Ian Porteous, Evgeniy Bart, Max Welling
Matrix factorization algorithms are frequently used in the machine learning community to find low dimensional representations of data. We introduce a novel generative Bayesian probabilistic model for unsupervised matrix and tensor factorization. The model consists of several interacting LDA models, one for each modality. We describe an efficient collapsed Gibbs sampler for inference. We also derive the nonparametric form of the model where interacting LDA models are replaced with interacting HDP models. Experiments demonstrate that the model is useful for prediction of missing data with two or more modalities as well as learning the latent structure in the data.
Subjects: 12. Machine Learning and Discovery; 3.4 Probabilistic Reasoning
Submitted: Apr 14, 2008