This paper provides a principled probabilistic co-clustering framework for missing value prediction and pattern discovery in users’ preference data. We extend the original dyadic formulation of the
Block Mixture Model(BMM)
in order to take into account explicit users’ preferences.
simultaneously identifies user communities and item categories: each user is modeled as a mixture over user communities, which is computed by taking into account users’ preferences on similar items. Dually, item categories are detected by considering preferences given by similar minded users. This recursive formulation highlights the mutual relationships between items and user, which are then used to uncover the hidden block-structure of the data. We next show how to characterize and summarize each block cluster by exploiting additional meta data information and by analyzing the underlying topic distribution, proving the effectiveness of the approach in pattern discovery tasks.