A Bagging-based ensemble method for recommendations under uncertain rating data

Abstract

Matrix factorization (MF) is one of the most-used techniques to build recommender systems. However, in practical use, the existence of noise in the training set brings some uncertainty, degrading the performance of MF approaches. In this paper, we propose a Bagging-based MF framework, an ensemble method of using multiple MF-based models to improve the stability and accuracy. Specifically, our framework first rebuilds new training sets by resampling on the original ratings, then takes advantage of the sets to train MF models and finally combines the predictions from the models in a ensemble way. The experiment results on real data show that our framework can achieve some performance improvement when having noise samples.