Relational Thematic Clustering with Mutually Preferred Neighbors

Abstract

Automatically learning thematic clusters in network data has long been a challenging task in machine learning community. A number of approaches have been proposed to accomplish it, utilizing edges, vertex features, or both aforementioned. However, few of them consider how the quantification of dichotomous inclination w.r.t. network topology and vertex features may influence vertex-cluster preferences, which deters previous methods from uncovering more interpretable latent groups in network data. To fill this void, we propose a novel probabilistic model, dubbed Relational Thematic Clustering with Mutually Preferred Neighbors (RTCMPN). Different from prevalent approaches which predetermine the learning significance of edge structure and vertex features, RTCMPN can further learn the latent preferences indicating which neighboring vertices are more possible to be in the same cluster, and the dichotomous inclinations describing how relative significance w.r.t. edge structure and vertex features may impact the association between pairwise vertices. Therefore, cluster structure implanted with edge structure, vertex features, neighboring preferences, and vertex-vertex dichotomous inclinations can be learned by RTCMPN. We additionally derive an effective Expectation-Maximization algorithm for RTCMPN to infer the optimal model parameters. RTCMPN has been compared with several strong baselines on various network data. The remarkable results validate the effectiveness of RTCMPN.