Latent Dirichlet Bayesian Co-Clustering

Co-clustering has emerged as an important technique for mining contingency data matrices. However, almost all existing coclustering algorithms are hard partitioning, assigning each row and column of the data matrix to one cluster. Recently a Bayesian co-clustering approach has been proposed which allows a probability distribution membership in row and column clusters. The approach uses variational inference for parameter estimation. In this work, we modify the Bayesian co-clustering model, and use collapsed Gibbs sampling and collapsed variational inference for parameter estimation. Our empirical evaluation on real data sets shows that both collapsed Gibbs sampling and collapsed variational inference are able to find more accurate likelihood estimates than the standard variational Bayesian co-clustering approach.