Active Learning with Bayesian Nonnegative Matrix Factorization for Recommender Systems

Gonul Ayci, Abdullatif Köksal, M. Melih Mutlu, Burak Suyunu, A. Taylan Cemgil

Abstract: Active learning is a method of analyzing the observed data such that choosing the next observation will give the most information about the variable to be predicted. However, when observations are costly, one needs strategies to obtain informative data to arrive at accurate predictions with less data. In this study, we compare various observation sequence selection strategies on the matrix completion problem. We used Gibbs Sampling and Variational Bayes as inference mechanisms on the MovieLens dataset. Our results suggest that the Gibbs sampler coupled with the selection of the element with minimal observations on a row and column is the superior approach for the Bayesian Nonnegative Matrix Factorization (NMF).