. 3 Graphical design of lattices with the Lattice–Maker tool

where the selected rule is replaced by the new fourth one agr_complex A B =
and_prod (min 1 (add A (0,6))) B. Finally, this last rule admits a final
unfolding step to reach our intended final definition for agr_complex (where no
calls to other connectives appear), represented by the fifth rule agr_complex A B
= prod (min 1 (add A (0,6))) B in program “asc 2”. Figure 4 shows
the original and final program of the transformation sequence, where the initial and
improved definitions of agr_complex appear darked in blue.
On the other hand, our “Fuzzy LOgic Programming Environment for Research”
FLOPER, is able to trace the execution of goals with respect to a given MALP pro-
gram and its associated lattice, by drawing derivation trees as the ones shown in
Figs. 1 and 5. When we choose option “ismode = small” then the system is
able to detail the set of connective calls performed along a derivation (sis1 steps)
as well as the set of primitive operators (sis2 steps) evaluated in the same deriva-
tion. Before explaining the three derivations collected on the tree drawn in Fig. 5,
remember that our MALP program looks like:

R1 : p(X ) ←godel &prod (|luka (q(X ), 0.6), r (X )) with 0.9

R2 : q(a) ← with 0.8
R3 : r (X ) ← with 0.7
R4 : p(X ) ←godel @complex (q(X ), r (X )) with 0.9
R5 : p(X ) ←godel @unfolded (q(X ), r (X )) with 0.9
70 P. J. Morcillo and G. Moreno

Fig. 4 The transformation system SYN T H

where @unfolded represents the improved definition (where no calls to other con-
nectives are performed) of @complex .
In the figure, each derivation starts with an admissible step exploiting a different
MALP rule defining predicate p (i.e., R1 , R4 and R5 , respectively) and continues
with rules R2 and R3 in all cases until finishing the admissible phase. From here, the
interpretive phase differs in each derivation since, even when all them evaluate exactly
the same set of primitive operators (so, the number of sis2 steps do coincide), they
perform a different number of direct/indirect calls to connectives (represented by
sis1 steps). So, the left-most branch in Fig. 5 is shorter than the branch in the
middle of the tree (as we have largely explained in the previous section), but note
that the right-most branch, which makes use of the connective whose definition has
been improved by means of our c-unfolding technique, is just the shortest one in the
whole figure, as wanted.
We have just seen by means of a very simple example that the “quality” of the con-
nective definitions accompanying a MALP program directly reverts on the efficiency
of the applications coded with this fuzzy language. Thus, the unfolding technique
applicable on fuzzy connectives described in this paper, is intended to play an impor-
tant role in the performance of large scale programs developed with our FLOPER
environment, as it is the case of the real-world application we have recently developed
with FLOPER in the field of the semantic web [2, 3] (see Fig. 6).
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 71

Fig. 5 Derivation tree using different definitions of fuzzy connectives

6 Related Work

The ideas managed in the previous sections have been mainly inspired by our previous
studies and experiences in the following two topics: fold/unfold transformations in
declarative programming and cost measures in multi-adjoint logic programming ([28,
30, 32]).
Focusing on primitive functional programs, the pioneer work [8] initiates a fer-
tile tradition in program optimization techniques based on fold/unfold transforma-
tions, which has highly attracted a wide audience in the declarative programming
research community during the last three decades (see the first introduction into logic
programming in [39], and then our adaptations to functional logic programming
72 P. J. Morcillo and G. Moreno

Fig. 6 An on-line work session with the fuzzyXPath application

in [4, 35] and fuzzy logic programming in [10, 16–18]). This approach is based
on the construction, by means of a strategy (heuristic), of a sequence of equivalent
programs—called transformation sequence and usually denoted by P 0 , . . . , P n such
that P n is more efficient than P 0 —where each program P i is obtained from the pre-
ceding ones P 0 , . . . , P i−1 by using an elementary transformation rule. The essential
rules are folding and unfolding, i.e., contraction and expansion of sub-expressions
of a program using the definitions of this program (or a preceding one).

Example 5 Consider for instance the following set of rewriting rules describing
classical operations for concatenating two and three lists via the app and double
function symbols, respectively (here we use the constructor symbols “[ ]” and “:” to
model empty and non-empty lists, respectively):

R1 : app([ ], L) =L
R2 : app(H : T, L) = H : app(T, L)
R3 : double(A, B, C) = app(app(A, B), C)

This is a classical example of program optimization by fold/unfold, where it is easy

to see that the underlined expression in rule R3 is rather inefficient because its
evaluation requires the double traversal of the first list (represented by variable A).
Here, we can apply two unfolding steps on rule R3 until obtaining the new pair of
rules:
R4 : double([ ], B, C) = app(B, C)
R6 : double(H : A, B, C) = H : app(app(A, B), C)

Now, observe that the underlined expression in R6 coincides with the body of the
right hand side of rule R3 , and thus, such expression admits a “contraction” in
terms of “double”. This is just what the final folding step produces, thus generating
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 73

the new rule R7 : double(H : A, B, C) = H : double(A, B, C). The reader can

easily check that this recursive definition of “double” is much more efficient than
the one proposed in the original program, since in order to concatenate three lists,
only one traversal of the first two lists is required now. So, the transformation process
successfully finishes with the following final program:

R1 : app([ ], L) =L
R2 : app(H : T, L) = H : app(T, L)
R4 : double([ ], B, C) = app(B, C)
R7 : double(H : A, B, C) = H : double(A, B, C)

On the other hand, in [10] we have developed a much more involved fuzzy fold/unfold
transformation system especially tailored for the MALP framework, whose advan-
tages were illustrated with the following program:

R1 : p(X, Y ) ←luka q(X, Z )&godel q(Z , Y ) with 0.93

R2 : q(a, b) ← with 0.9
R3 : q(b, c) ← with 0.85
R4 : q( f (X ), g(Y )) ←prod q(X, Y ) with 0.8
R5 : q(g(X ), h(Y )) ←prod q(X, Y ) with 0.95

After applying the transformation process, we obtain the improved version:

R11 : new(a, b, c) ←prod 0.9, 0.85 with 1

R16 : new( f (X ), g(Z ), h(Y )) ←prod @1 (new(X, Z , Y )) with 1
R18 : p(X, Y ) ←luka @2 (new(X, Z , Y )) with 0.93

where the new aggregators used in the body of rules R16 and R18 are
defined as @1 (x1 , x2 ) = &prod (0.8, x1 ), &prod (0.95, x2 ) and @2 (x1 , x2 ) =
&godel (x1 , x2 ). In this fuzzy setting, apart from the folding/unfolding operations,
it is mandatory to introduce two new transformations called “tupled definition intro-
duction” and “aggregation”, being this last operation strongly connected with the
notions introduced in this paper, as we are going to see.

As in our previous example, Burstall and Darlington considered in [8] a set of equa-
tions defining functions (also using variables and constructor symbols to build data
structures) as functional programs, with a shape very close to the connective defini-
tions used in this paper, by simply associating their notions of function symbols and
constructor symbols (or more properly, “defined function symbols” and “constructor
function symbols”, respectively), to our concepts of connective symbols and primi-
tive operators, respectively. In fact, our equations describing connectives can be seen
as a particular case of their equations defining program rules, since in the left hand
sides of connective definitions we only allow variables as arguments (they can also
use terms built with constructors).
74 P. J. Morcillo and G. Moreno

For this reason, our notion of c-unfolding does not only accomplish with the
original unfolding definition of [8] (and further, more complex extensions coping
with modern functional and functional-logic programs expressed as Term Rewriting
Systems, TRS’s [6, 12, 20]), but it is even easier, since we simply need a restricted
kind of equations to model connective definitions. Hence, apart from our initial, novel
use of unfolding described in this paper, we think that in the near future it would be
possible to extrapolate to our fuzzy setting many results on transformation techniques
(including folding) with a functional taste, but focusing now on the optimization of
connective definitions.
At this point, we think that it is mandatory to explain why in our initial fuzzy
fold/unfold transformation system described in [10, 18, 34] there exists the risk of
generating artificial and inefficient connective definitions when optimizing multi-
adjoint logic programs. The problem emerges in the fuzzy unfolding transformation
since this operation introduces a great amount of connectives and truth degrees on
unfolded program rules which obstruct the application of further folding steps (as
we have done in Example 5 when generating efficient recursive definitions).

Example 6 Such effects are easy to see even in the propositional case. So, in order
to unfold a fuzzy program rule like:

p ←luka q with 0.9

we must apply an admissible step (according to the procedural semantics described

in Definition 1) on its body, by using for instance a rule like:

q ←godel (r &prod s) with 0.8

Hence, we obtain the new unfolded rule:

p ←luka ( 0.8 &godel (r &prod s)) with 0.9

where the underlined elements confirm our previous comment, that is, the introduc-
tion of extra truth degrees and connectives on the body of unfolded rules.

It is important to note that the existence of such “noisy” elements only emerge in the
fuzzy setting (in contrast with other “crisp” languages) and, for manipulating them,
we require the application of auxiliary techniques which, as we are going to see, will
produce artificial connective definitions: the main motivation of this paper is just
the optimization of such connective definitions by following standard declarative
techniques classically proposed for the optimization of program rules.
So, if we revise the so-called “aggregation transformation rule” described in [10],
we observe that its main goal is to simplify the shape of program rules, by moving
several connective calls from their bodies to the definition of new connective symbols,
in order to give chances for further folding steps to proceed.
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 75

Example 7 Having a look to previous examples, the effects produced by the “aggre-
gation transformation rule” are:
– The program rule initially introduced in Example 1

R1 : p(X ) ←godel &prod (|luka (q(X ), 0.6), r (X )) with 0.9

becomes the transformed rule of Example 3

R∗1 : p(X ) ←godel @∗ (q(X ), r (X )) with 0.9

where remember that @∗ (x1 , x2 ) = &prod (|luka (x1 , 0.6), x2 ).

– Regarding now Example 6, we observe that the program rule obtained after an
unfolding step:

p ←luka (0.8 &godel (r &prod s)) with 0.9

could be replaced by application of an “aggregation step” by the new rule:

p ←luka @(r, s) with 0.9

where @(x1 , x2 ) = (0.8 &godel (x1 &prod x2 )).

Some important features of this transformation operation are:

– No folding step is allowed if a previous aggregation step has not been previously
performed.
– Each aggregation step introduces a new connective definition which necessarily
invokes other connectives previously known, thus producing nested definitions of
aggregators.
– When optimizing programs in practice, several folding/unfolding steps are often
required even on a same given program rule which, in our fuzzy setting, implies
several applications of the aggregation rule in an iterative way (which reinforces the
presence of nested connective definitions when fuzzy programs are manipulated
with these techniques).
Anyway, it must be taken into account that as in many other declarative paradigms,
the time spent once during the correct application of fold/unfold techniques for
optimizing program rules, is largely compensated forever by the fast execution of the
resulting refined programs. In the fuzzy setting, we have seen that this transformation
process focusing only on program rules, also requires a second stage for rearranging
the shape of such “artificial” connective definitions probably produced during the
first transformation phase. The techniques reported along this paper achieve this last
goal (just once, and in a finite time which is reduced as much as possible by the
use of call graphs) on connective definitions without disturbing the benefits initially
reached by fold/unfold on fuzzy program rules.
76 P. J. Morcillo and G. Moreno

7 Conclusions and Future Work

In this paper we were concerned with the optimization of fuzzy logic connectives
whose artificial, inefficient definitions could have been automatically produced by
previous transformation processes applied on fuzzy MALP programs. Our technique,
inspired by rewriting-based unfolding, takes profit from clear precedents in pure func-
tional programming. In this paper we have focused on the optimization of the proper
unfolding process (initially presented in [29]) by making use of dependency graphs
in order to decide the ordering in which several connective calls must be unfolded
inside a concrete connective definition. For the near future, we plan to implement our
technique inside the fuzzy logic programming environment FLOPER (visit http://
dectau.uclm.es/floper/) we have designed for developing applications coded with the
MALP language [1, 3, 40].

References

1. Almendros-Jiménez, J.M., Bofill, M., Luna Tedesqui, A., Moreno, G., Vázquez, C., Villaret,
M.: Fuzzy xpath for the automatic search of fuzzy formulae models. In: Beierle, C., Dekhtyar A.
(eds.) Scalable Uncertainty Management—9th International Conference, SUM 2015, Québec
City, QC, Canada, September 16–18, 2015. Proceedings, volume 9310 of Lecture Notes in
Computer Science, 385–398. Springer, 2015
2. Almendros-Jiménez, J.M., Luna, A., Moreno, G.: Fuzzy logic programming for implementing
a flexible xpath-based query language. Electron. Notes Theor. Comput. Sci. 282, 3–18 (2012)
3. Almendros-Jiménez, J.M., Luna, A., Moreno, G.: Fuzzy xpath through fuzzy logic program-
ming. New Generation Computing 33(2), 173–209 (2015)
4. Alpuente, M., Falaschi, M., Moreno, G., Vidal, G.: Rules + strategies for transforming lazy
functional logic programs. Theoretical Comput. Sci. Elsevier 311(1–3), 479–525 (2004)
5. Arts, T., Giesl, J.: Termination of term rewriting using dependency pairs. Theor. Comput. Sci.
236(1–2), 133–178 (2000)
6. Baader, F., Nipkow, T.: Term Rewriting and All That. Cambridge University Press, Cambridge
(1998)
7. Baldwin, J.F., Martin, T.P., Pilsworth, B.W.: Fril- Fuzzy and Evidential Reasoning in Artificial
Intelligence. Wiley, New Yark (1995)
8. Burstall, R.M., Darlington, J.: A transformation system for developing recursive programs. J.
ACM 24(1), 44–67 (1977)
9. Guerrero, J.A., Martínez, M.S., Moreno, G., Vázquez, C.: Designing lattices of truth degrees
for fuzzy logic programming environments. In: IEEE Symposium Series on Computational
Intelligence, SSCI 2015, Cape Town, South Africa, December 7–10, 2015, pp. 995–1004
IEEE, 2015
10. Guerrero, J.A., Moreno, G.: Optimizing fuzzy logic programs by unfolding, aggregation and
folding. Electron. Notes Theor. Comput. Sci. 219, 19–34 (2008)
11. Hanus, M. (ed.): Curry: An integrated functional logic language (vers. 0.8.3). Available at
http://www.curry-language.org, (2012)
12. Huet, G., Lévy, J.J.: Computations in orthogonal rewriting systems, Part I + II. In: Lassez, J.L.,
Plotkin, G.D. (eds.) Computational logic—essays in honor of alan robinson, pp. 395–443. The
MIT Press, Cambridge, MA (1992)
13. Hutton, Graham: Programming in Haskell. Cambridge University Press, Cambridge (2007)
Efficient Unfolding of Fuzzy Connectives for Multi-adjoint Logic Programs 77

14. Ishizuka, M., Kanai, N.: Prolog-ELF Incorporating Fuzzy Logic. In Joshi, A.K. (eds.) Proceed-
ings of the 9th International Joint Conference on Artificial Intelligence, IJCAI’85, 701–703.
Morgan Kaufmann, 1985
15. Julián Iranzo, P., Moreno, G., Penabad, J., Vázquez, C.: A declarative semantics for a fuzzy logic
language managing similarities and truth degrees. In: Alferes, J., Bertossi, L.E., Governatori,
G., Fodor, P., Roman, D. (eds.) Rule Technologies. Research, Tools, and Applications—10th
International Symposium, RuleML 2016, Stony Brook, NY, USA, July 6–9, 2016. Proceedings,
volume 9718 of Lecture Notes in Computer Science, 68–82. Springer, 2016
16. Julián, P., Medina, J., Morcillo, P.J., Moreno, G., Ojeda-Aciego, M.: An unfolding-based pre-
process for reinforcing thresholds in fuzzy tabulation. In: Proceedings of the 12th Interna-
tional Work-Conference on Artificial Neural Networks, IWANN’13, 647–655. Springer Verlag,
LNCS 7902, Part I, 2013
17. Julián, P., Moreno, G., Penabad, J.: On fuzzy unfolding. a multi-adjoint approach. Fuzzy Sets
Sys. 154, 16–33 (2005)
18. Julián, P., Moreno, G., Penabad, J.: Operational/interpretive unfolding of multi-adjoint logic
programs. J. Universal Comput. Sci. 12(11), 1679–1699 (2006)
19. Kifer, M., Subrahmanian, V.S.: Theory of generalized annotated logic programming and its
applications. J. Logic Program. 12, 335–367 (1992)
20. Klop, J.W., Middeldorp, A.: Sequentiality in Orthogonal Term Rewriting Systems. J. Symbol.
Comput. 161–195 (1991)
21. Lee, R.C.T.: Fuzzy logic and the resolution principle. J. ACM 19(1), 119–129 (1972)
22. Lee, C., Jones, N., Ben-Amram, A.: The size-change principle for program termination. SIG-
PLAN Not. 36(3), 81–92 (2001)
23. Li, D., Liu, D.: A fuzzy Prolog database system. Wiley, New Yark (1990)
24. Lloyd, J.W.: Foundations of Logic Programming. Springer-Verlag, Berlin (1987)
25. Medina, J., Ojeda-Aciego, M., Vojtáš, P.: Multi-adjoint logic programming with continuous
semantics. In: Proc of Logic Programming and Non-Monotonic Reasoning, LPNMR’01, vol.
2173, pp. 351–364. Springer-Verlag, LNAI, (2001)
26. Medina, J., Ojeda-Aciego, M., Vojtáš P.: A procedural semantics for multi-adjoint logic pro-
graming. In: Progress in Artificial Intelligence, EPIA’01, vol. 2258(1), pp. 290–297 Springer-
Verlag, LNAI, (2001)
27. Medina, J., Ojeda-Aciego, M., Vojtáš, P.: Similarity-based Unification: a multi-adjoint
approach. Fuzzy Sets Sys. 146, 43–62 (2004)
28. Morcillo, P.J., Moreno, G., Penabad, J., Vázquez, C.: Fuzzy Computed Answers Collecting
Proof Information. In: Cabestany, J., et al. (eds.) Advances in Computational Intelligence—
Proceedings of the 11th International Work-Conference on Artificial Neural Networks, IWANN
2011, 445–452. Springer Verlag, LNCS 6692, 2011
29. Morcillo, P.J., Moreno, G.: Improving multi-adjoint logic programs by unfolding fuzzy con-
nective definitions. In: Proceedings of the 13th International Work-Conference on Artificial
Neural Networks, IWANN 2015 Mallorca, Spain, June 10–12, pp. 511–524. Springer Verlag,
LNCS 9094, Part I, 2015
30. Morcillo, P.J., Moreno, G.: On cost estimations for executing fuzzy logic programs. In Arabnia,
H.R. et al. (ed.) Proceedings of the 11th International Conference on Artificial Intelligence,
ICAI 2009, July 13–16, 2009, Las Vegas (Nevada), USA
31. Morcillo, P.J., Moreno, G. Penabad, J., Vázquez, C.: A Practical Management of Fuzzy Truth
Degrees using FLOPER. In: Dean, M., et al. (eds.) Proceedings of 4nd Intl Symposium on
Rule Interchange and Applications, RuleML’10. Washington, USA, October 21–23, 20–34.
Springer Verlag, LNCS 6403, 2010
32. Morcillo, P.J. Moreno, G.: Modeling interpretive steps in fuzzy logic computations. In: Di Gesù
V. et al. (eds) Proceedings of the 8th International Workshop on Fuzzy Logic and Applications,
WILF 2009. Palermo, Italy, June 9-12, 44–51. Springer Verlag, LNAI 5571, 2009
33. Moreno, G., Penabad, J., Vidal. G.: Tuning fuzzy logic programs with symbolic execution.
CoRR, abs/1608.04688, 2016
78 P. J. Morcillo and G. Moreno

34. Moreno, G.: Building a fuzzy transformation system. In: Wiedermann, J. et al. (eds.) Proceed-
ings of the 32nd Conference on Current Trends in Theory and Practice of Computer Science,
SOFSEM’2006. Merin, Czech Republic, January 21–27, 409–418. Springer Verlag, LNCS
3831, 2006
35. Moreno, Ginés: A narrowing-based instantiation rule for rewriting-based fold/unfold transfor-
mations. Electr. Notes Theor. Comput. Sci. 86(3), 144–167 (2003)
36. Moreno, G., Vázquez, C.: Fuzzy logic programming in action with floper. J. Software Eng.
Appl. 7, 273–298 (2014)
37. Rodríguez-Artalejo, M., Romero-Díaz, C.: Quantitative logic programming revisited. In: Pro-
ceedings of 9th Functional and Logic Programming Symposium, FLOPS’08, 272–288. LNCS
4989, Springer Verlag, 2008
38. Straccia, U.: Managing uncertainty and vagueness in description logics, logic programs and
description logic programs. In: Reasoning Web, 4th International Summer School, Tutorial
Lectures, number 5224 in Lecture Notes in Computer Science, 54–103. Springer Verlag, 2008
39. Tamaki, H., Sato, T.: Unfold/Fold Transformations of Logic Programs. In: Tärnlund, S. (ed.)
Proceedings of Second International Conference on Logic Programming, 127–139, 1984
40. Vázquez, C., Moreno, G.. Tomás, L., Tordsson, J.: A cloud scheduler assisted by a fuzzy
affinity-aware engine. In: Yazici, A., Pal, N.R., Kaymak, U., Martin, T., Ishibuchi, H., Lin, C.-
T., Sousa, J.M.C., Tütmez, B. (eds.) 2015 IEEE International Conference on Fuzzy Systems,
FUZZ-IEEE 2015, Istanbul, Turkey, August 2–5, 2015, 1–8. IEEE, 2015
On Fuzzy Generalizations of Concept
Lattices

Lubomir Antoni, Stanislav Krajči and Ondrej Krídlo

Abstract We provide an overview of different generalizations of formal concept

analysis based on Fuzzy logic. Firstly, we recall a common platform for early fuzzy
approaches and then, we deal with the data heterogeneity and its various extensions.
A second-order formal context makes a bridge between the early fuzzy extensions
and the heterogeneous frameworks. A second-order formal context is based on the
bonds in L-fuzzy generalization. We present the connections between the related
approaches.

Keywords Formal concept analysis · Bonds · External formal contexts

1 Introduction

The study of structures and mappings which allow to analyze the data in various
forms is a challenging task. In this way, the first attempts to interpret the lattice
theory as concretely as possible and to promote the better communication between
lattice theorists and potential users of lattice theory represent the inception for data
analysis taking into account the binary relations on the objects and attributes sets
[94]. Since the concept hierarchies play an important role here, the term of formal
concept analysis has been adopted for this reasoning. Briefly, formal concept anal-
ysis scrutinizes an object-attribute block of relational data in bivalent form and the
complex foundations were built in [45].

L. Antoni (B) · S. Krajči · O. Krídlo

Institute of Computer Science, Faculty of Science, Pavol Jozef Šafárik
University in Košice, Jesenná 5, 040 01 Košice, Slovakia
e-mail: [email protected]
S. Krajči
e-mail: [email protected]
O. Krídlo
e-mail: [email protected]

© Springer International Publishing AG 2018 79

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_6
80 L. Antoni et al.

For the inception of formal concept analysis, very influential texts arise in [8, 9].
The efficient selection of relevant formal concepts is an interesting and important
issue for investigation and several studies have focused on this scalability question
in formal concept analysis. The stability index of Kuznetsov [67] represents the pro-
portion of subsets of attributes of a given concept whose closure is equal to the extent
of this concept (in an extensional formulation). A high stability index signalizes that
extent does not disappear if the extent of some of its attributes is modified. It helps to
isolate concepts that appear because of noisy objects in [52]. The complete restoring
of the original concept lattice is achieved by stability index in combination with two
other indices. The phenomenon of the basic level of concepts is advocated to select
important formal concepts in Bělohlávek et al. [22]. Five quantitative approaches on
the basic level of concepts and their metrics are comparatively analyzed in [23]. The
approaches on selecting of the formal concepts and simplifying the concept lattices
are examined by [18, 35, 40, 48–50, 66, 68], as well. A large survey on models and
techniques in knowledge processing based on formal concept analysis is given by
[84].
The applications of formal concept analysis which focus on the text retrieval and
text mining were published in the monograph [34]. The state of art of formal concept
analysis and its applications in linguistics, lexical databases, rule mining (iceberg
concept lattice), knowledge management, software engineering, object-engineering,
software development and data analysis are covered in the special LNAI 3626 Vol-
ume devoted to Formal concept analysis—Foundations and Applications and edited
by Berhard Ganter, Gerd Stumme and Rudolf Wille in 2005. An extensive overview
of applications of formal concept analysis in the various application domains includ-
ing software mining, web analytics, medicine, biology and chemistry data is given
by [83]. In addition, we mention the analysis how students learn to program [87], the
techniques for analyzing and improving integrated care pathways [82], evaluation
of questionnaires [20], or the morphological image and signal processing [2]. Our
applications aim at finding a set of representative symptoms for the disease [51],
clustering in a social network area [61] or exploring the educational tasks and objec-
tives of teachers who give the lessons of programming fundamentals [3]. We have
also applied formal concept analysis to explore the elements of MIT App Inventor 2
programming language [3].
Up to date, we remind that some other extensions called relational concept anal-
ysis, logical concept analysis, triadic concept analysis, temporal concept analysis,
rough formal concept analysis and pattern structures are sound and available for
thorough study and their application potential.
Conceptual scaling [45] and pattern structures [44] offer the possibility to process
the many-valued formal contexts. In this many-valued direction, the truth degrees
from Fuzzy logic were advocated by many researchers in effort to promote the
representation and interpretation of data in many-valued form. In this paper, we
recall early fuzzy extensions in formal concept analysis (Sect. 2), then we stress
the representation of fuzzy conjunctions (Sect. 3) in the various generalizations. We
start with the common generalized platform (Sect. 3.1), several novel heterogeneous
approaches (Sect. 3.2) and formal context of higher order (Sect. 3.3). The last one
On Fuzzy Generalizations of Concept Lattices 81

is based on the external formal contexts, harnessing the early fuzzy extensions to
the novel frameworks, and allows to explore the relationships between the formal
contexts by using the bonds in their L-fuzzy generalization.

2 Fuzzy Formal Context

The statements that people use to communicate facts about the world are usually not
bivalent. The truth of such statements is a matter of degree, rather than being only
true or false. Fuzzy logic and fuzzy set theory are frameworks which extend formal
concept analysis in various independent ways. Here, we recall the basic definitions
of fuzzy formal context. The structures of partially ordered set, complete lattice or
residuated lattice are applied here to represent data.

Definition 1 Consider two nonempty sets A and B, an ordered set of truth degrees
T and a mapping R such that R : A × B −→ T . Then, the triple A, B, R is called
a (T )-fuzzy formal context. The elements of the sets A and B are called attributes
and objects, respectively. The mapping R is called a (T )-fuzzy incidence relation.

Definition 2 A residuated lattice is an algebra L , ∧, ∨, ⊗, →, 0, 1 where 0 is the

least element and 1 is the greatest element; ⊗ is associative, commutative, and the
identity x ⊗ 1 holds; and adjointness property, i.e. x ≤ y → z if and only if x ⊗ y ≤
z holds for each x, y, z ∈ L. A residuated lattice is called complete if L , ∧, ∨, 0, 1
is a complete lattice. The operations ⊗ and → are called fuzzy conjunction and
residuum, respectively.

In the definition of (T )-fuzzy formal context, we often take the interval T = [0, 1],
because it is a frequent scale of truth degrees in many applications. For such replace-
ment, the shortened notion of fuzzy formal context has been adopted. Analogously,
we can consider L-fuzzy formal context, or P-fuzzy formal context, having replaced
the interval [0, 1] by the algebraic structures of complete residuated lattice L, par-
tially ordered set P or other plausible scale of truth degrees.

Definition 3 Let A be a nonempty set and let T be an ordered set of truth degrees.
The mapping f : A −→ T is called a fuzzy membership function (also called a T -
fuzzy set or a fuzzy set). We denote a set of all fuzzy membership functions A −→ T
by T A .

The list of early fuzzy approaches is described in Table 1. It summarizes the

names of approaches, types of fuzzy formal contexts and encloses references on
related extensions.
The proper (symmetric) generalization of the classical concept-forming operators
(called also derivation operators or polars, see Ganter and Wille [45]) was introduced
by Bělohlávek [11].
82 L. Antoni et al.

Table 1 Early fuzzy approaches

Name of approach Formal context Authors
L-fuzzy extension L-fuzzy Burusco et al. [28], Pollandt [85], Bělohlávek [11]
One-sided extension Fuzzy Krajči [56], Yahia et al. [27]
α-cuts Fuzzy Snášel, Vojtáš et al. [91]
Crisply generated ext. L-fuzzy Bělohlávek, Sklenář, Zacpal [21]

Definition 4 Let L , ∨, ∧, ⊗, →, 0, 1 be a complete residuated lattice and let

A, B, R be an L-fuzzy formal context. The mappings : L B −→ L A and :
L A −→ L B given by

( (g))(a) = g (a) = (g(b) → R(a, b)), (1)
b∈B

and 
( ( f ))(b) = f (b) = ( f (a) → R(a, b)). (2)
a∈A

are called L-fuzzy concept-forming operators (or L-fuzzy derivation operators, short-
ened L-fuzzy polars).
The set of pairs

L-FCL( A, B, R) = { f, g : f = g, g = f} (3)

is called the set of all L-fuzzy formal concepts of A, B, R. The elements of the
pairs  f, g ∈ L-FCL( A, B, R) are called the extents and intents of L-fuzzy formal
concepts, respectively. The set of all extents is denoted by Ext(A, B, R), the set of
all intents by Int(A, B, R).
Bělohlávek [11] proved that L-fuzzy concept-forming operators form an antitone
Galois connection. In order theory [39, 45, 78], an antitone Galois connection is given
by two opposite mappings between ordered sets. These mappings are order-inverting
and a composition of these two mappings yields two closure operators.
Definition 5 An antitone Galois connection between two ordered sets (P, ≤ P ) and
(Q, ≤ Q ) is a pair g, f  of mappings g : P → Q, f : Q → P such that for all p ∈ P
and q ∈ Q
g( p) ≤ Q q i f andonlyi f p ≥ P f (q).

Other useful properties of L-fuzzy formal concepts have been intensively inves-
tigated since their first inception in [11]. The structure of L-fuzzy concept lattice is
given by the set of all L-fuzzy formal concepts equipped with a crisp order here.
In the research line, the papers [12–14] are oriented to fuzzy Galois connections
and L-fuzzy concept lattices which are enhanced by a fuzzy order of L-fuzzy formal
On Fuzzy Generalizations of Concept Lattices 83

concepts. More extensive results about these issues have been proliferated in [15, 17,
19, 24–26] and the more thorough findings can be found in the recently published
literature.

2.1 One-Sided Concept Lattices

The extents and intents of L-fuzzy formal concepts are both represented by fuzzy
membership functions. In [11], Bělohlávek explains the examples of empirically
empty fuzzy extents, partially empty fuzzy extents (which can be interpreted appar-
ently natural) and full crisp extents in effort to interpret L-fuzzy concepts given by
planets and their attributes.
Full crisp extents with the maintenance of fuzzy intents became the origin of our
approach which was elaborated theoretically [56], as well as practically in our special
oriented applications for management of teams [60, 61].
Definition 6 Let X ⊆ B and ↑: P(B) −→ [0, 1] A . Then, ↑ is a mapping that assigns
to every crisp set X of objects a fuzzy membership function X ↑ of attributes, such
that a value in a point a ∈ A is:

X ↑ (a) = R(a, b). (4)
b∈X

Definition 7 Let f : A → [0, 1] and ↓: [0, 1] A −→ P(B). Then, ↓ is a mapping

that assigns to each fuzzy membership function f of attributes the crisp set

f ↓ = {b ∈ B : (∀a ∈ A)R(a, b) ≥ f (a)}. (5)

The mappings ↑ and ↓ are called the one-sided concept-forming operators. In

addition, the composition of (4) and (5) allows us to define the notion of one-sided
fuzzy concept.
Definition 8 Let X ⊆ B and f ∈ [0, 1] A . The pair X, f  is called a one-sided fuzzy
concept, if X ↑ = f and f ↓ = X . The crisp set of objects X is called the extent and
the fuzzy membership function X ↑ is called the intent of one-sided fuzzy concept.
The set of all one-sided fuzzy concepts ordered by inclusion of extents forms
a complete lattice, called one-sided fuzzy concept lattice, as introduced in [56].
Moreover, in [56] we proved that this construction is a generalization of classical
approach from [45]. The one-sided fuzzy concept lattices for a fuzzy formal context
from Fig. 1 and for P-fuzzy formal context from Fig. 2 are illustrated in Figs. 3 and 4.
The selection of relevant concepts from the set of all one-sided fuzzy concepts is an
important issue. In this direction, the method of subset quality measure [91] takes into
account the idea of α-cuts from Fuzzy logic for searching of significant clusters. The
modified Rice-Siff algorithm provides the efficient generation of clusters and gathers
84 L. Antoni et al.

Fig. 1 Example of fuzzy R a1 a2 a3

formal context
b1 1 0.9 0.8
b2 0.8 0.7 0.7
b3 0.3 0.3 0.3
b4 0.8 0.6 0.9

Fig. 2 Example of P-fuzzy R a1 a2 a3

formal context P 1 b1 1 0 t
b2 s 1 1
s t
b3 0 t s
0 b4 1 s 0

Fig. 3 One-sided fuzzy {b1 ,b2,b3,b4 }, 0.3, 0.3, 0.3

concept lattice for formal
context from Fig. 1 {b1 ,b2,b4 }, 0.8, 0.6, 0.7

{b1 ,b2}, 0.8, 0.7, 0.7 {b1 ,b4 }, 0.8, 0.6, 0.8

{b1 }, 1, 0.9, 0.8 {b4 }, 0.8, 0.6, 0.9

∅, 1, 1, 1

Fig. 4 One-sided fuzzy I – X:

concept lattice for formal X ∅, 1,1,1
context from Fig. 2 {b2 }, s, 1, 1
IX {b1 }, 1, 0, t
V {b4 }, 1, s, 0
VII {b2 ,b3 }, 0, t, s
VI VIII
{b1 ,b2 }, s, 0, t
III
II IV {b2 ,b4 }, s, s, 0
{b1 ,b4 }, 1, 0, 0
{b1 ,b2 ,b4 }, s, 0, 0
I {b1 ,b2 ,b3 ,b4 }, 0, 0, 0

the clusters of objects due to their minimal distance function which is specially
defined. The metric properties of this distance function are proved in [56] and the
efficient algorithm is presented and illustrated. Recently, the Gaussian probabilistic
index [5] was introduced in effort to describe the stability of the extents due to random
fluctuation of values in [0, 1]-fuzzy formal context.
The applications of one-sided approach in a social network area, specially in a
secondary school class in Slovakia [60, 61] help to understand the structure of the
managed team or real school classes. The school data are collected persistently and
the mentioned methods for selection of relevant clusters are applied to serve the
complex information for the class teacher about relationships between schoolmates.
On Fuzzy Generalizations of Concept Lattices 85

Last but not least, the fuzzy subsets in one coordinate and the crisp subsets in the
other coordinate of a formal concept are independently investigated in [21, 27], as
well. Authors of [21] accentuate the crisply generated fuzzy concepts in effort to deal
with the problem of a possibly large number of concepts in an L-fuzzy approach.

3 Representation of Fuzzy Conjunctions and Their

Diversification

The relationships between the objects and attributes are represented by a fuzzy
incidence relation. Fuzzy conjunctions describe the relationships between the truth
degrees of objects and the truth degrees of attributes. In L-fuzzy approach, the fuzzy
conjunctions L × L → L usually describe these relationships.
The inevitability to represent a diverse form of fuzzy conjunctions in a fuzzy for-
mal context leads to the novel definitions of the fuzzy formal contexts in a broader
perspective. Therefore, the notions of generalized formal context, multi-adjoint for-
mal context, heterogeneous formal context, connectional formal context and formal
context of second-order were recently introduced. Each of them allows to represent
data in a specific form and offers the interpretation on the examples from the real
world. The list of novel fuzzy approaches follows in Table 2.
The understanding the way how the fuzzy conjunctions are represented is an
important issue for potential applications. In the following subsections, we present
the cornerstone of generalized, heterogeneous and second-order version of a formal
context.

3.1 Common Generalized Platform

As we have mentioned previously, the fuzzy conjunctions in a fuzzy formal context

express the way in which the degrees of objects and degrees of attributes are mutually

Table 2 Novel fuzzy approaches

Name of approach Formal context Authors
Generalized extension Generalized Krajči [57, 58]
Multi-adjoint extension Multi-adjoint Medina and Ojeda-Aciego [69, 71,
74, 75]
Heterogeneous extension Heterogeneous Krajči et al. [7], Medina et al. [70],
Pócs [79, 80], Popescu [86]
Heterog. one-sided ext. Heterogeneous one-sided Butka and Pócs [29, 32]
Higher order extension Second-order Krídlo et al. [62]
86 L. Antoni et al.

L L L
L 0 0.5 1 L 0 0.5 1 L 0 0.5 1
0 0 0 0 0 0 0 0 0 0 0 0
0.5 0 0.5 0.5 0.5 0 0 0.5 0.5 0 0.25 0.5
1 0 0.5 1 1 0 0.5 1 1 0 0.5 1

Fig. 5 Gödel, Łukasiewicz and product fuzzy conjunctions

connected. In L-fuzzy approach, the frequently used formulas for fuzzy conjunction
⊗ : L × L −→ L are:
– Gödel fuzzy conjunction given by a ⊗ b = min(a, b),
– Łukasiewicz fuzzy conjunction given by a ⊗ b = max(0, a + b − 1),
– product fuzzy conjunction given by a ⊗ b = a · b.
The Gödel fuzzy conjunction points to a smaller number of formal concepts in
comparison with the Łukasiewicz fuzzy conjunction. Moreover, a smaller fluctuation
of fuzzy membership function usually causes a smaller fluctuation of its closure com-
puted by the Łukasiewicz fuzzy conjunction in comparison with the Gödel structure.
The product fuzzy conjuntion can be discretized to compute the formal concepts in
L-fuzzy approach. However, notice that the Gödel, Łukasiewicz and product fuzzy
conjunctions are commutative, monotone and left-continuous operations. The exam-
ples of Gödel, Łukasiewicz and product fuzzy conjunctions are shown in Fig. 5.
The first prospects to non-commutativity of fuzzy conjunctions L × L −→ L is
given by Georgescu and Popescu in [47]. This non-commutative extension requires
the definition of L-fuzzy concept with one extent and two intents.
Analogously in [58], our motivation is based on a more generalized formula-
tion of fuzzy conjunctions in comparison with L-fuzzy formal context (also for
non-commutative cases). The aim is to formulate fuzzy conjunctions by three-
sorted structures of truth degrees which are not necessarily numerical, see for
instance ({s, t, u}, s ≤ t ≤ u) or ({⊥, }, ⊥ ≤ ). Another possibility is to con-
sider not linearly ordered set of truth degrees, e. g. ({0, s, t, 1}, 0 ≤ s ≤ 1, 0 ≤ t ≤
1, s is incomparable with t). We aim at preserving the one extent and one intent in a
formal concept.

Definition 9 Let A, B, R be a P-fuzzy formal context for a poset (P,  P ). Let
(C, C ) and (D,  D ) be the complete lattices.
Let • be an operation (a fuzzy conjunction) such that • is from C × D to P and
is monotone and left-continuous in both arguments, that is:
(1a) c1 C c2 implies c1 • d  P c2 • d for all c1 , c2 ∈ C and d ∈ D.
(1b) d1  D d2 implies c • d1  P c • d2 for all c ∈ C and d1 , d2 ∈D.
(2a) If c • d  P p for d ∈ D, p ∈ P and for all c ∈ S ⊆ C, then S• d  P p.
(2b) If c • d  P p for c ∈ C, p ∈ P and for all d ∈ T ⊆ D, then c • T  P p.
Then, the tuple A, B, P, R, C, D, • is called a generalized formal context.
On Fuzzy Generalizations of Concept Lattices 87

Fig. 6 General form of two C C

fuzzy conjunctions D 0 0.5 1 D s t u
0 0 0 0 0 ⊥ ⊥ ⊥
0.5 0 0 0 0.5 ⊥  
1 0 0.5 1 1 ⊥  

The left part of Fig. 6 indicates non-commutative fuzzy conjunction. The right
part expresses a fuzzy conjunction in a generalized formal context.

Definition 10 Let A, B, P, R, C, D, • be a generalized formal context. Let g :

B −→ D and f : A −→ C. Then, the mappings G : D −→ C
B A
and G :
C −→ D given by
A B


( G (g))(a) = {c ∈ C : (∀b ∈ B)c • g(b)  P R(a, b)},

( G( f ))(b) = {d ∈ D : (∀a ∈ A) f (a) • d  P R(a, b)}

are called the generalized concept-forming operators.

In [57, 58], we proved that a generalized approach is a generalization of L-

fuzzy extension from Bělohlávek [14]. It is sufficient to translate the notions of L-
fuzzy approach to our generalized approach (see Table 3). In the generalized concept-
forming operators, we need to replace a complete lattice (C, C ), a complete lattice
(D,  D ) and a poset (P,  P ) by the complete residuated lattices L , ∨, ∧, ⊗, →
, 0, 1. Then, we can replace the operation • by ⊗. In this setting, it can be proved
that generalized concept-forming operators and L-fuzzy concept-forming operators
coincide (for more details see [58]).
In Fuzzy logic, the hedges are unary connectives which intensify the meaning of
the statements. Bělohlávek and Vychodil [25] applied the hedges for the L-fuzzy
approach to demonstrate the reduction in size of the L-fuzzy concept lattice. By

Table 3 Translation of L-fuzzy approach to generalized approach

Generalized approach [57, 58] L-fuzzy approach [14]
Poset (P,  P ) Complete residuated lattice
L , ∨, ∧, ⊗, →, 0, 1
Complete lattice (C, C ) Complete residuated lattice
L , ∨, ∧, ⊗, →, 0, 1
Complete lattice (D,  D ) Complete residuated lattice
L , ∨, ∧, ⊗, →, 0, 1
R : A×B → P R : A×B → L
•:C×D→ P ⊗:L×L→L
88 L. Antoni et al.

experiments, they claim that a stronger hedge lead to a smaller L-fuzzy concept
lattice. The close relationship between the generalized concept lattices and L-fuzzy
concept lattices constrained by truth-stressing hedges is explored in [59]. The foun-
dations of isotone fuzzy Galois connections with a truth-stressing hedge and a truth-
depressing hedge were developed by Konečný [53] and the reduction of trilattices
with truth-stressing hedges is described by Konečný and Osička in [55].

3.2 Heterogeneous Approaches

The reasoning about heterogeneous data in a formal context appears in [36–38, 70,
72, 79, 86]. These works inspired us to define the heterogeneous extension based on
the full diversification of structures, whereby the construction of concept lattices is
still sound and the proposed extension covers the well-known approaches.
Particularly, the main idea is based on a diversification of all data structures that
can be diversified within a formal context. We use different sets of truth degrees across
a set of objects, different sets of truth degrees across a set of attributes and different
sets of truth degrees across the object-attribute table fields. In addition, for each
object-attribute pair, one can formulate the original fuzzy conjunctions according
to the three-sorted sets of truth degrees of each particular object, attribute and type
of values of corresponding table field. The full definition of heterogeneous formal
context follows.

Definition 11 Consider the set of attributes A and the set of objects B. Let P =
((Pa,b ,  Pa,b ) : a ∈ A, b ∈ B) be a system of posets and let R be a function from
A × B such that R(a, b) ∈ Pa,b for all a ∈ A, b ∈ B. Let C = ((Ca , Ca ) : a ∈ A)
and D = ((Db ,  Db ) : b ∈ B) be systems of complete lattices.
Let  = (•a,b : a ∈ A, b ∈ B) be a system of operations such that •a,b is from
Ca × Db to Pa,b and is monotone and left-continuous in both arguments, that is:
(1a) c1  c2 implies c1 •a,b d  c2 •a,b d for all c1 , c2 ∈ Ca and d ∈ Db .
(1b) d1  d2 implies c •a,b d1  c •a,b d2 for all c ∈ Ca and d1 , d2 ∈ Db.
(2a) If c •a,b d  p for d ∈ Db , p ∈ Pa,b and for all c ∈ S ⊆ Ca , then S •a,b
d  p.
(2b) If c •a,b d  p for c ∈ Ca , p ∈ Pa,b and for all d ∈ T ⊆ Db , then c •a,b
T  p.
Then, we call the tuple A, B, P, R, C, D,  a heterogeneous formal context.

 direct product of complete lattices C1 × C2 × . . . × Cn can be denoted

The
by i∈{1,2,...,n} Ci . Then, the heterogeneous concept-forming operators are defined
between the direct products of complete lattices.
  
Definition 12 Let g ∈ b∈B Db and H : b∈B Db −→ a∈A Ca . Then, H is a
mapping that assigns to every function g a function H (g), such that its value in the
attribute a is given by
On Fuzzy Generalizations of Concept Lattices 89

Ca1 Ca1 Ca1

D b1 s t u Db2 s t u Db 3 s t u
0 0 0 0 0 i i i ⊥ 0 0 0
0.5 0 0.5 0.5 0.33 i ii ii  0 1 1
1 0 0.5 1 0.66 i ii ii
1 i ii ii

Fig. 7 Heterogeneous form of fuzzy conjunctions


( H (g))(a) = {c ∈ Ca : (∀b ∈ B)c •a,b g(b)  R(a, b)}.
  
Symmetrically, let f ∈ a∈A Ca and H : a∈A Ca −→ b∈B Db . Then, H is
a mapping that assigns to every function f a function H ( f ), such that its value in
the object b is given by

( H( f ))(b) = {d ∈ Db : (∀a ∈ A) f (a) •a,b d  R(a, b)}.

The mappings H and H are called heterogeneous concept-forming operators.

A pair g, f  such that H (g) = f and H ( f ) = g is called a heterogeneous for-
mal concept. The ordering of heterogeneous formal concepts g1 , f 1  ≤ g2 , f 2  is
defined by g1 ≤ g2 (or equivalently f 1 ≥ f 2 ).
Then, the poset of all heterogeneous formal concepts ordered by ≤ will be called
a heterogeneous concept lattice. We denote a heterogeneous concept lattice by
HCL(A, B, P, R, C, D, , H , H , ≤).

The fuzzy conjunctions in a heterogeneous form are exemplified in Fig. 7. It

represents the fuzzy conjunctions between one particular attribute and three different
objects.
Illustration of such heterogeneous formal context on an example of longterm and
shortterm preferences of people can be found in [7]. Moreover, the basic theorem on
heterogeneous concept lattices is formulated and proved there.
Theorem 1 (Basic theorem on heterogeneous concept lattices)
1. A heterogeneous concept lattice HCL(A, B, P, R, C, D, , H, H , ≤) is a
complete lattice in which
   
  
gi , f i  = gi , H H fi
i∈I i∈I i∈I

and    
  
gi , f i  = H H gi , fi .
i∈I i∈I i∈I
90 L. Antoni et al.

2. For each a ∈ A and b ∈ B, let 0 Pa,b be the least element of Pa,b such that
0Ca •a,b d = c •a,b 0 Db = 0 Pa,b for all c ∈ Ca , d ∈ Db . Then, a complete lattice
L is isomorphic to HCL(A, B, P, R, C, D, , H , H , ≤) if and only if there
are mappings α : a∈A ({a} × Ca ) −→ L and β : b∈B ({b} × Db ) −→ L such
that:
(1a) α does not increase in the second argument (for a fixed first argument);
(1b) β does not decrease in the second argument (for a fixed first argument);
(2a) Rng(α) is ∧-dense in L;
(2b) Rng(β) is ∨-dense in L; and
(3) For every a ∈ A, b ∈ B, c ∈ Ca and d ∈ Db ,

α(a, c) ≥ β(b, d) if and only if c •a,b d ≤ R(a, b).

Proof The proof is presented in [7].

Other inspiring approaches describe the possibility to apply formal concept anal-
ysis for heterogeneous data, as well. Pócs introduces the connectional approach in
[79, 80] which is based on the antitone Galois connections (Definition 5) between
each pair of object and attribute. Fuzzy conjunctions are replaced by the antitone
Galois connections. Note that the definition of connectional formal context does not
contain the fuzzy incidence relation.

Definition 13 Let A and B be non-empty sets. Let C = ((Ca , Ca ) : a ∈ A) and

D = ((Db ,  Db ) : b ∈ B) be systems of complete lattices. Let G = ((φa,b , ψa,b ) :
a ∈ A, b ∈ B) be a system of antitone Galois connections, such that (φa,b , ψa,b ) is
an antitone Galois connection between (Ca , Ca ) and (Db ,  Db ). Then, the tuple
A, B, G, C,D is called a connectional formal
 context. 
For g ∈ b∈B Db , the mapping G : b∈B Db −→ a∈A Ca is defined by

( G (g))(a) = φa,b (g(b)).
b∈B

  
For f ∈ a∈A Ca , the mapping G : a∈A Ca −→ b∈B Db is defined by

( G( f ))(b) = ψa,b ( f (a)).
a∈A

The mappings G and G are called connectional concept-forming operators.

The connectional formal concepts can be constructed by the connectional concept-

forming operators. An ordered set of connectional formal concepts forms a lattice
structure [79, 80]. Relationship between heterogeneous approach and connectional
approach is described formally in [6].
Figure 8 illustrates an example of antitone Galois connection between the object-
attribute pair (a1 , b1 ) for a1 ∈ A, b1 ∈ B. It is uniquely determined from Fig. 8 and
On Fuzzy Generalizations of Concept Lattices 91

Fig. 8 Antitone Galois (φa1 ,b1 , ψa1 ,b1 )

connection from n u
connectional approach

l m t

k s

from the properties of an antitone Galois connection that φa1 ,b1 (k) = u, φa1 ,b1 (l) = u,
φa1 ,b1 (m) = t, φa1 ,b1 (n) = t, ψa1 ,b1 (s) = n, ψa1 ,b1 (t) = n and ψa1 ,b1 (u) = l.
Medina and Ojeda-Aciego applied the idea of multi-adjointness from logic pro-
gramming [76, 77] to formal concept analysis. The multi-adjoint formal context is
characterized by the adjoints pertaining to the operations • in a common generalized
platform. The first prospects, motivating results and examples were published in [73,
75]. We recall a multi-adjoint formal context here. Two central notions are an adjoint
triple and a multi-adjoint frame.
Definition 14 Let (U, U ), (V, V ) and (P,  P ) be posets. Then, the triple (•, →1
, →2 ) is called an adjoint triple (or implication triple) if • : (U × V ) −→ P, →1 :
(V × P) −→ U and →2 : (U × P) −→ V and moreover

(u • v)  P p iff u U (v →1 p) iff v V (u →2 p).

Let (C, C ), (D,  D ) be complete lattices and (P,  P ) be a poset. A multi-adjoint

frame is a tuple C, D, P, •b : b ∈ B in which (•b , →1b , →2b ) is an adjoint triple
with respect to (C, C ), (D,  D ) and (P,  P ) for all b ∈ B.
Finally, a multi-adjoint context is a tuple A, B, R, σ  such that A and B are the
sets of attributes and objects, respectively, R is a function from A × B such that
R(a, b) ∈ P for all a ∈ A and b ∈ B, and σ is a mapping that associates any object
b ∈ B with a particular adjoint triple (shortly denoted by σ (b)) from the multi-adjoint
frame.
For g ∈ D B , the mapping σ : D B −→ C A is defined by

( σ (g))(a) = (g(b) →1σ (b) R(a, b)).
b∈B

For f ∈ C A , the mapping σ: C A −→ D B is defined by


( σ ( f ))(b) = ( f (a) →2σ (b) R(a, b)).
a∈A

The mappings σ and σ are called multi-adjoint concept-forming operators.

The multi-adjoint formal concepts are defined by means of the multi-adjoint
concept-forming operators, similarly as in a heterogeneous framework. Moreover,
the ordered set of all multi-adjoint formal concepts forms a lattice structure [73, 75].
92 L. Antoni et al.

We remind that the framework of multi-adjoint concept lattices represents the

point of interest of many researchers. The extended approaches of t-concept lattices
[69], multi-adjoint concept lattices based on heterogeneous conjunctors [70], dual
multi-adjoint concept lattices [71], or constraints with hedges [54] are sound and
interpretable. The weaker conditions which are sufficient to generate multi-adjoint
concept lattices have been recently introduced in [41]. In general, the necessary and
sufficient conditions to generate adjunctions [46] and the algebraic properties of
multilattices [33] are the point of interest of current research which is closely related
to these issues.
The close relationships between the multi-adjoint, heterogeneous and connec-
tional approaches are thoroughly presented in [4, 6, 7] and transformations between
them are advocated. Moreover, some other general algebraic aspects of forming fuzzy
concept lattices are described in [81]. A biresiduated mapping (acting on two com-
plete lattices and a poset) is defined for every object-attribute pair in a fuzzy formal
context. Then, Pócs and Pócsová have proved that for every biresiduated mapping
(acting on two complete lattices and a poset), one can construct Galois connection and
moreover, the construction of Galois connection works also between direct products
of complete lattices, which uniquely determines the corresponding fuzzy concept-
forming operators. The theoretical details regarding the representation of the fuzzy
concept lattices in the framework of classical concept lattices are described in [30].
The ideas of heterogeneity applied for the one-sided fuzzy approach (see Sect. 2.1)
are introduced by Butka and Pócs in [29], whereby the possible applications of hetero-
geneous one-sided fuzzy concept lattices are advocated in merging of heterogeneous
subcontexts.

Definition 15 Consider the set of attributes A and the set of objects B. Let C =
((Ca , Ca ) : a ∈ A) be a system of complete lattices. Let R be a function such that
R(a, b) ∈ Ca for all a ∈ A, b ∈ B. Then, we call a tuple A, B, R, C a heterogeneous
one-sided fuzzy formal context (or shortly C-fuzzy formal context).

By a proper generalization of one-sided fuzzy concept-forming operators from

Equation (4) and Equation (5) can be constructed the heterogeneous one-sided fuzzy
concept lattices (or shortly C-fuzzy concept lattices). An example of C-fuzzy formal
context is shown in Fig. 9.
The relationship between the conceptual scaling and the heterogeneous one-sided
fuzzy concept lattices is proved and the method for selecting the relevant heteroge-
neous one-sided fuzzy formal concepts is presented in [31, 32]. The term of general-

Fig. 9 Example of C -fuzzy R a1 a2 a3

formal context
C1 = C2 C3 1 b1 1 0 t
1
b2 0.5 1 1
0.5 s t
b3 0 0.5 s
0
0 b4 1 0.5 0
On Fuzzy Generalizations of Concept Lattices 93

ized one-sided fuzzy approach has been adopted there, but we call it heterogeneous
in effort to emphasize the heterogeneous types of attributes in this extension.
An alternative approach to solve the presence of heterogeneous data in a fuzzy
formal context, taking into account the structures of idempotent semirings, can be
seen in the papers [92, 93] of Valverde-Albacete and Peláez-Moreno.

3.3 Second-Order Formal Context

The main contribution in this subsection is the presentation of novel extended

approach in order to merge L-fuzzy framework and heterogeneous framework. The
name second-order comes from the idea that one can take an auxiliary concept lattice
of some L-fuzzy formal context (entitled as external formal context) and then, the
elements of auxiliary concept lattice are taken for the ordered set of truth degrees
for some object of the main formal context. This is performed for each object and
each attribute of the main formal context. This means that several concept lattices of
external formal contexts are used to compute the summary concept lattice.
We define a second order formal context and its derivation operators, moreover
present the structural properties of this extension in [62].

Definition 16 Consider two non-empty index sets J , I and L-fuzzy formal context
 j∈J A j , i∈I Bi , R, whereby
– A j1 ∩ A j2 = ∅ for any j1 , j2 ∈ J , j1  = j2 ,
– Bi1 ∩ Bi2 = ∅ for any i 1 , i 2 ∈ I , i 1  = i 2 ,
– R : j∈J A j × i∈I Bi −→ L.
Moreover, consider two non-empty sets of L-fuzzy formal contexts (external formal
contexts) notated by
– {O j , A j , Q j  : j ∈ J }, whereby D j = O j , A j , Q j ,
– {Bi , Ti , Pi  : i ∈ I }, whereby Ci = Bi , Ti , Pi .
A second-order formal context is a tuple

A j , {D j ; j ∈ J }, Bi , {Ci ; i ∈ I }, R ,
j∈J i∈I

whereby a subrelation R j,i : A j × Bi −→ L is defined as R j,i (a, b) = R(a, b) for

any a ∈ A j and b ∈ Bi .

Example of a second-order formal context for two external formal contexts on

both objects and attributes sides is shown in Fig. 10.
Particularly, the objects (and attributes) of the second order formal context can
be collected by merging the objects (and attributes) of two or more L-fuzzy formal
contexts (i. e. external formal contexts). Nevertheless, Definition 16 shows that we
94 L. Antoni et al.

Fig. 10 Example of D1 D2
second-order formal context
o1 1 0.5 0 1 o4
o2 0.5 0.5 1 0 o5
o3 1 1 1 1 o6
t3 t2 t1 a1 a2 a3 a4
0.5 0.5 1 b1 1 0 1 1
C1 1 1 0.5 b2 0.5 0.5 1 1
1 1 0.5 b3 0.5 1 1 0.5
0 1 b4 1 1 0.5 0
C2
1 1 b5 1 1 0.5 1
t5 t4

can also consider only one external formal context for objects and only one external
formal context for attributes in a very special case. Furthermore, the extents and
intents of these external formal contexts represent the inputs for the computations of
second-order formal concepts.
The direct product of external
 L-fuzzy lattices L-FCL(C1 ) × L-FCL(C2 ) × . . . ×
L-FCL(Cn ) is denoted by i∈{1,2,...,n} L-FCL(Ci ). To find the second-order formal
concepts, the second-order concept-forming operators are defined between direct
products of the two previously described sets of L-fuzzy concept lattices. The precise
form of the concept-forming operators are introduced in [62] and their formulation
suppose the equivalence functor between the categories of fuzzy Chu correspon-
dences and completely lattice L-ordered sets. For more details, see [62, 65].
Nevertheless, to compute the second-order formal concepts in a more shortened
way, we can use the additional results presented in [62]. To do that, we need to
recall that given two arbitrary sets A and B, the mapping β : B → L A is called
L-multimapping. Then, a notion of L-bond is defined as follows.
Definition 17 Let Ci = Ai , Bi , Ri  for i ∈ {1, 2} be two L-fuzzy formal contexts.
An L-bond is an L-multimapping β : B1 −→ Int(C2 ) such that β t : A2 −→ Ext(C1 ),
where β t (a2 )(b1 ) = β(b1 )(a2 ) for any (b1 , a2 ) ∈ B1 × A2 . The set of all L-bonds
between L-fuzzy formal contexts C1 and C2 is denoted by L-Bonds(C1 , C2 ).
Remark 1 The set of all L-bonds between two L-fuzzy formal contexts forms the
structure of complete lattice. For the proof and the precise formulation of supremum
and infimum of L-bonds see [64]. Relationships between L-bonds and extents of
direct products of L-fuzzy contexts are drawn in [63].
The complete lattice of the nine L-bonds between D1 and C1 (from Fig. 10) can
be seen in Fig. 11. For the calculations of the extents and intents of L-fuzzy formal
contexts by Equation (1) and (2), we selected the Łukasiewicz fuzzy implications
given by a → b = min{1, 1 − a + b} for a, b ∈ L.
Having selected one particular L-bond β from the set of all L-bonds between two
L-fuzzy formal contexts, one can introduce the concept-forming operators between
On Fuzzy Generalizations of Concept Lattices 95

Fig. 11 Concept lattice of

L-bonds between D1 and C1
β9
from Fig. 10 1 1
1 1
1 1

β7 β8
1 0.5 1 1
1 1 1 0.5
1 1 1 0.5

β4 β5 β6
0.5 0.5 1 0.5 1 1
1 1 1 0.5 0.5 0.5
1 1 1 0.5 0.5 0.5

β2 β3
0.5 0.5 1 0.5
1 0.5 0.5 0.5
1 0.5 0.5 0.5

β1
0.5 0.5
0.5 0.5
0.5 0.5
96 L. Antoni et al.

two external sets of all L-fuzzy membership functions (L B1 and L A2 ) as the following
definition states.
Definition 18 Let Ci = Ai , Bi , Ri  for i ∈ {1, 2} be two L-fuzzy contexts and let
β be an arbitrary L-bond between C1 and C2 . The mappings ↑β : L B1 −→ L A2 and
↓β : L A2 −→ L B1 such that

(↑β ( f ))(a) = ( f (b) → β(b)(a)) (6)
b∈B1

and 
(↓β (g))(b) = (g(a) → β(b)(a)) (7)
a∈A2

for any f ∈ L B1 and g ∈ L A2 are called the second-order concept-forming operators

pertaining the L-bond β .
Remark 2 From [62], we know that a pair ↑β , ↓β  forms an antitone Galois con-
nection between complete lattices L-Ext(C1 ), ≤ and L-Int(C2 ), ≤, where ≤ is an
ordering based on fuzzy sets inclusion.
It can be proved that the second-order concept lattices and the special L-fuzzy
concept lattices are isomorphic as the following theorem states.
Theorem 2 Let

K= A j , {D j : j ∈ J }, Bi , {Ci : i ∈ I }, R
j∈J i∈I

be a second-order formal context and let

=
K Aj, Bi , ρ ji ,
j∈J i∈I ( j,i)∈J ×I


ρ ji = {β ∈ L-Bonds(D j , Ci ) : (∀(a j , bi ) ∈ A j × Bi )β(a j )(bi ) ≥ R ji (a j , bi )}

 are isomorphic.
be an L-fuzzy formal context. Then, the concept lattices of K and K
Proof The proof is presented in [62].
In an L-fuzzy formal context K  given by Theorem 2, the subrelation R ji of a
second-order formal context K is replaced by an L-bond ρ ji for each pair ( j, i) ∈
J × I . The bond ρ ji is constructed as the closest bond with respect to the subrelation
R ji . For more details and examples, see [62]. In the second-order formal context given
by Fig. 10, we have that ρ1,1 = β7 , whereby β7 is described in Fig. 11.
We have described the way how to find the second-order formal concepts. The
heterogeneous formal contexts from Definition 11 can be seen in terms of the second-
order formal contexts as the following translation indicates:
On Fuzzy Generalizations of Concept Lattices 97

Int (D1 ) Int (D2 )

Ext (C1 )
1,1,1

Ext (C1 ) 0.5,1,1 1,0.5,0.5 

r r
1,1 1,2

Ext (C2 ) r r 0.5,0.5,0.5 

2,1 2,2

Fig. 12 Heterogeneous formal context derived from second-order formal context

– for attributes and objects take index sets J and I ,

– for the sets of truth degrees D j , Ci for any ( j, i) ∈ J × I take the complete lattices
Int(D j ), ≤ and Ext(Ci ), ≤,
– for the set of truth degrees of table field P j,i take a complete lattice of all fuzzy
relations L A j ×Bi for any ( j, i) ∈ J × I ,
– for the values of fuzzy incidence relation R take R( j, i) = ρ Rj,i , whereby ρ Rj,i :
A j × Bi −→ L such that ρ Rj,i (a, b) = ρ j,i (a)(b); i. e. the closest covering bond
ρ j,i is converted to a subrelation ρ Rj,i for any ( j, i) ∈ J × I ,
– for operations • j,i : Int(D j ) × Ext(Ci ) −→ L A j ×Bi take

(g • j,i f )(a, b) = g(a) ⊗ f (b)

for any g ∈ Int(D j ) and f ∈ Ext(Ci ) and any (a, b) ∈ A j × Bi .

Since the operation ⊗ (Definition 2) is isotone, then • j,i is isotone for ( j, i) ∈

J × I . In [62], we proved that • j,i is left-continuous for ( j, i) ∈ J × I , hence the
assumptions on fuzzy conjunctions from heterogeneous approach [7] are fulfilled.
Heterogeneous formal context in Fig. 12 is derived from the second-order formal
context in Fig. 10. We illustrate the exact values of Ext(C1 ), however the form of
Ext(Ci ), Int(D j ), ρ Rj,i for j, i ∈ {1, 2} can be computed from Fig. 10.

4 Conclusion

The information hidden within data can help to solve the many pending issues within
community, enterprise or science. Turning of data into knowledge and wisdom is
beneficial and necessary, considering either the simple computing in the spread-
sheet calculators or various methods of data analysis which are more complex. Data
collecting, preprocessing, reduction, visualization and dependencies exploration are
important parts of the scientific research, as well.
98 L. Antoni et al.

Fig. 13 Scheme of second-order

relationships between
approaches

heterogeneous connectional
( P ócs)

multi-adjoint generalized
(Medina, Ojeda-Aciego) (Krajči)

one-sided
( B ěloh lá vek et a l., Ya h ia et a l., K r a jči)

classical
(Gan ter and Wille)

We have presented the various approaches in formal concept analysis—from the

early fuzzy approaches to the most recent heterogeneous extensions and the second-
order platform. We presented the one-sided fuzzy approach, the common generalized
platform, the heterogeneous approaches, the second-order platform and a translation
to find the second-order formal concepts. The extensive references on papers deal-
ing with the related generalizations and close areas are included and commented
throughout the paper.
We have also recalled the frameworks of multi-adjoint formal context and con-
nectional formal context which offer the base for the various inspiring extensions.
For mathematical aspects of relationships between the presented extensions, we refer
to the papers [4, 6, 7, 31, 32, 56]. A scheme of these relationships are illustrated in
Fig. 13.
An interval-valued fuzzy setting [1] and the bipolar fuzzy membership functions
which represent the positive and negative information in L-fuzzy formal contexts
[88, 89] seem to be the potential areas for a continuation of a current research line
in the heterogeneous extensions. The results of possibility theory [42], probability
theory [43, 90] or fuzzy relational equations [10, 16] can be harnessed to characterize
the (fuzzy) concept lattices.
On Fuzzy Generalizations of Concept Lattices 99

In conclusion, we are convinced that the area of concept lattices plays an important
role in the research of both theoreticians and practitioners and thus, the research from
a broader perspective is still beneficial.

Acknowledgements We thank the anonymous reviewers for their careful reading of our manuscript
and their many fruitful comments and suggestions. This work was partially supported by the Sci-
entific Grant Agency of the Ministry of Education of Slovak Republic and the Slovak Academy
of Sciences under the contract VEGA 1/0073/15 and by the Slovak Research and Development
Agency under the contract No. APVV–15–0091. This work was partially supported by the Agency
of the Ministry of Education, Science, Research and Sport of the Slovak Republic for the Structural
Funds of EU under the project Center of knowledge and information systems in Košice, CeZIS,
ITMS: 26220220158.

References

1. Alcalde, C., Burusco, A., Fuentes-González, R., Zubia, I.: The use of linguistic variables and
fuzzy propositions in the L-fuzzy concept theory. Comput. Math. Appl. 62(8), 3111–3122
(2011)
2. Alcalde, C., Burusco, A., Fuentes-González, R.: Application of the L-fuzzy concept analysis
in the morphological image and signal processing. Ann. Math. Artif. Intell. 72(1–2), 115–128
(2014)
3. Antoni, L., Guniš, J., Krajči, S., Krídlo, O., Šnajder, L.: The educational tasks and objectives
system within a formal context. In: Bertet, K., Rudolph, S. (eds.) CLA 2014. CEUR Workshop
Proceedings, vol. 1252, pp. 35–46. CEUR-WS.org (2014)
4. Antoni, L., Krajči, S., Krídlo, O.: On different types of heterogeneous formal contexts. In: Pasi,
G., Montero, J., Ciucci, D. (eds.) Proceedings of the 8th Conference of the European Society
for Fuzzy Logic and Technology, pp. 302–309. Atlantis Press (2013)
5. Antoni, L., Krajči, S., Krídlo, O.: Randomized fuzzy formal contexts and relevance of one-
sided concepts. In: Baixeries, J., Sacarea, Ch. (eds.) ICFCA 2015. LNCS (LNAI), vol. 9113.
Springer, Heidelberg (2015)
6. Antoni, L., Krajči, S., Krídlo, O., Macek, B., Pisková, L.: Relationship between two FCA
approaches on heterogeneous formal contexts. In: Szathmary, L., Priss, U. (eds.) CLA 2012.
CEUR Workshop Proceedings, vol. 972, pp. 93–102. CEUR-WS.org (2012)
7. Antoni, L., Krajči, S., Krídlo, O., Macek, B., Pisková, L.: On heterogeneous formal contexts.
Fuzzy Sets Syst. 234, 22–33 (2014)
8. Arnauld, A., Nicole, P.: Logic, or the Art of Thinking. Cambridge University Press, Cambridge
(1996)
9. Barbut, M., Monjardet, B.: Ordre et classification. Algbre et combinatoire. Collection Hachette
Universit, Librairie Hachette (1970)
10. Bartl, E.: Minimal solutions of generalized fuzzy relational equations: probabilistic algorithm
based on greedy approach. Fuzzy Sets Syst. 260, 25–42 (2015)
11. Bělohlávek, R.: Lattices generated by binary fuzzy relations. Tatra Mt. Math. Publ. 16, 11–19
(1999)
12. Bělohlávek, R.: Fuzzy galois connections. Math. Logic Q. 45(4), 497–504 (1999)
13. Bělohlávek, R.: Reduction and a simple proof of characterization of fuzzy concept lattices.
Fundam. Informaticae 46(4), 277–285 (2001)
14. Bělohlávek, R.: Concept lattices and order in fuzzy logic. Ann. Pure Appl. Logic 128, 277–298
(2004)
15. Bělohlávek, R.: What is fuzzy concept lattice? II. In: Kuznetsov, S.O., Slezak, D., Hepting,
D.H., Mirkin, B.G. (eds.) Rough Sets, Fuzzy Sets, Data Mining and Granular Computing.
LNCS, vol. 6743, pp. 19–26. Springer, Heidelberg (2011)
100 L. Antoni et al.

16. Bělohlávek, R.: Sup-t-norm and inf-residuum are one type of relational product: Unifying
framework and consequences. Fuzzy Sets Syst. 197, 45–58 (2012)
17. Bělohlávek, R.: Ordinally equivalent data: a measurement-theoretic look at formal concept
analysis of fuzzy attributes. Int. J. Approx. Reason. 54(9), 1496–1506 (2013)
18. Bělohlávek, R., De Baets, B., Konečný, J.: Granularity of attributes in formal concept analysis.
Inf. Sci. 260, 149–170 (2014)
19. Bělohlávek, R., Klir, G.J.: Concepts and Fuzzy Logic. MIT Press, Cambridge (2011)
20. Bělohlávek, R., Sigmund, E., Zacpal, J.: Evaluation of IPAQ questionnaires supported by formal
concept analysis. Inf. Sci. 181(10), 1774–1786 (2011)
21. Bělohlávek, R., Sklenář, V., Zacpal, J.: Crisply generated fuzzy concepts. In: Ganter, B., Godin,
R. (eds.) ICFCA 2005. LNCS (LNAI), vol. 3403, pp. 269–284. Springer, Heidelberg (2005)
22. Bělohlávek, R., Trnečka, M.: Basic level of concepts in formal concept analysis. In: Domenach,
F., Ignatov, D., Poelmans, J. (eds.) ICFCA 2012. LNCS (LNAI), vol. 7278, pp. 28–44. Springer,
Heidelberg (2012)
23. Bělohlávek, R., Trnečka, M.: Basic Level in Formal Concept Analysis: Interesting Concepts
and Psychological Ramifications. In: Rossi, F. (eds.) IJCAI 2013, pp. 1233–1239. AAAI Press
(2013)
24. Bělohlávek, R., Vychodil, V.: What is fuzzy concept lattice?. In: Bělohlávek, R., Snášel, V.
(eds.) CLA 2005. CEUR Workshop Proceedings, vol. 162, pp. 34–45. CEUR-WS.org (2005)
25. Bělohlávek, R., Vychodil, V.: Fuzzy concept lattices constrained by hedges. J. Adv. Comput.
Intell. Intell. Inform. 11(6), 536–545 (2007)
26. Bělohlávek, R., Vychodil, V.: Formal concept analysis and linguistic hedges. Int. J. Gen. Syst.
41(5), 503–532 (2012)
27. Ben Yahia, S., Jaoua, A.,: Discovering knowledge from fuzzy concept lattice. In: Kandel, A.,
Last, M., Bunke, H. (eds.) Data Mining and Computational Intelligence. Studies in Fuzziness
and Soft Computing, vol. 68, pp. 167–190. Physica-Verlag, Heidelberg (2001)
28. Burusco, A., Fuentes-González, R.: The study of L-fuzzy concept lattice. Mathw. Soft Comput.
3, 209–218 (1994)
29. Butka, P., Pócs, J.: Generalization of one-sided concept lattices. Comput. Inform. 32(2), 355–
370 (2013)
30. Butka, P., Pócs, J., Pócsová, J.: Representation of fuzzy concept lattices in the framework of
classical FCA. J. Appl. Math. 2013, Article ID 236725, 7 pages. (2013)
31. Butka, P., Pócs, J., Pócsová, J.: On equivalence of conceptual scaling and generalized one-sided
concept lattices. Inf. Sci. 259, 57–70 (2014)
32. Butka, P., Pócs, J., Pócsová, J.: Reduction of concepts from generalized one-sided concept
lattice based on subsets quality measure. In: Zgrzywa, A., Choros, A., Sieminski, A. (eds.) New
Research in Multimedia and Internet Systems. Advances in Intelligent Systems and Computing,
vol. 314, pp. 101–111. Springer (2015)
33. Cabrera, I.P., Cordero, P., Gutiérez, G., Martinez, J., Ojeda-Aciego, M.: On residuation in
multilattices: filters, congruences, and homorphisms. Fuzzy Sets Syst. 234, 1–21 (2014)
34. Carpineto, C., Romano, G.: Concept Data Analysis. Wiley, Theory and Applications. J (2004)
35. Cellier, P., Ferré, S., Ridoux, O., Ducassé, M.: A parameterized algorithm to explore formal
contexts with a taxonomy. Int. J. Found. of Comput. Sci. 2, 319–343 (2008)
36. Cornejo, M.E., Medina, J., Ramírez-Poussa, E.: Attribute and size reduction mechanisms in
multi-adjoint concept lattices. J. Comput. Appl. Math. 318, 388–402 (2017)
37. Cornejo, M.E., Medina, J., Ramírez-Poussa, E.: A comparative study of adjoint triples. Fuzzy
Sets Syst. 211, 1–14 (2013)
38. Cornejo, M.E., Medina, J., Ramírez-Poussa, E.: Adjoint negations, more than residuated nega-
tions. Inf. Sci. 345, 355–371 (2016)
39. Davey, B.A., Priestley, H.A.: Introduction to Lattices and Order. Cambridge University Press,
Cambridge (2002)
40. Dias, S.M., Vieira, N.J.: Concept lattices reduction: Definition, analysis and classification. In:
Expert Syst. Appl., https://doi.org/10.1016/j.eswa.2015.04.044
On Fuzzy Generalizations of Concept Lattices 101

41. Diáz-Moreno, J.C., Medina, J., Ojeda-Aciego, M.: On basic conditions to generate multi-adjoint
concept lattices via Galois connections. Int. J. Gen. Syst. 43(2), 149–161 (2014)
42. Dubois, D., Prade, H.: Possibility theory and formal concept analysis: Characterizing indepen-
dent sub-contexts. Fuzzy Sets Syst. 196, 4–16 (2012)
43. Frič, R., Papčo, M.: A categorical approach to probability theory. Stud. Logica 94(2), 215–230
(2010)
44. Ganter, B., Kuznetsov, S.: Pattern structures and their projections. In: Delugach, H.S., Stumme,
G. (eds.) ICCS 2001. LNCS, vol. 2120, pp. 129–142. Springer, Heidelberg (2001)
45. Ganter, B., Wille, R.: Formal Concept Analysis: Mathematical Foundation. Springer, Heidel-
berg (1999)
46. Garciá-Pardo, F., Cabrera, I.P., Cordero, P., Ojeda-Aciego, M., Rodríguez-Sanchez, F.J.: On
the definition of suitable orderings to generate adjunctions over an unstructured codomain. Inf.
Sci. 286, 173–187 (2014)
47. Georgescu, G., Popescu, A.: Concept lattices and similarity in non-commutative fuzzy logic.
Fundamenta Informaticae 53(1), 23–54 (2002)
48. Halaš, R., Mesiar, R., Pócs, J.: Description of sup- and inf-preserving aggregation functions
via families of clusters in data tables, Inf. Sci., to appear. https://doi.org/10.1016/j.ins.2017.
02.060
49. Halaš, R., Pócs, J.: Generalized one-sided concept lattices with attribute preferences. Inf. Sci.
303, 50–60 (2015)
50. Kardoš, F., Pócs, J., Pócsova, J.: On concept reduction based on some graph properties. Knowl.-
Based Syst. 93, 67–74 (2016)
51. Kiseliová, T., Krajči, S.: Generation of representative symptoms based on fuzzy concept lattices.
Adv. Soft Computing 33, 349–354 (2005)
52. Klimushkin, M., Obiedkov, S., Roth, C.: Approaches to the selection of relevant concepts in
the case of noisy data. In: Kwuida, L., Sertkaya, B. (eds.) ICFCA 2010. LNCS, vol. 5986, pp.
255–266. Springer, Heidelberg (2010)
53. Konečný, J.: Isotone fuzzy Galois connections with hedges. Inf. Sci. 181, 1804–1817 (2011)
54. Konečný, J., Medina, J., Ojeda-Aciego, M.: Multi-adjoint concept lattices with heterogeneous
conjunctors and hedges. Ann. Math. Artif. Intell. 72(1–2), 73–89 (2014)
55. Konečný, J., Osička, P.: Triadic concept lattices in the framework of aggregation structures.
Inf. Sci. 279, 512–527 (2014)
56. Krajči, S.: Cluster based efficient generation of fuzzy concepts. Neural Netw. World 13(5),
521–530 (2003)
57. Krajči, S.: The basic theorem on generalized concept lattice. In Bělohlávek R., Snášel V. (eds.)
CLA 2004. CEUR Workshop Proceedings, vol. 110, pp. 25–33. CEUR-WS.org (2004)
58. Krajči, S.: A generalized concept lattice. Log. J. IGPL 13(5), 543–550 (2005)
59. Krajči, S.: Every concept lattice with hedges is isomorphic to some generalized concept lattice.
In Bělohlávek R., Snášel V. (eds.) CLA 2005. CEUR Workshop Proceedings, vol. 162, pp. 1–9.
CEUR-WS.org (2005)
60. Krajči, S.: Social Network and Formal Concept Analysis. In: Pedrycz, W., Chen, S.M. (eds.)
Social Networks: A framework of Computational Intelligence. Studies in Computational Intel-
ligence, vol 526, pp. 41–62. Springer (2014)
61. Krajči, S., Krajčiová, J.: Social network and one-sided fuzzy concept lattices. In: Spyropou-
los, C. (eds.) Fuzz-IEEE 2007. Proceedings of the IEEE International Conference on Fuzzy
Systems, pp. 1–6. IEEE Press (2007)
62. Krídlo, O., Krajči, S., Antoni, L.: Formal concept analysis of higher order. Int. J. Gen. Syst.
45(2), 116–134 (2016)
63. Krídlo, O., Krajči, S., Ojeda-Aciego, M.: The category of L-Chu correspondences and the
structure of L-bonds. Fund. Inform. 115(4), 297–325 (2012)
64. Krídlo, O., Ojeda-Aciego, M.: On L-fuzzy Chu correspondences. Int. J. Comput. Math. 88(9),
1808–1818 (2011)
65. Krídlo, O., Ojeda-Aciego, M.: Revising the link between L-Chu correspondences and com-
pletely lattice L-ordered sets. Ann. Math. Artif. Intell. 72(1–2), 91–113 (2014)
102 L. Antoni et al.

66. Krupka, M.: On complexity reduction of concept lattices: three counterexamples. Inf. Retr.
15(2), 151–156 (2012)
67. Kuznetsov, S.O.: On stability of a formal concept. Ann. Math. Artif. Intell. 49, 101–115 (2007)
68. Kwuida, L., Missaoui, R., Ben Amor, B., Boumedjout, L., Vaillancourt, J.: Restrictions on
concept lattices for pattern management. In: Kryszkiewicz, M., Obiedkov, S. (eds.) CLA 2010.
CEUR Workshop Proceedings, vol. 672, pp. 235–246. CEUR-WS.org (2010)
69. Medina, J., Ojeda-Aciego, M.: Multi-adjoint t-concept lattices. Inf. Sci. 180(5), 712–725 (2010)
70. Medina, J., Ojeda-Aciego, M.: On multi-adjoint concept lattices based on heterogeneous con-
junctors. Fuzzy Sets Syst. 208, 95–110 (2012)
71. Medina, J., Ojeda-Aciego, M.: Dual multi-adjoint concept lattices. Inf. Sci. 225, 47–54 (2013)
72. Medina, J., Ojeda-Aciego, M., Pócs, J., Ramiréz-Poussa, E.: On the Dedekind-MacNeille
completion and formal concept analysis based on multilattices. Fuzzy Sets Syst. 303, 1–20
(2016)
73. Medina, J., Ojeda-Aciego, M., Ruiz-Calviño, J.: On multi-adjoint concept lattices: definition
and representation theorem. Lect. Notes Artif. Intell 4390, 197–209 (2007)
74. Medina, J., Ojeda-Aciego, M., Ruiz-Calviño, J.: Relating generalized concept lattices and
concept lattices for non-commutative conjunctors. Appl. Math. Lett. 21, 1296–1300 (2008)
75. Medina, J., Ojeda-Aciego, M., Ruiz-Calviño, J.: Formal concept analysis via multi-adjoint
concept lattices. Fuzzy Sets Syst. 160(2), 130–144 (2009)
76. Medina, J., Ojeda-Aciego, M., Valverde, A., Vojtáš, P.: Towards biresiduated multi-adjoint logic
programming. In: Conejo, R., Urretavizcaya, M., Pérez-de-la-Cruz, J.-L. (eds.) CAEPIA-TTIA
2003. LNCS (LNAI), vol. 3040, pp. 608–617. Springer, Heidelberg (2004)
77. Medina, J., Ojeda-Aciego, M., Vojtáš, P.: Similarity-based unification: a multi-adjoint approach.
Fuzzy Sets Syst. 146, 43–62 (2004)
78. Ore, Ø.: Galois connexions. Trans. Am. Mathe. Soc. 55, 493–513 (1944)
79. Pócs, J.: Note on generating fuzzy concept lattices via Galois connections. Inf. Sci. 185(1),
128–136 (2012)
80. Pócs, J.: On possible generalization of fuzzy concept lattices using dually isomorphic retracts.
Inf. Sci. 210, 89–98 (2012)
81. Pócs, J., Pócsová, J.: On some general aspects of forming fuzzy concept lattices. Appl. Math.
Sci. 7(112), 5559–5605 (2013)
82. Poelmans, J., Dedene, G., Verheyden, G., Van der Mussele, H., Viaene, S., Peters, E.: Combining
business process and data discovery techniques for analyzing and improving integrated care
pathways. In: Perner, P. (ed.) ICDM 2010, pp. 505–517. Springer-Verlag, Berlin Heidelberg
(2010)
83. Poelmans, J., Ignatov, D.I., Kuznetsov, S.O., Dedene, G.: Formal concept analysis in knowledge
processing: a survey on applications. Expert Syst. Appl. 40(16), 6538–6560 (2013)
84. Poelmans, J., Kuznetsov, S.O., Ignatov, D.I., Dedene, G.: Formal concept analysis in knowledge
processing: a survey on models and techniques. Expert Syst. Appl. 40, 6601–6623 (2013)
85. Pollandt, S.: Fuzzy Begriffe. Springer, Berlin (1997)
86. Popescu, A.: A general approach to fuzzy concepts. Math. Logic Quart. 50(3), 265–280 (2004)
87. Priss, U.: Using FCA to Analyse How Students Learn to Program. In: Cellier, P., Distel, F.,
Ganter, B. (eds.) ICFCA 2013. LNCS, vol. 7880, pp. 216–227. Springer, Heidelberg (2013)
88. Singh, P.K., Kumar, ChA: Bipolar fuzzy graph representation of concept lattice. Inf. Sci. 288,
437–448 (2014)
89. Singh, P.K., Kumar, ChA: A note on bipolar fuzzy graph representation of concept lattice. Int.
J. Comp. Sci. Math. 5(4), 381–393 (2014)
90. Skřivánek, V., Frič, R.: Generalized random events. Int. J. Theor. Phys. 54, 4386–4396 (2015)
91. Snášel, V., Duráková, D., Krajči, S., Vojtáš, P.: Merging concept lattices of α-cuts of fuzzy
contexts. Contrib. Gen. Algebra 14, 155–166 (2004)
92. Valverde-Albacete, F.J., Peláez-Moreno, C.: Towards a generalisation of formal concept anal-
ysis for data mining purposes. In: Missaoui, R., Schmid, J. (eds.) ICFCA 2006. LNCS, vol.
3874, pp. 161–176. Springer, Heidelberg (2006)
On Fuzzy Generalizations of Concept Lattices 103

93. Valverde-Albacete, F.J., Peláez-Moreno, C.: Extending conceptualisation modes for gener-
alised formal concept analysis. Inf. Sci. 181, 1888–1909 (2011)
94. Wille, R.: Restructuring lattice theory: an approach based on hierarchies of concepts. In: Rival,
I. (ed.) Ordered Sets, pp. 445–470. Reidel, Dodrecht-Boston (1982)
Generating Fuzzy Attribute Rules Via
Fuzzy Formal Concept Analysis

Valentín Liñeiro-Barea, Jesús Medina and Inmaculada Medina-Bulo

Abstract Extracting knowledge from databases is a procedure which interest has

increased in a wide variety of areas like stock market, medicine or census data, to
name a few. A compact representation of this knowledge is given by rules. Formal
concept analysis plays an important role in this area. This paper introduces a new
kind of attribute implications considering the fuzzy notions of support and confi-
dence and is also focused on the particular case in which the set of attributes are
intensions. Moreover, an application to clustering for size reduction of concept lat-
tices is included.

Keywords Fuzzy rule · Fuzzy formal concept analysis · Fuzzy set

1 Introduction

The hidden knowledge databases store is a valuable asset in a wide variety of areas
like stock market prediction [16, 22], disease diagnosis [20, 21] or census data
analysis [7, 18], among others. One of the main techniques in order to represent
the knowledge extracted from a database is by rules, summarizing completely the
information stored in the database. These rules are usually extracted via APRIORI

Partially supported by the State Research Agency (AEI) and the European Regional Develop-
ment Fund (ERDF) projects TIN2015-65845-C3-3-R and TIN2016-76653-P.

V. Liñeiro-Barea · J. Medina (B)

Department of Mathematics, University of Cádiz, Cadiz, Spain
e-mail: [email protected]
J. Medina
e-mail: [email protected]
I. Medina-Bulo
Department of Computer Engineering, University of Cádiz, Cadiz, Spain
e-mail: [email protected]

© Springer International Publishing AG 2018 105

L. T. Kóczy and J. Medina (eds.), Interactions Between Computational
Intelligence and Mathematics, Studies in Computational Intelligence 758,
https://doi.org/10.1007/978-3-319-74681-4_7
106 V. Liñeiro-Barea et al.

algorithms [1], which explore frequent itemsets to select the most frequent and con-
fident rules of the database.
Formal Concept Analysis (FCA) [8] is a mathematical technique which helps
to discover relationships between sets of attributes and objects inside a database,
known as concepts. FCA retrieves the main concepts that a database has, and that
can be useful to obtain a rule set. In the classical FCA [1], only boolean attributes
are considered, leading to a set of rules which consider attributes fully true or false.
A most accurate framework is given by fuzzy generalizations where uncertainty and
noise are present.
The computation of attribute implications in one of the most important research
topics in FCA [4, 5, 10, 11, 15, 28]. The main goal of this paper is to define, in
any fuzzy concept lattice framework, a new kind of fuzzy attribute implications (that
is a rule with attributes that might not be fully true or false) and a base of them,
as a minimum set of rules needed to summarize the whole information present in a
database. A particular case in which intensions are considered in the rules is studied
and an application to clustering the concepts of the concept lattice, which provides
a size reduction mechanism of the concept lattice, is presented.
This paper is structured as follows. In Sect. 2 the preliminary definitions are pre-
sented. Then, Sect. 3 presents fuzzy sc-attribute rules, properties and the notion of
base. Next, Sect. 4 studies the particular case in which intensions are considered in
the rules and Sect. 5 applies these rules in order to provide a clustering in the concept
lattice. Finally, conclusions and future work are shown.

2 Preliminaries

This section presents the required preliminary definitions. In all the definitions a
complete lattice (L , ) and a finite universe U are considered.
The following definition introduces the notion of spanning tree, which will be
used in Sect. 4.

Definition 1 ([24]) A spanning tree S of a connected graph V, E is a minimal

subset of E which interconnects all elements in V.

The notion of cardinality is given in a fuzzy framework as follows.

([31]) Given a fuzzy set f : U → L, the fuzzy cardinality of f defined

Definition 2 
as card(f ) = i∈U f (i) is called sigma count.

Formal Concept Analysis (FCA) is applied to relational databases in order to

extract pieces of information, which are called concepts, and they are hierarchized
forming a complete lattice. The concepts are obtained from two concept-forming
operators, which form a Galois connection. There exist different extensions of formal
concept analysis to a fuzzy framework and almost all of them are based on a context
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 107

(a set of attributes A, a set of objects B and a fuzzy relation R between them) and a
Galois connection [3, 14, 19] or a family of Galois connections [2, 23].
The theory and results proposed in this paper are applied to all of them. In order
to recall a particular fuzzy concept lattice, mainly for the examples, we will present
the notion of context and the definitions of the concept-forming operators introduced
in this framework.
The first definition introduces the basic operators considered in the concept-
forming operators.
Definition 3 Let (P1 , ≤1 ), (P2 , ≤2 ), (P3 , ≤3 ) be posets and & : P1 × P2 → P3 ,
 : P3 × P2 → P1 , : P3 × P1 → P2 be mappings, then(&, , ) is an adjoint
triple with respect to P1 , P2 , P3 if:

Adjoint property: x ≤1 z  y iff x&y ≤3 z iff y ≤2 z x

where x ∈ P1 , y ∈ P2 and z ∈ P3 .
The posets (P1 , ≤1 ) and (P2 , ≤2 ) considered in the previous definition are com-
plete lattices in the multi-adjoint concept lattice framework [19] and they are denoted
as (L 1 , 1 ) and (L 2 , 2 ). Now the notion of multi-adjoint frame is presented.
Definition 4 A multi-adjoint frame L is a tuple

(L 1 , L 2 , P, 1 , 2 , ≤, &1 , 1 , 1 , . . . , &n ,  ,
n
n)

where (L 1 , 1 ) and (L 2 , 2 ) are complete lattices, (P, ≤) is a poset and (&i , i , i)

is an adjoint triple with respect to L 1 , L 2 , P, for all i ∈ {1, . . . , n}.
Given a multi-adjoint frame, a context can be introduced.
Definition 5 Let (L 1 , L 2 , P, &1 , . . . , &n ) be a multi-adjoint frame, a context is a
tuple (A, B, R, σ ) such that A and B are non-empty sets (interpreted as attributes and
objects, respectively), R is a P-fuzzy relation R : A × B → P and σ : A × B →
{1, . . . , n} is a mapping which associates any element in A × B with a particular
adjoint triple in the frame.
The symbols L 2B and L 1A denote the set of fuzzy subsets g : B → L 2 , f : A → L 1
respectively. On these sets a pointwise partial order can be considered from the partial
orders in (L 1 , 1 ) and (L 2 , 2 ), which provides L 2B and L 1A the structure of complete
lattice.
After introducing the notions of multi-adjoint frame and context are introduced,
the definition of the concept-forming operators in this framework can be given. The
mappings ↑ : L 2B → L 1A and ↓ : L 1A → L 2B are defined as:

g ↑ (a) = inf{R(a, b) σ (a,b) g(b) | b ∈ B}

f ↓ (b) = inf{R(a, b) σ (a,b) f (a) | a ∈ A}
108 V. Liñeiro-Barea et al.

Table 1 Relation R in Example 1

R a1 a2 a3
b1 0.5 0.0 1.0
b2 1.0 0.5 0.0

for all g ∈ L 2B , f ∈ L 1A , a ∈ A, b ∈ B. These two operators form a Galois connec-

tion [19]. The notion of concept can be defined as usual: A multi-adjoint concept is
a pair g, f  satisfying that g ∈ L 2B , f ∈ L 1A and g ↑ = f , f ↓ = g, with (↑ , ↓ ) being
the Galois connection defined above. The fuzzy subsets g and f in a concept are
usually known as the extent and intent of the concept, respectively.

Definition 6 The multi-adjoint concept lattice associated with a multi-adjoint frame

(L 1 , L 2 , P, &1 , . . . , &n ) and a context (A, B, R, σ ) is the set

M = {g, f  | g ∈ L 2B , f ∈ L 1A and g ↑ = f, f ↓ = g}

in which the ordering is defined by g1 , f 1   g2 , f 2  if and only if g1  g2 ( f 2 

f 1 ). This ordering provides M the structure of a complete lattice [19].

Example 1 Let (L , , &G ) be a multi-adjoint frame, where &G is the Gödel con-
junctor with respect to L = {0, 0.5, 1}. The context (A, B, R, σ ) is formed by the
sets A = {a1 , a2 , a3 } and B = {b1 , b2 }, the relation R is defined from Table 1 and the
constant mapping σ . In the fuzzy set notation, for each a ∈ A, the expression a/1
will simply be written as a and a/0 will be omitted.
The concept lattice related to the given context is shown in Fig. 1. We can see we
have seven concepts, and the Hasse diagram shows the hierarchy among them.

3 Fuzzy sc-Attribute Rules

This section presents a special kind of fuzzy attribute rule and several properties.
From now on, a frame and a context (A, B, R, σ ) will be fixed from which a Galois
connection (↑ , ↓ ) or a family of Galois connections (as in [2, 6, 23]) is defined.
Specifically, we will consider a Galois connection instead of a family, in order to
simplify the notation.
First of all, the classical definitions of support and confidence are provided.
Definition 7 ([1]) The support is defined as the ratio of objects which are related to
a given subset of attributes.

Definition 8 ([1]) Given a set of attributes A and subsets Y1 , Y2 ∈ A, the confidence

is defined as the probability of, given any object that has the attributes in Y1 , it also
has the attributes of Y2 .
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 109

Fig. 1 Concept lattice

related to the context given
in Example 1

These measures on rules will be extended to the fuzzy one next. The first notion
we will consider is the definition of support, which is a key measure of a fuzzy rule.
Definition 9 The support of f ∈ L 1A in (A, B, R, σ ) is defined as

card(f ↓ )
supp(f ) = (1)
|B|

This definition is applied in the following example.

Example 2 Continuing with Example 1, the support of the intent {a1 /0.5, a3 } is
given by:

card({a1 /0.5, a3 }↓ )
supp({a1 /0.5, a3 }) =
|B|
card(b1 ) 1
= =
2 2

The support satisfies interesting properties that are shown next. Since the proof
can straightforwardly be obtained, it is not included.

Proposition 1 Given f 1 , f 2 , f ⊥ ∈ L 1A , with f ⊥ (a) = ⊥ for all a ∈ A, the following

properties hold:
1. If f 1 ≤ f 2 then supp(f2 ) ≤ supp(f1 ).
2. supp(f⊥ ) = 1.
3. 0 ≤ supp( f 1 ) ≤ 1.
110 V. Liñeiro-Barea et al.

Once the notion of support has been introduced, the following definition provides
the definition of fuzzy sc-attribute rule and a specific truth value associated with it.
Definition 10 Given two fuzzy subsets of attributes f 1 , f 2 ∈ L 1A , the fuzzy attribute
rule over A from f 1 to f 2 is given by the expression f 2 ←(s,c) f 1 , where s = supp(f1 )
and the confidence c is defined by

supp(f1 ∪ f 2 )
c= (2)
supp(f1 )

If the confidence is 1, the fuzzy sc-attribute rule is called fuzzy sc-attribute implica-
tion, which is denoted by f 2 ⇐ f 1 .
Now, an example of the application of the notion of confidence will be shown.

Example 3 From Example 1, the fuzzy attribute rule {a1 } ←(s,c) {a1 /0.5} can be
considered, where the confidence is computed as:

supp({a1 /0.5} ∪ {a1 })

c=
supp({a1 /0.5})
card({a1 }↓ )
supp({a1 }) |B| card({a1 }↓ )
= = =
supp({a1 /0.5}) card({a1 /0.5}↓ ) card({a1 /0.5}↓ )
|B|
card({b1 /0.5, b2 }) 1.5
= = = 0.75
card({b1 , b2 }) 2

This mapping satisfies several interesting properties as the following one, which
trivially holds.

Proposition 2 The confidence c of a fuzzy attribute rule verify 0 ≤ c ≤ 1.

The following section will present different properties of the introduced rules.

3.1 Properties of the Fuzzy sc-Attribute Rules

Given two ordering related fuzzy subsets of attributes, a trivial fuzzy attribute rule
always arises.
Proposition 3 Let f 1 , f 2 ∈ L 1A , where f 1 ≺ f 2 , and the rule f 1 ←(s,c) f 2 , we have
that c = 1.

supp(f1 ∪ f 2 ) supp( f 2 )
Proof The confidence of the rule can be expressed as supp(f2 )
= supp(f2 )
= 1. 

A relation in the subsets of attributes implies a relation in the confidences.

Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 111

Proposition 4 Let f 1 , f 2 , f 3 ∈ L 1A , where f 1  f 2 and f 1  f 3 , and the rules

f 2 ←(s12 ,c12 ) f 1 , f 3 ←(s13 ,c13 ) f 1 . If f 2  f 3 , then c13 ≤ c12 .

Proof The confidences c13 and c12 arise from Eq. 2:

supp(f1 ∪ f 2 ) supp( f 2 )
c12 = =
supp(f1 ) supp(f1 )
supp(f1 ∪ f 3 ) supp( f 3 )
c13 = =
supp(f1 ) supp(f1 )

Assuming that f 2  f 3 , we also know that supp(f3 ) ≤ supp( f 2 ). Applying this fact,
the following inequalities are equivalents:

supp(f3 ) ≤ supp( f 2 )
supp(f3 ) supp(f2 )
≤
supp( f 1 ) supp( f 1 )
c13 ≤ c12

This establishes that, if f 2  f 3 , then c13 ≤ c12 . 

Example 4 In the context of Example 1, we will see the previous property in the
rules {a1 } ←(s1 ,c1 ) {a1 /0.5} and {a1 , a3 } ←(s2 ,c2 ) {a1 /0.5}. Hence, we compute the
supports and confidences of both rules:

s1 = supp({a1 /0.5}) = 1
supp({a1 }) 0.75
c1 = = = 0.75
supp({a1 /0.5}) 1
s2 = supp({a1 /0.5}) = 1
supp({a1 , a3 }) 0.25
c2 = = = 0.25
supp({a1 /0.5}) 1

As we can see, c2 ≤ c1 . Now, we will consider other two rules that do not satisfy
the hypotheses in Proposition 4 and we will see that the thesis does not hold either.
Given the rules {a1 , a3 } ←(s3 ,c3 ) {a1 } and {a1 , a2 /0.5} ←(s4 ,c4 ) {a1 }, we have:

s3 = supp({a1 }) = 0.75
supp({a1 , a3 }) 0.25
c3 = = = 0.33
supp({a1 }) 0.75
s4 = supp({a1 }) = 0.75
supp({a1 , a2 /0.5}) 0.5
c4 = = = 0.67
supp({a1 /0.5}) 0.75
112 V. Liñeiro-Barea et al.

which implies that c4  c3 .

From Proposition 4, the following corollary arises.

Corollary 1 Let f 1 , f 2 , f 3 ∈ L 1A , where f 1 ≺ f 2 ≺ f 3 , and the rules f 2 ←(s,c) f 1 ,
f 3 ←(s  ,1) f 1 , we have that c = 1.

Proof Since f 2 ≺ f 3 , by Proposition 4, we obtain that 1 ≤ c, therefore, as 1 is the

greatest possible confidence, c = 1 holds. 

The following result shows how the confidence can be derived via transitivity,
based on the idea given in [17].
Theorem 1 Let f 1 , f 2 , f 3 ∈ L 1A , where f 1 ≺ f 2 ≺ f 3 , and the rules f 2 ←(s,c) f 1 ,
f 3 ←(s  ,c ) f 2 , f 3 ←(s,c ) f 1 , then we have that c · c = c .

Proof Let f 3 ←(s  ,c ) f 1 . The support of the new rule, s  can be obtained applying
Eq. 1:

s  = supp(f1 ) = s

The confidence of the new rule c arises from Eq. 2:

supp(f1 ∪ f 3 ) supp( f 3 )
c = =
supp(f1 ) supp(f1 )

Applying again Eq. 2, c and c are actually:

supp(f1 ∪ f 2 ) supp( f 2 )
c= =
supp(f1 ) supp(f1 )
supp(f ∪ f ) supp( f3 )
c =
2 3
=
supp(f2 ) supp(f2 )

If c and c are multiplied:

supp(f2 ) supp(f3 ) supp(f3 )

c · c = · = = c
supp( f 1 ) supp( f 2 ) supp( f 1 )

This establishes that s  = s and c = c · c . 

Finally, the following theorem extends the previous transitivity property.

Theorem 2 Let f 1 , f 2 , f 3 , f 4 ∈ L 1A , where f 1  f 2 , f 1  f 3 and f 3  f 4 , f 2  f 4
and the rules f 2 ←(s1 ,c1 ) f 1 , f 3 ←(s2 ,c2 ) f 1 , f 4 ←(s3 ,c3 ) f 3 and f 4 ←(s4 ,c4 ) f 2 , then
we obtain that c4 = c2 · c3 · c11 .
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 113

Proof Let f 4 ←(s4 ,c4 ) f 2 . The confidence of the new rule c4 arises from Eq. 2:

supp(f2 ∪ f 4 ) supp( f 4 )
c4 = =
supp(f2 ) supp(f2 )

In other hand, we know that:

1 supp(f1 ∪ f 3 ) supp(f3 ∪ f 4 ) supp(f1 )

c2 · c3 · = · ·
c1 supp(f1 ) supp(f3 ) supp( f 1 ∪ f 2 )
supp(f3 ) supp(f4 ) supp(f1 ) supp(f4 )
= · · = = c4
supp( f 1 ) supp( f 3 ) supp( f 2 ) supp( f 2 )

This establishes that c4 = c2 · c3 · 1

c1
. 
This property allows to remove cycles in the concept lattice, which will be fun-
damental in order to compute a minimal base of fuzzy sc-attribute rules. Note that a
concept lattice can be considered as a graph, in which the concepts are the vertices
and the edges are given by the relations among them in the Hasse diagram of the
concept lattice. This relation will be considered throughout this paper.

3.2 Fuzzy Sc-Attribute Rule Base

Given a context, the set of fuzzy attribute rules computed may be huge and it also may
include redundant and not interesting rules for the purpose they are being obtained.
In order to fix that, a base of rules should be obtained.
Definition 11 A fuzzy attribute rule f  ←(s,c) f is derived from a set of fuzzy
attribute rules M if a succession { f 1 , f 2 , · · · , f n } of different fuzzy subset of
attributes exists, such that f = f 1 , f  = f n , and f i+1 ←(si ,ci ) f i ∈ M or f i ←(si ,ci )
f i+1 ∈ M, for all i ∈ {1, . . . , n − 1}.
Example 5 Considering in Example 1 the set of fuzzy attribute rules M = {{a1 }
←(s12 =1,c12 =0.75) {a1 /0.5}, {a1 , a3 } ←(s23 =0.75,c23 =0.3̂) {a1 }}, we can derive a new rule
from it, {a1 , a3 } ←(s,c) {a1 /0.5}, with:

s = s12 = 1
c = c12 · c23 = 0.75 · 0.3̂ = 0.25

From the notion of derivation, the definition of minimal base is introduced.

Definition 12 A set of fuzzy rules T is a minimal base if the following properties
hold:
114 V. Liñeiro-Barea et al.

– Completeness. Any rule f 2 ←(s,c) f 1 ∈

/ T can be derived from the rules in T .
– Non-redundancy. No rule f 2 ←(s,c) f 1 ∈ T can be derived from the rest of rules
in T .
A minimal base of fuzzy rules T is formed by two special subsets, depending on
the confidence threshold considered: the set of fuzzy rules with confidence c = 1,
which is called fuzzy attribute implication base and, the set of rules with c < 1, which
is called fuzzy sc-attribute rule base.
Current papers in fuzzy attribute implications [26, 27] only consider fuzzy rules
that have the greatest confidence value (c = 1), however considering implication
with c < 1 is also interesting since some noise in the data can be mitigated and
outlier data are also considered. The following section will be focused on this kind
of implications.

4 Fuzzy sc-Intension Rules

This section defines, in a given concept lattice framework, an sc-attribute rule base
extending the classical case [17] and relates this base to the given concept lattice.
First of all, the notion of fuzzy sc-intension rule is introduced.
Definition 13 Given the set of all intents in the context Int(A, B, R, σ ) = { f ∈
L 1A | f = f ↓↑ }, the fuzzy rule f 2 ←(s,c) f 1 , in which the fuzzy subset of attributes
are intents, that is f 1 , f 2 ∈ Int( A, B, R, σ ), and f 1 ≺ f 2 is called fuzzy sc-intension
rule.
Note that Theorems 1 and 2 show how new fuzzy sc-intension rules can be derived
from a base via transitivity. For example, given f 1 , f 2 , f 3 ∈ Int( A, B, R, σ ), satisfy-
ing f 1 ≺ f 2 ≺ f 3 , and the rules f 2 ←(s,c) f 1 and f 3 ←(s  ,c ) f 2 , the rule f 3 ←(s,c ) f 1
can straightforwardly be derived considering c = c · c .
This result provides that only the rules between neighbor intents need to be consid-
ered in a base and gives us a mechanism in order to compute a fuzzy sc-intension rule
base, only considering a minimal subset of rules between neighbor intents removing
cycles in the concept lattice, in other words, obtaining a spanning tree of the Hasse
diagram of the concept lattice.
If we need to compute the confidence of a new rule, we just need to identify the
path followed in the spanning tree and then multiply the confidences of the involved
rules, inverting the value in the rules that we consider in the reverse form. This is
shown in the following example.

Example 6 Given a spanning tree S of a concept lattice M formed by the partial

implications { f 2 ←(0.5,0.4) f 1 , f 3 ←(0.4,0.3) f 1 , f 4 ←(0.2,0.25) f 2 }, the derived partial
implication f 4 ←(s,c) f 3 can be obtained transiting in the spanning tree as follows:
First of all, we go from the node related to f 3 to f 1 considering the rule f 3 ←(0.4,0.3)
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 115

f 1 , then we go from f 1 to f 2 by f 2 ←(0.5,0.4) f 1 and, finally, we pass from f 2 to f 3

considering the rule f 4 ←(0.2,0.25) f 2 .
Hence, the support of the derived rule is trivially supp(f3 ) and the confidence is
1
0.3
· 0.4 · 0.25 = 0.3̂.
Once a mechanism for deriving rules has been presented, the definition of fuzzy
sc-intension rule based is introduced.
Definition 14 A subset TL ⊆ { f 2 ←(s,c) f 1 | f 1 , f 2 ∈ Int( A, B, R, σ ) and f 1 ≺ f 2 }
is called fuzzy sc-intension rule base of the context if it is a minimal base.
As it was commented previously, this base is related to the notion of spanning
tree.
Proposition 5 Given a fuzzy sc-intension rule base of the context TL , the graph
V, E, where V = Int(A, B, R, σ ) and E = {( f 1 , f 2 ) | f 2 ←(s,c) f 1 ∈ TL }, is a
spanning tree of the concept lattice M associated with the context.
Since multiple spanning trees of a concept lattice M may exist, TL is not unique,
although the size of the base is fixed, as the following result shows.
Theorem 3 Given the concept lattice M associated with the context (A, B, R, σ ),
and a fuzzy sc-intension rule base TL of the context, we have that its size is always
|M| − 1.
Proof Applying Definition 14, the set of edges of the rules in the base forms a
spanning tree of M. As the number of edges of any spanning tree of a graph is
|V | − 1, begin V the number of nodes in the graph, the size of the spanning tree
associated with M is |M| − 1. 
The most interesting spanning trees are those associated with the rules with bigger
confidence. Given a minimal base TL , the spanning tree related to TL , in which each
edge is labeled with the confidence of the corresponding fuzzy sc-intension rule, will
be called TL -weighted spanning tree.

5 Application to Clustering

The base of fuzzy sc-intension rules has immediate application in a clustering concept
task. Given a concept lattice M associated with a context (A, B, R), we can naturally
see the confidence of a rule f 2 ←(s,c) f 1 as a ratio of similarity between the concepts
C2 =  f 2 ↓ , f 2  and C1 =  f 1 ↓ , f 1 . If we establish a threshold δ of confidence, we
can clustered the concepts of M obtaining a set of the similar ones. Note that one
concept can be part of more than one cluster. Hence, the obtained clustering can be
a covering [9, 30].
Given a concept C, the set of concepts which has a confidence greater or equal
than δ with respect to C will be called as the cover in M of C and will be denoted
as [C].
The process of clustering can be summarized in the following steps:
116 V. Liñeiro-Barea et al.

1. Obtaining a fuzzy sc-intension rule base of the context.

2. Traversing the fuzzy rules base from the top as follows:
– Considering two sets, the set of concepts M and the set in which the computed
clusters will be added, N , which is empty at the first.
– Step 1. Fix the top concept C1 ∈ M.
– Step 2. Consider in the cover of C1 , [C1 ], with respect to δ.
– Step 3. Add [C1 ] to N and remove the set [C1 ] from M.
– Step 4. After removing [C1 ] from M the rest of concepts form a meet-
semilattice. Apply Steps 3–4 to each maximal element C M1 , . . . , C Mn .
– Step 5. Apply Steps 2–4 until M be empty.
It is clear that depending on Step 4, the clustering provides a partition or a covering.

(a) If Step 4 is doing in a sequential process, that is, firstly Steps 2 and 3 are applied
to the maximal element C M1 , then to C M2 , and so on, a partition of M is obtained.
The main weakness is that this partition depend on the considered ordering in
the maximal elements.
(b) If Step 4 is doing in a parallel process, that is, firstly Step 2 is applied to
each maximal element C M1 , . . . , C Mn , and the different, not necessary disjoint
classes [C M1 ], . . . , [C Mn ] are computed, and then Step 3 is applied, a covering
is obtained. In this case, a partition is not obtained, in general.
In the following example, the first option of the clustering algorithm is applied to
a particular concept lattice.
Example 7 Following Example 1, we obtain the fuzzy sc-intension rule base TL
shown in Fig. 2, which provides a maximal TL -weighted spanning tree. The con-
fidence value of each rule is indicated in the edges of the diagram. In each node,
together with the extension and intension of the corresponding concept, a number is
written representing the label of the concept.
If we assume δ = 0.66, we obtain the following partition:
– [C1 ] = {1, 2}
– [C3 ] = {3}
– [C4 ] = {4}
– [C5 ] = {5}
– [C6 ] = {6}
– [C7 ] = {7}
In case of δ = 0.5, we obtain the following partition:
– [C1 ] = {1, 2, 3, 4}
– [C5 ] = {5}
– [C6 ] = {6}
– [C7 ] = {7}
Considering now δ = 0.25, the partition is:
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 117

Fig. 2 Fuzzy sc-intension

rule base related to the
context given in Example 1

– [C1 ] = {1, 2, 3, 4, 5, 6}
– [C7 ] = {7}
Note that, in this particular case, since a small concept lattice has been considered,
the algorithm with the second option (b) provides the same (partition) covering.
An interesting consequence of the proposed clustering mechanism is that it pro-
vides a reduction procedure based on the confidence of the fuzzy sc-intension rules.
This is possible because the set of classes always contains the minimum concept.
Moreover, we have that the concepts closer to the top concept provides greater confi-
dences. Therefore, the computation of the top part of the concept lattice is interesting
in order to obtain the most representative fuzzy sc-intension rules.

6 Conclusions and Future Work

A new kind of attribute implication has been introduced considering the fuzzy notions
of support and confidence, and the properties of these rules have been studied. These
rules allow to consider fuzzy rules that have a confidence value less than one, c < 1,
which is also interesting since some noise in the data can be mitigated and outlier
data are also considered. The particular case of using intensions in the rules provides
attribute implications with confidence less than 1, which can complement the usual
attribute implications with truth degree 1. Moreover, these rules provide a proce-
dure for clustering the concept lattices and, as a consequence, offer a size reduction
mechanism of the concept lattice.
In the following, a set of future work lines will be presented. First of all, the
performance and the potential of the procedure for information retrieval will be done
118 V. Liñeiro-Barea et al.

with an experimental survey. This will be the main line for future work, allowing to
search for optimizations in big data sets.
Extracting information from a database is the main task fuzzy rules are designed
for. Fuzzy rules can be utilized in Machine Learning scenarios such classification [13,
29] or prediction [12, 25]. Another line for future work will be to consider such
scenarios in order to study the potential of our work in them.

References

1. Agrawal, R., Imielinski, T., Swami, A.: Mining association rules between sets of items in large
databases. In: ACM SIGMOD International Conference on Management of Data, pp. 207–216
(1993)
2. Antoni, L., Krajci, S., Kridlo, O., Macek, B., Pisková, L.: On heterogeneous formal contexts.
Fuzzy Sets Syst. 234, 22–33 (2014)
3. Bělohlávek, R.: Lattices of fixed points of fuzzy Galois connections. Math. Logic Q. 47(1),
111–116 (2001)
4. Belohlavek, R., Cordero, P., Enciso, M., Mora, A., Vychodil, V.: Automated prover for attribute
dependencies in data with grades. Int. J. Approx. Reasoning 70, 51–67 (2016)
5. Belohlávek, R., Vychodil, V.: Attribute dependencies for data with grades ii. Int. J. Gen. Syst.
46(1), 66–92 (2017)
6. Butka, P., Pócs, J.: Generalization of one-sided concept lattices. Comput. Inform. 32(2), 355–
370 (2013)
7. Chertov, O., Aleksandrova, M.: Using association rules for searching levers of influence in
census data. Procedia Soc. Behav. Sci. 73(0), 475–478 (2013), In: Proceedings of the 2nd
International Conference on Integrated Information (IC-ININFO 2012), Budapest, Hungary,
30 Aug–3 Sept 2012
8. Ganter, B., Wille, R.: Formal Concept Analysis: Mathematical Foundations, Springer, New
York (2012)
9. Ge, X., Wang, P., Yun, Z.: The rough membership functions on four types of covering-based
rough sets and their applications. Inf. Sci. 390, 1–14 (2017)
10. Glodeanu, C.V.: Knowledge discovery in data sets with graded attributes. Int. J. Gen. Syst.
45(2), 232–249 (2016)
11. Hill, J., Walkington, H., France, D.: Graduate attributes: implications for higher education
practice and policy. J. Geogr. High. Educ. 40(2), 155–163 (2016)
12. Ikram, A., Qamar, U.: Developing an expert system based on association rules and predicate
logic for earthquake prediction. Knowl. Based Syst. 75, 87–103 (2015)
13. Kianmehr, K., Alhajj, R.: Carsvm: a class association rule-based classification framework and
its application to gene expression data. Artif. Intell. Med. 44(1), 7–25 (2008)
14. Krajči, S.: A generalized concept lattice. Logic J. IGPL 13(5), 543–550 (2005)
15. Kuhr, T., Vychodil, V.: Fuzzy logic programming reduced to reasoning with attribute implica-
tions. Fuzzy Sets Syst. 262, 1–20 (2015)
16. Lu, H., Han, J., Feng, L.: Stock movement prediction and n-dimensional inter-transaction
association rules. In: Proceedings of the ACM SIGMOD Workshop on Research Issues in Data
Mining and Knowledge Discovery, p. 12 (1998)
17. Luxenburger, M.: Implications partielles dans un contexte. Mathématiques, Informatique et
Sciences Humaines 29(113), 35–55 (1991)
18. Malerba, D., Lisi, F.A., Sblendorio, F.: Mining spatial association rules in census data: a rela-
tional approach. In: Proceeding of the ECML/PKDD?02 workshop on Mining Official Data,
University Printing House, Helsinki, Citeseer (2002)
Generating Fuzzy Attribute Rules Via Fuzzy Formal Concept Analysis 119

19. Medina, J., Ojeda-Aciego, M., Ruiz-Calviño, J.: Formal concept analysis via multi-adjoint
concept lattices. Fuzzy Sets Syst. 160(2), 130–144 (2009)
20. Nahar, J., Imam, T., Tickle, K.S., Chen, Y.-P.P.: Association rule mining to detect factors which
contribute to heart disease in males and females. Expert Syst. Appl. 40(4), 1086–1093 (2013)
21. Ordonez, C., Santana, C., de Braal, L.: Discovering interesting association rules in medical
data. In: Proceedings of ACM SIGMOD Workshop on Research Issues on Data Mining and
Knowledge Discovery, pp. 78–85 (2000)
22. Paranjape-Voditel, P., Deshpande, U.: An association rule mining based stock market rec-
ommender system. In: 2011 Second International Conference on Emerging Applications of
Information Technology (EAIT), pp. 21–24, Feb 2011
23. Popescu, A.: A general approach to fuzzy concepts. Math. Logic Q. 50(3), 265–280 (2004)
24. Rai, S., Sharma, S.: Determining minimum spanning tree in an undirected weighted graph.
In: 2015 International Conference on Advances in Computer Engineering and Applications
(ICACEA), pp. 637–642, Mar 2015
25. Rudin, C., Letham, B., Salleb-Aouissi, A., Kogan, E., Madigan, D.: Sequential event prediction
with association rules. In: 24th Annual Conference on Learning Theory (COLT 2011), pp. 615–
634, July 2011
26. Vychodil, V.: Computing sets of graded attribute implications with witnessed non-redundancy
(2015). arxiv: CoRRabs/1511.01640
27. Vychodil, V.: Rational fuzzy attribute logic (2015). arxiv: CoRRabs/1502.07326
28. Vychodil, V.: Computing sets of graded attribute implications with witnessed non-redundancy.
Inf. Sci. 351, 90–100 (2016)
29. Wang, W., Wang, Y., Bañares-Alcántara, R., Cui, Z., Coenen, F.: Application of classification
association rule mining for mammalian mesenchymal stem cell differentiation. In: Perner, P.
(ed.) Advances in Data Mining. Applications and Theoretical Aspects, Volume 5633 of Lecture
Notes in Computer Science, pp. 51–61. Springer, Berlin Heidelberg (2009)
30. Yang, B., Hu, B.O.: On some types of fuzzy covering-based rough sets. Fuzzy Sets Syst. 312,
36–65 (2017) (Theme: Fuzzy Rough Sets)
31. Zadeh, L.A.: Fuzzy logic. Computer 21(4), 83–93 (1988)