Learning Factor Selection for Boosted Matrix Factorisation in Recommender Systems

Matrix factorisation (MF), an effective recommendation algorithm, predicts user preferences on items based on the historical preferences of other like-minded users. Classical MF methods do not explicitly distinguish the significances across the underlying factors that determine a users' preference on an item. The identical contribution of latent factors during learning results unnecessary updates on unimportant variables that leads to slower and suboptimal convergence. In this paper, we propose a new matrix factorisation method that not only seeks the intrinsic and outstanding factors that determine the users' preferences but also systematically reinforces the contribution generated by these factors. Based on boosting, a factor selection mechanism is developed to account the variable importance of latent factors to generate an ensemble recommender on the selected subspace of the latent factors by the principle of model uncertainty reduction. The proposed method is evaluated against a variety of the state-of-the-art methods of recommender systems on three publicly available benchmark datasets. The results confirm the effectiveness and efficiency of the proposed method.