A Novel Approach with Multi-view Features for Recommendation

Recommender system is a major technology for information filtering. Sparseness of rating data is one of the vital factors that deteriorate its quality. Features of auxiliary textual information, such as reviews or synopses, extracted more efficiently can effectively alleviate data sparsity, and better boost prediction accuracy of recommender system. However, due to the inherent limitation of traditional probability matrix factorization (PMF), they have difficulties in effectively utilizing contextual information. This paper proposes a novel multi-view feature extraction model, i.e. SLCPMF, that integrates textual convolutional neural network (TCNN) and latent dirichlet allocation (LDA) to comprehensively capture semantic features of items from reviews. In addition, as an unsupervised neural network, stacked denoising autoencoder (SDAE) can compress the user-item rating data and then reconstruct them, which can obtain users' latent feature. Extensive experiments show that compared with PMF, CDL and ConvMF both root mean square error (RMSE) and mean absolute error (MAE) of the proposed model are reduced on two Movielens datasets and Amazon dataset. Experimental results indicate that the SLCPMF model is effective and feasible in the field of recommender systems.