Personalized PageRank community detection

pprcd.py

import collections
import sys

# setup the graph
G = {
  1:set([ 2, 3, 5, 6,]),
  2:set([ 1, 4,]),
  3:set([ 1, 6, 7,]),
  4:set([ 2, 5, 7, 8,]),
  5:set([ 1, 4, 6, 8, 9, 10,]),
  6:set([ 1, 3, 5, 7,]),
  7:set([ 3, 4, 6, 9,]),
  8:set([ 4, 5, 9,]),
  9:set([ 5, 7, 8, 20,]),
  10:set([ 5, 11, 12, 14, 15,]),
  11:set([ 10, 12, 13, 14,]),
  12:set([ 10, 11, 13, 14, 15,]),
  13:set([ 11, 12, 15,]),
  14:set([ 10, 11, 12, 25,]),
  15:set([ 10, 12, 13,]),
  16:set([ 17, 19, 20, 21, 22,]),
  17:set([ 16, 18, 19, 20,]),
  18:set([ 17, 20, 21, 22,]),
  19:set([ 16, 17,]),
  20:set([ 9, 16, 17, 18,]),
  21:set([ 16, 18,]),
  22:set([ 16, 18, 23,]),
  23:set([ 22, 24, 25, 26, 27,]),
  24:set([ 23, 25, 26, 27,]),
  25:set([ 14, 23, 24, 26, 27,]),
  26:set([ 23, 24, 25,]),
  27:set([ 23, 24, 25,]),
}
Gvol = 102

# G is graph as dictionary-of-sets
alpha=0.99
tol=0.01
seed=[1]

x = {} # Store x, r as dictionaries
r = {} # initialize residual
Q = collections.deque() # initialize queue
for s in seed: 
  r[s] = 1/len(seed)
  Q.append(s)
while len(Q) > 0:
  v = Q.popleft() # v has r[v] > tol*deg(v)
  if v not in x: x[v] = 0.
  x[v] += (1-alpha)*r[v]
  mass = alpha*r[v]/(2*len(G[v])) 
  for u in G[v]: # for neighbors of u
    assert u is not v, "contact [email protected] for self-links"
    if u not in r: r[u] = 0.
    if r[u] < len(G[u])*tol and \
       r[u] + mass >= len(G[u])*tol:
       Q.append(u) # add u to queue if large
    r[u] = r[u] + mass
  r[v] = mass*len(G[v]) 
  if r[v] >= len(G[v])*tol: Q.append(v)
print str(x)

  
# Find cluster, first normalize by degree
for v in x: x[v] = x[v]/len(G[v])  
# now sort x's keys by value, decreasing
sv = sorted(x.iteritems(), key=lambda x: x[1], reverse=True)
S = set()
volS = 0.
cutS = 0.
bestcond = 1.
bestset = sv[0]
for p in sv:
  s = p[0] # get the vertex
  volS += len(G[s]) # add degree to volume
  for v in G[s]:
    if v in S:
      cutS -= 1
    else:
      cutS += 1
  print "v: %4i  cut: %4f  vol: %4f"%(s, cutS,volS)
  S.add(s)
  if cutS/min(volS,Gvol-volS) < bestcond:
    bestcond = cutS/min(volS,Gvol-volS)
    bestset = set(S) # make a copy
print "Best set conductance: %f"%(bestcond)
print "  set = ", str(bestset)

Hey, got here from a really interesting talk of yours. Quick note, I think you want to change to r[s] = 1.0 / len(seed) here, otherwise the residuals will all be zero when one has multiple seeds.

Thank you for your sharing. Do you have any publication for this implementation? How do you know the improvement of this model with Personalized PageRank compared with the normal community detection?

This is the ACL method. See Andersen Chung, Lang, FOCS 2006. Many studies have validated this as a good algorithm, e.g. see Kloumann and Kleinberg who study PPR for seeded community detection, or any of our papers on Neighborhood Inflated Seed Expansion -- there is no best algorithm.