Discriminative infinite latent feature models

Latent feature models (LFMs) have been widely used to model ordinal rating data and relational network data in various tasks such as collaborative filtering and link prediction, typically in a generative way. Alternatively, one might incorporate max-margin learning into the model via the principle of Maximum Entropy Discrimination (MED) to learn a more discriminative latent feature space that favors the supervised learning task. Another dimension to extend LFMs is to employ Bayesian nonparametric methods to make LFMs self-adaptive to the number of latent features, which is crucial for model complexity control. In this paper we review several recent progresses that have been made in the above two extensions for the task of collaborative filtering and link prediction.