Bayesian Optimization for Recommender System

Abstract

Many web services have a Recommender System to help the users in their choices such as movies to watch or products to buy. The aim is to make accurate predictions on the user preferences depending on his/her past choices. Matrix-factorization is one of the most widely adopted method to build a Recommender System. Like many Machine Learning algorithms, matrix-factorization has a set of hyper-parameters to tune, leading to a complex expensive black-box optimization problem. The objective function maps any possible hyper-parameter configuration to a numeric score quantifying the quality of predictions. In this work, we show how Bayesian Optimization can efficiently optimize three hyper-parameters of a Recommender System: number of latent factors, regularization term and learning rate. A widely adopted acquisition function, namely Expected Improvement, is compared with a variant of Thompson Sampling. Numerical for both a benchmark 2-dimensional test function and a Recommender System evaluated on a benchmark dataset proved that Bayesian Optimization is an efficient tool for improving the predictions of a Recommendation System, but a clear choice between the two acquisition function is not evident.