Ensemble-Based Discovery of Disjoint, Overlapping and Fuzzy Community Structures in Networks

TL;DR

This paper introduces two novel ensemble methods, ENDISCO and MEDOC, for community detection in networks, capable of identifying disjoint, overlapping, and fuzzy communities with improved accuracy and robustness over existing algorithms.

Contribution

The paper presents the first ensemble methods for fuzzy and overlapping community detection, demonstrating superior performance and robustness compared to existing algorithms.

Findings

Both algorithms outperform existing standalone community detection methods.

ENDISCO and MEDOC outperform existing ensemble algorithms in accuracy and runtime.

Ensemble methods aid in exploring core-periphery structures and stable communities.

Abstract

Though much work has been done on ensemble clustering in data mining, the application of ensemble methods to community detection in networks is in its infancy. In this paper, we propose two ensemble methods: ENDISCO and MEDOC. ENDISCO performs disjoint community detection. In contrast, MEDOC performs disjoint, overlapping, and fuzzy community detection and represents the first ever ensemble method for fuzzy and overlapping community detection. We run extensive experiments with both algorithms against both synthetic and several real-world datasets for which community structures are known. We show that ENDISCO and MEDOC both beat the best-known existing standalone community detection algorithms (though we emphasize that they leverage them). In the case of disjoint community detection, we show that both ENDISCO and MEDOC beat an existing ensemble community detection algorithm both in terms of multiple accuracy measures and run-time. We further show that our ensemble algorithms can help explore core-periphery structure of network communities, identify stable communities in dynamic networks and help solve the "degeneracy of solutions" problem, generating robust results.