Item recommendation task predicts a personalized ranking over a set of itemsfor individual user. One paradigm is the rating-based methods that concentrateon explicit feedbacks and hence face the difficulties in collecting them.Meanwhile, the ranking-based methods are presented with rated items and thenrank the rated above the unrated. This paradigm uses widely available implicitfeedback but it usually ignores some important information: item reviews. Itemreviews not only justify the preferences of users, but also help alleviate thecold-start problem that fails the collaborative filtering. In this paper, wepropose two novel and simple models to integrate item reviews into matrixfactorization based Bayesian personalized ranking (BPR-MF). In each model, wemake use of text features extracted from item reviews via word embeddings. Ontop of text features we uncover the review dimensions that explain thevariation in users' feedback and these review factors represent a priorpreference of a user. Experiments on real-world data sets show the benefits ofleveraging item reviews on ranking prediction. We also conduct analyses tounderstand the proposed models.
展开▼