Random walk with restart

From Cohen Courses
Revision as of 10:42, 3 September 2010 by WikiAdmin (talk | contribs) (1 revision)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

This is a technical method discussed in Social Media Analysis 10-802 in Spring 2010.

What problem does it address

Given a weighted graph, its often asked "how closely related are two nodes in a graph?". Random walk with restart (RWR) provides a good relevance score between two nodes in a weighted graph.

Algorithm

  • Input -
         A : Adjecency matrix
         q : Query node 
         c : probability of restarting the random walk from node q 
  • Output - u : steady state probability vector
 - Let v = query vector where all 0 except qth node = 1
 - A = Column_Norm(A) i.e. Normalize A such that each column sums to 1
 - Intialize u = v
 - while (u has not converged)
       u = (1-c)*A*u + c*v

There are many possible convergence criteria. e.g delta(norm(u)) < 1e-9

Used in

This technique is widely used practice. e.g personalized pagerank, automatic image captioning etc.

Relevant Papers