Distributed Algorithms for Finding Local Clusters Using Heat Kernel Pagerank

TL;DR

This paper introduces a distributed algorithm for efficiently finding local clusters in massive graphs, leveraging heat kernel pagerank to achieve logarithmic or sublinear runtime depending on the model.

Contribution

The paper presents the first distributed algorithms for local cluster detection that operate efficiently in both CONGEST and k-machine models, utilizing heat kernel pagerank.

Findings

Algorithm runs in logarithmic time in the CONGEST model.

Algorithm runs in sublinear time in the k-machine model.

Performance depends only on approximation error bounds, not graph size.

Abstract

A distributed algorithm performs local computations on pieces of input and communicates the results through given communication links. When processing a massive graph in a distributed algorithm, local outputs must be configured as a solution to a graph problem without shared memory and with few rounds of communication. In this paper we consider the problem of computing a local cluster in a massive graph in the distributed setting. Computing local clusters are of certain application-specific interests, such as detecting communities in social networks or groups of interacting proteins in biological networks. When the graph models the computer network itself, detecting local clusters can help to prevent communication bottlenecks. We give a distributed algorithm that computes a local cluster in time that depends only logarithmically on the size of the graph in the CONGEST model. In particular, when the value of the optimal local cluster is known, the algorithm runs in time entirely independent of the size of the graph and depends only on error bounds for approximation. We also show that the local cluster problem can be computed in the k-machine distributed model in sublinear time. The speedup of our local cluster algorithms is mainly due to the use of our distributed algorithm for heat kernel pagerank.