UC San Diego

A recommender system uses information from a user's past behavior to present items of interest to him. A fundamental problem in recommender systems is approximating a full user-item matrix where most of the entries are missing. The rows of the matrix represent the users and the columns represent the items. The entries indicate the plausibility that the user will enjoy the item. In this thesis the items are movies and the entries ratings. In this thesis I compare three statistical models that how a user will rate a movie. The first two are Bernoulli models that predict whether a rating is greater than three out of five. The first Bernoulli model uses logistic regression. The second Bernoulli model is a latent factor model. The third model extends the latent factor model to use a five class multinomial. A five class multinomial is chosen to predict a rating on a scale of one to five. The results show that latent factor model that uses a Bernoulli distribution has a better accuracy than a model trained by logistic regression. The latent factor model is extended to use a multinomial. The accuracy of variants of the multinomial model are evaluated. A technique to initialize the multinomial model is shown to improve the accuracy. However the accuracy is lower than other models used in the Netflix competition. The Bernoulli and multinomial latent factor models are compared against each other. The Bernoulli model is more accurate