Computing Heat Kernel Pagerank and a Local Clustering Algorithm

TL;DR

This paper introduces a fast, sublinear time algorithm for approximating heat kernel pagerank in graphs, enabling efficient local clustering by identifying low-conductance cuts using heat kernel pagerank vectors.

Contribution

The work presents the first sublinear time algorithm for approximating heat kernel pagerank and integrates it into a local clustering method for graph partitioning.

Findings

Algorithm runs in time $O(rac{ ext{log}(rac{1}{ ext{epsilon}}) ext{log} n}{ ext{epsilon}^3 ext{log} ext{log}(rac{1}{ ext{epsilon}})})$

Heat kernel pagerank can identify low-conductance cuts with many seed vertices

Effective local clustering achieved through heat kernel pagerank approximation

Abstract

Heat kernel pagerank is a variation of Personalized PageRank given in an exponential formulation. In this work, we present a sublinear time algorithm for approximating the heat kernel pagerank of a graph. The algorithm works by simulating random walks of bounded length and runs in time $O\big(\frac{\log(\epsilon^{-1})\log n}{\epsilon^3\log\log(\epsilon^{-1})}\big)$, assuming performing a random walk step and sampling from a distribution with bounded support take constant time. The quantitative ranking of vertices obtained with heat kernel pagerank can be used for local clustering algorithms. We present an efficient local clustering algorithm that finds cuts by performing a sweep over a heat kernel pagerank vector, using the heat kernel pagerank approximation algorithm as a subroutine. Specifically, we show that for a subset $S$ of Cheeger ratio $\phi$, many vertices in $S$ may serve as seeds for a heat kernel pagerank vector which will find a cut of conductance $O(\sqrt{\phi})$.