Infinite mixtures for multi-relational categorical data

Large relational datasets are prevalent in many fields. We propose an unsupervised component model for relational data, i.e., for heterogeneous collections of categorical co-occurrences. The co-occurrences can be dyadic or n-adic, and over the same or different categorical variables. Graphs are a special case, as collections of dyadic co occurrences (edges) over a set of vertices. The model is simple, with only one latent variable. This allows wide applicability as long as a global latent component solution is preferred, and the generative process fits the application. Estimation with a collapsed Gibbs sampler is straightforward. We demonstrate the model with graphs enriched with multinomial vertex properties, or more concretely, with two sets of scientific papers, with both content and citation information available.