Hybrid Learning Model with Barzilai-Borwein Optimization for Context-aware Recommendations

We propose an improved learning model for non-negative matrix factorization in the context-aware recommendation. We extend the collective non-negative matrix factorization through hybrid regularization method by combining multiplicative update rules with Barzilai-Borwein optimization. This provides new improved way of learning factorized matrices. We combine ratings, content features, and contextual information in three different 2-dimensional matrices. We study the performance of the proposed method on recommending top-N items. The method was empirically tested on 4 datasets, including movies, music, and mobile apps, showing an improvement in comparison with other state-of-the-art for top-N recommendations, and time convergence to the stationary point for larger datasets.