Deep neural review text interaction for recommendation systems

Abstract Users’ reviews contain valuable information which are not taken into account in most recommender systems. According to the latest studies this field, using review texts could only improve performance of recommendation, but it can also alleviate impact data sparsity and help tackle problem. In paper, we present a neural model recommends items by leveraging user reviews. order predict rating for each item, our proposed model, named M t c h P y r m i d R e o n S s (MPRS), represents item with their corresponding texts. Thus, problem recommendation is viewed as text matching such that score obtained from be considered good representative joint extent similarity. To solve problem, inspired MatchPyramid (Pang et al., 2016), employed an interaction-based approach according matrix constructed given pair input The matrix, has property hierarchical patterns, then fed Convolutional Neural Network (CNN) compute user–item pair. Our experiments on small categories Amazon dataset show gains 1.76% 21.72% relative improvement compared DeepCoNN (Zheng 2017), 0.83% 3.15% TransNets (Catherine Cohen, 2017). Also, two large categories, namely AZ-CSJ AZ-Mov , achieves improvements 8.08% 7.56% 1.74% 0.86% respectively.