Poisson Matrix Factorization
Probabilistic models for Bayesian recommender systems.
Recommendation systems are a vital component of the modern web. They help readers navigate otherwise unwieldy archives of information and help websites direct users to items, movies, articles, songs, and products that they will like. Collaborative filtering is one technique for building them, inferring user preferences and item attributes from data.
In this project I studied various models for Bayesian recommender systems, including Poisson Matrix Factorization and its extensions, Hierarchical Poisson Matrix Factorization and Bayesian Non-parametric Poisson Matrix Factorization. I analyzed the effect of latent dimensions on the models, and learned how auxiliary variables in variational inference make the models locally conjugate to facilitate inference. I evaluated their performance on the MovieLens 1M dataset.