Towards fast algorithms for estimating Personalized PageRank using commonly generated random walks

摘要

Personalized PageRank (PPR) is a measure of the importance of the nodes in graph G = (V, E) from the perspective of some s E V . PPR has been used in many applications, such as recommending who a user should follow on Twitter. Due to the scale of networks of interest, precomputing PPR requires prohibitive storage, while computing exactly at query time is slow. As such, many PPR estimation algorithms have been proposed. In this extended abstract, we focus on the related PPR search problem, wherein one aims to estimate PPR of T = {v E V : v relevant to query from s}. We begin with a description of PPR and PPR search using Bidirectional-PPR. We provide a path interpretation of this algorithm, which leads us to define a class of similar algorithms. We conclude with two such algorithms intended for use in PPR search.