Predicting NBA games with matrix factorization

Abstract

In my thesis, I present the methods I use to predict NBA games using matrix factorization. Matrix factorization is popular through the Netflix recommendation problem, but in general, one can apply it to data that are best modeled as the result of pairwise interaction. My thesis contains three parts. First, I explain how I model NBA prediction as a matrix factorization problem and use the basic low-rank matrix factorization approach to discover structure in the data. I also explain some differences between using matrix factorization for NBA prediction versus that in the Netflix recommendation problem. Second, I use probabilistic matrix factorization (PMF) to incorporate the fact that when two teams play each other, the scores will be different each time. Lastly, I incorporate supplementary information such as the date of the game by combining multiple PMF problems using Gaussian process priors. I replace the scalar latent features with functions of this supplementary information to aid with prediction.