Interpretable Robust Recommender Systems with Side Information

Abstract

In this paper, we propose two methods, namely Trace-norm regression (TNR) and Stable Trace-norm Analysis (StaTNA), to improve performances of recommender systems with side information. Our trace-norm regression approach extracts low-rank latent factors underlying the side information that drives user preference under different context. Furthermore, our novel recommender framework StaTNA not only captures latent low-rank common drivers for user preferences, but also considers idiosyncratic taste for individual users. We compare performances of TNR and StaTNA on the MovieLens datasets against state-of-the-art models, and demonstrate that StaTNA and TNR in general outperforms these methods.