GridRep is a feature transformation-mapping tool that wraps DBSCAN and enables more efficient clustering.
It is effective for large volumes of low-cardinality input data containing multiple repeating unique sets of feature values.
The GridRep transformer generates a representative input subset based on DBSCAN's min_samples parameter that participates in the clustering procedure. The generated labels can then be re-mapped back to the original input data.
For data with high-cardinality, the GridRep transformer allows the mitigation of potential false precision (e.g. lots of meaningless decimals) by passing a rounding_decimals parameter value.
ClippedDBSCAN wraps the FeaturesTransformer around sklearn's DBSCAN, in a sklearn.pipeline compatible Estimator.
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.datasets.samples_generator import make_blobs
from sklearn.preprocessing import FunctionTransformer
from sklearn.cluster import DBSCAN
import numpy as np
from gridrep.cluster import ClippedDBSCANcenters = [(-2, -2), (0, 0), (4.2, 5)]
X, _ = make_blobs(n_samples=20000, centers=centers, n_features=2, random_state=0)
radius = 0.1
min_samples = 7
round_decimals = 1
# ClippedDBSCAN
pipeline_clip = make_pipeline(StandardScaler(),
ClippedDBSCAN(eps=radius,
min_samples=min_samples,
round_decimals=round_decimals))
# DBSCAN
pipeline_noClip = make_pipeline(StandardScaler(),
FunctionTransformer(np.round,
validate=False,
kw_args={"decimals": round_decimals}),
DBSCAN(eps=radius, min_samples=min_samples))%%timeit
pipeline_noClip.fit_predict(X)317 ms ± 9.21 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
pipeline_clip.fit_predict(X)36.4 ms ± 479 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)