info-cluster 0.5.post1

a hierachical clustering algorithm based on information theory

Python binding

How to build

The binding uses Cython. To package the library, use python setup.py bdist_wheel. Install the package by pip install --user info_cluster. Below is the prebuilt binary packages:

Platform	py3.6	py3.7
Windows	T	T
MacOS	T	T
Linux	T	T

Demo code

We provide a high-level wrapper of info-clustering algorithm. After installing info_cluster, you can use it as follows:

from info_cluster import InfoCluster
import networkx as nx
g = nx.Graph() # undirected graph
g.add_edge(0, 1, weight=1)
g.add_edge(1, 2, weight=1)
g.add_edge(0, 2, weight=5)
ic = InfoCluster(affinity='precomputed') # use precomputed graph structure
ic.fit(g)
ic.print_hierachical_tree()

The output is like

      /-0
   /-|
--|   \-2
  |
   \-1

import psp # classify the three data points shown in the above figure
g = psp.PyGraph(3, [(0,1,1),(1,2,1),(0,2,5)]) # index started from zero, similarity is 5 for vertex 0 and 2
g.run() # use maximal flow algorithm to classify them
print(g.get_critical_values()) # [2,5]
print(g.get_category(2)) # get the result which has at least 2 categories, which is [0,1,0]

Parametric Dilworth Truncation(pdt) implementation

We provide another alternative implementation, which can be used similar to PyGraph. To make it work, you should apply a patch preflow.patch to preflow.h before building, which belongs to lemon library 1.3.1, see #625.

import psp
g = psp.PyGraphPDT(3, [(0,1,1),(1,2,1),(0,2,5)]) # index started from zero, similarity is 5 for vertex 0 and 2
g.run() # use maximal flow algorithm to classify them
print(g.get_critical_values()) # [2,5]
print(g.get_category(2)) # get the result which has at least 2 categories, which is [0,1,0]