Probabilistic Matrix Factorization (PMF)
Ruslan Salakhutdinov and Andriy Mnih, University of Toronto, Poster ID T21

· Many approaches to collaborative filtering have difficulty handling very large datasets and dealing with users who have entered very few ratings. · We present the PMF model that scales linearly with the number of observations and performs well on the Netflix dataset, which is large, sparse, and imbalanced. · We propose two extensions to the PMF model: ­ Adding adaptive priors on the model parameters to provide automatic model capacity control; ­ Constraining the PMF model using the rated/unrated information to make it learn similar preferences for users who have rated similar sets of movies. This extension significantly improves model performance for users with very few ratings. · We achieve an error rate of 0.8861 (nearly 7% better than the score of Netflix's own system) by linearly combining the predictions of multiple PMF and Restricted Boltzmann Machine models.