Sequential Monte Carlo Inference Based on Activities for Overlapping Community Models

Abstract

Various kinds of data such as social media can be represented as a network or graph. Latent variable models using Bayesian statistical inference are powerful tools to represent such networks. One such latent variable network model is a Mixed Membership Stochastic Blockmodel (MMSB), which can discover overlapping communities in a network and has high predictive power. Previous inference methods estimate the latent variables and unknown parameters of the MMSB on the basis of the whole observed network. Therefore, dynamic changes in network structure over time are hard to track. Thus, we first present an incremental Gibbs sampler based on node activities that focuses only on observations within a fixed term length for online sequential estimation of the MMSB. We further present a particle filter based on node activities with various term lengths. For instance, in an e-mail communication network, each particle only considers e-mail accounts that sent or received a message within a specific term length, where the length may be different from those of other particles. We show through experiments with two link prediction datasets that our proposed methods achieve both high prediction performance and computational efficiency.