Just a small, simple Python package I made to help people simulate, import, and analyze network data -- specifically, network data represented in the form of call detail records (CDR) aggregated up to the day level.
Why? Well, it turns out getting actual CDR data is incredibly difficult and a lot of people simply want the experience of playing with "real" network data. This package will help create files that look and feel like real CDR files using an underlying network that is known to (and specified by) the user. It's derived mostly from my own simulation exercise, but enough people expressed interest that I'm putting it up here in a refactored, probably-still-buggy form.
It uses Python 2.7 and hasn't been tested in 3+.
CDRs come in a wide range of formats and styles, but the most basic (aggregated) CDR will look something like this:
20130101;26;357;9;40.7;23;9
20130101;52;504;15;73.3;30;24
20130101;89;158;24;64.6;22;14
20130101;235;447;19;122.7;30;12
20130101;293;849;15;18.7;31;12
Representing date, caller id, recipient id, number of calls, minutes, text messages, and mms messages, respectively.
Each column is delimited by a semicolon (;) and missingness is represented by a blank space ( ). Each column is integer
except for minutes which is float.
If you're lucky, you'll sometimes get an attribute file which will look like this:
1;1554;F;46
2;1179;M;45
3;4795;M;46
4;2066;F;
5;1356;F;59
Representing caller id, zip code, sex, and age, respectively. Again, delimited by semicolon with a space to represent a missing
value.
Download all files as a zip and put the cdrhelper directory (the subfolder containing __init__.py) into your working directory. There may be some issues on Windows machines --- looking into a workaround.
from cdrhelper import *
Note, if you have actual postcode or age populatoin data, use the sourcefile option for both commands.
postcodes = generatePostcode()
ageweight = generatePopulation()
There are two ways to generate call information. The first way (easy way) is just a wrapper for the second way.
The second way offers more control. At the end of either process, you will have G the underlying network, df a dataframe of call information, and attr attribute information for each node.
G, df, attr = makeData(postcodes = postcodes, ageweight = ageweight)
Note, makeData() has quite a few parameters that allow you to control most of the interesting bits.
def makeData(postcodes, ageweight, nodes = 30, edges = 10, days = 10,
callsperday = 20, startdate = "20130101",
mean_calls = 5, call_dur = 1.15,
mean_sms = 25, mean_mms = 10,
r_prob = .33, seed = None,
maleid = "M", femaleid = "F"):
Make a graph representing your "ground truth" network --- i.e., the network in the whole world.
G = nx.barabasi_albert_graph(n = 1000, m = 25)
This step isn't necessary, but in my experience, CDR indices are hashed or start at 1, never 0. Remove it for authenticity.
G.remove_node(0)
Now make a dataframe with mobile interactions and dates. Note, you may not be able to get back to your "ground truth" network if your observations of interactions are sparse (relative to the size of the network). This is true in the real world as well where you may only get data from a single telecommunications company and cannot observe the "true" network.
df = generateDateCallers(G = G, days = 30, callsperday = 10,
startdate = "20130101")
Now populate this dataframe with random interactions (calls, texts, sms)
df = generateCallData(data = df, mean_calls = 20, call_dur = 1.15,
mean_sms = 30, mean_mms = 15)
Finally, for the call dataset, reciprocate some of these calls.
df = reciprocate(df_call = df, r_prob = .4)
Now we need to generate attribute information. You can add more information (e.g., height, weight, etc), but this is the minimum requirement for the rest of the module to work.
attr = pd.DataFrame(data = G.nodes(), columns = ['A_num'])
attrrows = len(attr)
attr['postcode'] = np.random.choice(postcodes['Postal Code'], attrrows)
attr['gender'] = np.random.choice(["F", "M"], attrrows)
attr['age'] = np.random.choice(range(18, 106),
p = ageweight, size = attrrows)
Most CDR attribute data will contain (sometimes significant) missingness in the attribute data -- missingness in the call data is possible, but less common. Add some missingness to the attribute data by specifying what proportion of missingness there should be for each column.
attr = insertMissing(attr = attr, p_post = .1, p_age = .15, p_gender = .2)
folderCheck("../data")
exportAttrData(attr, filename = "../data/myfakeattrdata.txt")
exportCallData(df, filename = "../data/myfakecalldata.txt")
Once you have data, you can import it in two ways -- (1) as a NetworkX object or (2) as a pandas dataframe.
from cdrhelper import *
This will import three NetworkX objects --- (1) Nodes which is a node-only network, (2) G which is an undirected graph,
and (3) D which is a directed graph. Obviously, only D and G will have edge information. Note that the undirected graph,
G, is not the same as turning D into an undirected graph using NetworkX. Specifically, my code will take the sum
of weights between (u, v) and (v, u) directed edges while the NetworkX code just takes the weight of the first (u, v)
edge.
Nodes = importNodes("../data/myfakeattrdata.txt")
G = importEdges("../data/myfakecalldata.txt", Nodes)
D = importEdges("../data/myfakecalldata.txt", Nodes, directed = True)
Pretty straightforward. Use attr.head() and call.head() to see the basic structure.
attr = importAttr("../data/myfakeattrdata.txt")
call = importCalls("../data/myfakecalldata.txt")
I'll keep building the analyze portion of the module as I get more time, but here are some basic things you can do now:
olap = overlapDistribution(G)
relativeLWCCsize(D)
summaryStats(attr)
## Directed network stats
DNetworkSummary(D, qtr = 1)
## Undirected network stats
GNetworkSummary(G, qtr = 1)
## Get a list of nodes that subsets out age and sex (see ?agexsexsubset)
agexsex = agexsexsubset(G)