Dynamic relational topic model for social network analysis with noisy links

A probabilistic framework is presented for joint analysis of text and links between nodes (e.g., people) in a time-evolving social network. Unlike existing approaches, the proposed model is able to handle “noisy” links, i.e., observed links between nodes for which there is limited or no similarity in the associated text. This decoupling between links and text is made possible by incorporating random effects in the probabilistic model, and leads to improved text modeling and link prediction performance. The model allows efficient inference using fully conjugate Gibbs sampling, obviating the need for any maximum-likelihood parameter setting. Experiments are conducted using scientific paper citation and co-authorship network datasets, with the proposed approach outperforming previous state-of-the-art results.