Skip to content

Instantly share code, notes, and snippets.

@ntamas
Created November 19, 2013 09:54
Show Gist options
  • Save ntamas/7542961 to your computer and use it in GitHub Desktop.
Save ntamas/7542961 to your computer and use it in GitHub Desktop.
Hierarchical link clustering algorithm of Ahn et al (see http://barabasilab.neu.edu/projects/linkcommunities/), implemented using Python and igraph for fun.
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Hierarchical link clustering
============================
:Author: Tamás Nepusz
This is an implementation of the hierarchical link clustering algorithm of Ahn
et al. The results provided by this implementation match those of the original
Python implementation of the authors, but it is somewhat faster. At least I
hope so. Also, it handles all the input graph formats that igraph_ handles.
.. _igraph: http://igraph.sourceforge.net
"""
from __future__ import division
from collections import defaultdict
from igraph import load
from itertools import izip
from operator import itemgetter
from optparse import OptionParser
from textwrap import dedent
import logging
import sys
__author__ = "Tamas Nepusz"
__license__ = "MIT"
__docformat__ = "restructuredtext en"
__version__ = "0.1"
class JaccardSimilarityCalculator(object):
"""Calculates pairwise Jaccard similarities on a given unweighted
graph. When calculating the similarities, it is assumed that every
vertex is linked to itself.
"""
def __init__(self, graph):
self._adjlist = [set(graph.neighbors(i)).union([i])
for i in xrange(graph.vcount())]
def get_similarity(self, v1, v2):
"""Returns the Jaccard similarity between the two given vertices,
assuming that both of them are linked to themselves."""
set1, set2 = self._adjlist[v1], self._adjlist[v2]
isect = len(set1.intersection(set2))
return isect / (len(set1) + len(set2) - isect)
class TanimotoSimilarityCalculator(object):
"""Calculates pairwise Tanimoto coefficients on a given weighted
graph. When calculating the similarities, it is assumed that every
vertex is linked to itself with an edge whose weight is equal to the
average weight of edges adjacent to the vertex."""
def __init__(self, graph, attr="weight"):
degrees = graph.degree()
strengths = graph.strength(weights=attr)
weights = graph.es[attr]
self._adjedgelist = []
get_eid = graph.get_eid # prelookup
for i in xrange(graph.vcount()):
weis = dict((j, weights[get_eid(i, j)]) for j in graph.neighbors(i))
weis[i] = strengths[i] / degrees[i]
self._adjedgelist.append(weis)
self._sqsums = [sum(value * value for value in vec.itervalues())
for vec in self._adjedgelist]
def get_similarity(self, v1, v2):
"""Returns the Tanimoto coefficient of the two given vertices,
assuming that both of them are linked to themselves."""
vec1, vec2 = self._adjedgelist[v1], self._adjedgelist[v2]
if len(vec1) > len(vec2):
# vec1 must always be the smaller
vec1, vec2 = vec2, vec1
numerator = sum(value * vec2.get(key, 0)
for key, value in vec1.iteritems())
return numerator / (self._sqsums[v1] + self._sqsums[v2] - numerator)
class EdgeCluster(object):
"""Class representing a group of edges (i.e. a group of vertices
in the line graph)
This class also keeps track of the original vertices the edges
refer to."""
__slots__ = ("vertices", "edges")
def __init__(self, vertices, edges):
self.vertices = set(vertices)
self.edges = set(edges)
def is_smaller_than(self, other):
"""Compares this group of edges with another one based on
size."""
return len(self.edges) < len(other.edges)
def partition_density(self):
"""Returns the number of edges times the relative density
of this group. This value is used in the calculation of
the overall partition density, used to select the best
threshold."""
m, n = len(self.edges), len(self.vertices)
if n <= 2:
return 0.
return m * (m-n+1) / (n-2) / (n-1)
def merge_from(self, other):
"""Merges another group of edges into this one, updating
self.vertices and self.edges"""
self.vertices |= other.vertices
self.edges |= other.edges
def __repr__(self):
return "%s(%r, %r)" % (self.__class__.__name__,
self.vertices, self.edges)
class EdgeClustering(object):
"""Class representing an edge clustering of a graph as a whole.
This class is essentially a list of `EdgeCluster` instances
plus some additional bookkeeping to facilitate the easy lookup
of the cluster of a given edge.
"""
def __init__(self, edgelist):
"""Constructs an initial edge clustering of the given graph
where each edge belongs to its own cluster.
The graph is given by its edge list in the `edgelist`
parameter."""
self.clusters = [EdgeCluster(edge, (i, ))
for i, edge in enumerate(edgelist)]
self.membership = range(len(edgelist))
self.d = 0.0
def lookup(self, edge):
"""Returns the cluster of a given edge"""
return self.clusters[self.membership[edge]]
def merge_edges(self, edge1, edge2):
"""Merges the clusters corresponding to the given edges."""
cid1, cid2 = self.membership[edge1], self.membership[edge2]
# Are they the same cluster?
if cid1 == cid2:
return
cl1, cl2 = self.clusters[cid1], self.clusters[cid2]
# We will always merge the smaller into the larger cluster
if cl1.is_smaller_than(cl2):
cl1, cl2 = cl2, cl1
cid1, cid2 = cid2, cid1
# Save the partition densities
dc1, dc2 = cl1.partition_density(), cl2.partition_density()
# Merge the smaller cluster into the larger one
for edge in cl2.edges:
self.membership[edge] = cid1
cl1.merge_from(cl2)
self.clusters[cid2] = cl1
# Update D
self.d += cl1.partition_density() - dc1 - dc2
def partition_density(self):
"""Returns the overall partition density of the clustering."""
return self.d * 2.0 / len(self.membership)
class HLC(object):
"""Hierarchical link clustering algorithm on a given graph.
This class implements the algorithm outlined in Ahn et al: Link communities
reveal multiscale complexity in networks, Nature, 2010. 10.1038/nature09182
The implementation supports undirected and unweighted networks only at the
moment, and it is assumed that the graph does not contain multiple or loop
edges. This is not ensured within the class for sake of efficiency.
The class provides the following attributes:
- `graph` contains the graph being analysed
- `min_size` contains the minimum size of the clusters one is interested
in. It is advised to set this to at least 3 (which is the default value)
to ensure that pseudo-clusters containing only two nodes do not turn up
in the results.
The algorithm may be run with or without a similarity threshold. When no
similarity threshold is passed to the `run()` method, the algorithm will
scan over the possible range of similarities and return a partition that
corresponds to the similarity with the highest partition density. In this
case, the similarity threshold and the partition density is recorded in
the `last_threshold` and `last_partition_density` attributes. The former
is also set properly when a single similarity threshold is used.
"""
def __init__(self, graph = None, min_size = 3):
"""Constructs an instance of the algorithm. The algorithm
will be run on the given `graph` with the given minimum
community size `min_size`."""
self._graph = None
self._edgelist = None
self.last_threshold = None
self.last_partition_density = None
self.graph = graph
self.min_size = int(min_size)
@property
def graph(self):
"""Returns the graph being clustered."""
return self._graph
@graph.setter
def graph(self, graph):
"""Sets the graph being clustered."""
self._graph = graph
self._edgelist = graph.get_edgelist()
def run(self, threshold = None):
"""Runs the hierarchical link clustering algorithm on the
graph associated to this `HLC` instance, cutting the dendrogram
at the given `threshold`. If the threshold is `None`, the
optimal threshold will be selected using the partition density
method described in Ahn et al, 2010. Returns a generator that
will generate the clusters one by one.
"""
if threshold is None:
return self._run_iterative()
else:
return self._run_single(threshold)
def get_edge_similarity_graph(self):
"""Calculates the edge similarity graph of the graph assigned
to this `HLC` instance."""
# Construct the line graph
linegraph = self.graph.linegraph()
# Create an adjacency list representation (we already have an edgelist)
# and select the appropriate similarity function
if "weight" in self.graph.edge_attributes():
similarity = TanimotoSimilarityCalculator(self.graph).get_similarity
else:
similarity = JaccardSimilarityCalculator(self.graph).get_similarity
# For each edge in the line graph, compute a similarity score
scores = []
append, edgelist = scores.append, self._edgelist # prelookups
for edge in linegraph.es:
(a, b), (c, d) = edgelist[edge.source], edgelist[edge.target]
if a == c:
append(similarity(b, d))
elif a == d:
append(similarity(b, c))
elif b == c:
append(similarity(a, d))
else: # b == d
append(similarity(a, c))
linegraph.es["score"] = scores
return linegraph
def _run_single(self, threshold):
"""Runs the hierarchical link clustering algorithm on the
graph associated to this `HLC` instance, cutting the dendrogram
at the given threshold. Returns a generator that will generate the
clusters one by one.
:Parameters:
- threshold: the level where the dendrogram will be cut
"""
# Record the threshold in last_threshold
self.last_threshold = threshold
self.last_partition_density = None
# Construct the edge similarity graph
linegraph = self.get_edge_similarity_graph()
# Remove unnecessary edges
linegraph.es(score_le=threshold).delete()
# Process the connected components of the linegraph and build the result
clusters = linegraph.clusters()
result = [set() for _ in xrange(len(clusters))]
for edge, cluster_index in izip(self._edgelist, clusters.membership):
result[cluster_index].update(edge)
return (list(cluster) for cluster in result
if len(cluster) >= self.min_size)
def _run_iterative(self):
"""Runs the hierarchical link clustering algorithm on the given graph,
cutting the dendrogram at the place where the average weighted partition
density is maximal. Returns a generator that will generate the clusters
one by one.
:Parameters:
- graph: the graph being clustered
- min_size: minimum size of clusters
"""
# Construct the line graph
linegraph = self.get_edge_similarity_graph()
# Sort the scores
sorted_edges = sorted(linegraph.es, key=itemgetter("score"),
reverse=True)
# From now on, we only need the edge list of the original graph
del linegraph
# Set up initial configuration: every edge is a separate cluster
clusters = EdgeClustering(self._edgelist)
# Merge clusters, keep track of D, find maximal D
max_d, best_threshold, best_membership = -1, None, None
prev_score = None
merge_edges = clusters.merge_edges # prelookup
for edge in sorted_edges:
score = edge["score"]
if prev_score != score:
# Check whether the current D score is better than the best
# so far
if clusters.d >= max_d:
max_d, best_threshold = clusters.d, score
best_membership = list(clusters.membership)
prev_score = score
# Merge the clusters
merge_edges(edge.source, edge.target)
del clusters
max_d *= 2 / self.graph.ecount()
# Record the best threshold and partition density
self.last_threshold = best_threshold
self.last_partition_density = max_d
# Build the result
result = defaultdict(set)
for edge, cluster_index in izip(self._edgelist, best_membership):
result[cluster_index].update(edge)
return (list(cluster) for cluster in result.itervalues()
if len(cluster) >= self.min_size)
def hlc_single(graph, threshold = 0.66, min_size = 3):
"""Runs the hierarchical link clustering algorithm on the given graph,
cutting the dendrogram at the given threshold. Returns a generator
that will generate the clusters one by one.
:Parameters:
- graph: the graph being clustered
- threshold: the level where the dendrogram will be cut
- min_size: minimum size of clusters
"""
# Construct the line graph
linegraph = graph.linegraph()
# Create an adjacency list and an edge list representation
edgelist = graph.get_edgelist()
adjlist = [set(graph.neighbors(i)).union([i])
for i in xrange(graph.vcount())]
# For each edge in the line graph, compute a similarity score
scores = []
for edge in linegraph.es:
(a, b), (c, d) = edgelist[edge.source], edgelist[edge.target]
if a == c:
scores.append(jaccard(adjlist[b], adjlist[d]))
elif a == d:
scores.append(jaccard(adjlist[b], adjlist[c]))
elif b == c:
scores.append(jaccard(adjlist[a], adjlist[d]))
else: # b == d
scores.append(jaccard(adjlist[a], adjlist[c]))
# Assign the similarity scores, remove unnecessary edges
linegraph.es["score"] = scores
linegraph.es(score_le=threshold).delete()
# Process the connected components of the linegraph and build the result
clusters = linegraph.clusters()
result = [set() for _ in xrange(len(clusters))]
for edge, cluster_index in izip(edgelist, clusters.membership):
result[cluster_index].update(edge)
return (list(cluster) for cluster in result if len(cluster) >= min_size)
class HierarchicalLinkClusteringApp(object):
"""\
Usage: %prog [options] input_file
Runs a hierarchical link clustering on the given graph according
to the method of Ahn et al.
"""
def __init__(self):
self.parser = OptionParser(usage=dedent(self.__doc__).strip())
self.parser.add_option("-f", "--format", metavar="FORMAT",
dest="format",
help="assume that the input graph is in the given FORMAT")
self.parser.add_option("-s", "--min-size", metavar="K",
dest="min_size", default=3,
help="print only clusters containing at least K nodes")
self.parser.add_option("-t", "--threshold", metavar="THRESHOLD",
dest="threshold", type=float,
help="use the given THRESHOLD to cut the dendrogram. "
"If not specified, the threshold will be determined "
"automatically.")
self.parser.add_option("-q", "--quiet", action="store_true",
dest="quiet", help="quiet mode, print the result only")
self.parser.add_option("-W", "--no-weights", action="store_true",
dest="no_weights", help="ignore edge weights even if they "
"are present in the original graph")
self.log = logging.getLogger("hlc")
self.options = None
def run(self):
"""Runs the application."""
self.options, args = self.parser.parse_args()
level = logging.WARNING if self.options.quiet else logging.INFO
logging.basicConfig(level=level, format="%(message)s")
if not args:
self.parser.print_help()
sys.exit(1)
for arg in args:
self.process_file(arg)
def process_file(self, filename):
"""Loads a graph from the given file, runs the clustering
algorithm on it and prints the clusters to the standard
output."""
self.log.info("Processing %s..." % filename)
graph = load(filename, format=self.options.format)
# If the graph has weights and we want to ignore them, delete them
if self.options.no_weights and "weight" in graph.edge_attributes():
del graph.es["weight"]
# If the graph is directed, we have to make it undirecteed
if graph.is_directed():
graph.to_undirected(combine_edges="sum")
self.log.warning("Converted directed graph to undirected.")
# Set up the "name" attribute properly
if "label" in graph.vertex_attributes():
graph.vs["name"] = graph.vs["label"]
del graph.vs["label"]
elif "name" not in graph.vertex_attributes():
graph.vs["name"] = [str(i) for i in xrange(graph.vcount())]
# Run the algorithm, get the result generator
self.log.info("Calculating clusters, please wait...")
algorithm = HLC(graph, self.options.min_size)
results = algorithm.run(self.options.threshold)
# Print the optimal threshold if we determined it automatically
if self.options.threshold is None:
self.log.info("Threshold = %.6f" % algorithm.last_threshold)
self.log.info("D = %.6f" % algorithm.last_partition_density)
# Print the results
for community in results:
print "\t".join(graph.vs[community]["name"])
def main():
"""Main entry point for the application when run from the command line"""
return HierarchicalLinkClusteringApp().run()
if __name__ == "__main__":
sys.exit(main())
@odoluca
Copy link

odoluca commented Jan 1, 2018

Hello ntamas,
I wanted to compare this with the original python code but I can not seem to find it. Could you let me know where did you find the original code? Cheers
Osman

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment