Effective and Efficient Conductance-based Community Search at Billion Scale

TL;DR

This paper introduces a conductance-based community search problem and proposes an efficient four-stage algorithm, SCCS, to find high-quality communities in billion-scale graphs, addressing limitations of previous methods.

Contribution

The paper formulates the conductance-based community search problem, proves its NP-hardness, and develops a scalable four-stage algorithm, SCCS, for billion-scale graph community detection.

Findings

SCCS effectively finds communities with low conductance.

The method scales to billion-scale graphs.

Experiments demonstrate high effectiveness and efficiency.

Abstract

Community search is a widely studied semi-supervised graph clustering problem, retrieving a high-quality connected subgraph containing the user-specified query vertex. However, existing methods primarily focus on cohesiveness within the community but ignore the sparsity outside the community, obtaining sub-par results. Inspired by this, we adopt the well-known conductance metric to measure the quality of a community and introduce a novel problem of conductance-based community search (CCS). CCS aims at finding a subgraph with the smallest conductance among all connected subgraphs that contain the query vertex. We prove that the CCS problem is NP-hard. To efficiently query CCS, a four-stage subgraph-conductance-based community search algorithm, SCCS, is proposed. Specifically, we first greatly reduce the entire graph using local sampling techniques. Then, a three-stage local optimization strategy is employed to continuously refine the community quality. Namely, we first utilize a seeding strategy to obtain an initial community to enhance its internal cohesiveness. Then, we iteratively add qualified vertices in the expansion stage to guarantee the internal cohesiveness and external sparsity of the community. Finally, we gradually remove unqualified vertices during the verification stage. Extensive experiments on real-world datasets containing one billion-scale graph and synthetic datasets show the effectiveness, efficiency, and scalability of our solutions.