Bayesian Modeling of Random Walker for Community Detection in Networks

TL;DR

This paper introduces a Bayesian generative model utilizing random walks and advanced MCMC algorithms to improve community detection in networks, especially for overlapping communities, overcoming EM limitations.

Contribution

It develops a Bayesian model with MCMC algorithms for better community detection, surpassing EM-based methods in accuracy and robustness.

Findings

Gibbs samplers outperform previous methods in detecting overlapping communities.

Markovian dynamics are key to robust community detection.

Algorithms show high precision and robustness on synthetic and real networks.

Abstract

We propose a generative model to detect globally optimal community structures in networks by utilizing random walks. Sophisticated parameter optimization algorithms are developed based on the Markov chain Monte Carlo methods to overcome limitations of the EM algorithm, which has been used in previous works but is sometimes trapped in local optima depending on initial conditions. We apply the algorithms to synthetic and real-world networks to examine their performance in terms of precision and robustness of detected communities. It is found that the Gibbs samplers outperform the previous approaches especially in detecting overlapping communities. The Markovian dynamics of random walkers is crucial to robustly detect the optimal community structures.