Exploiting review embedding and user attention for item recommendation

As a valuable source of user preferences and item properties, reviews have been widely leveraged in many approaches to enhance the performance of recommender systems. Although encouraging success has been obtained, there are two more weaknesses need to be addressed. (1) Most approaches represent users or items merely based on the modeling of review texts, but ignore the potential and latent preferences beyond textual information. (2) Existing methods tend to blindly merge all the previous reviews for user profiling. However, it may be less effective because different interacted items may play distinct roles. Hence, indiscriminately aligning interacted items may limit the model flexibility and performance. In this paper, with the desire to fill these gaps, we design a novel attentive deep review-based recommendation method. In specific, we complement the item representation by an auxiliary vector, based on which a user is then attentively profiled by her posted items to predict the likeness for the target item. Extensive experiments on five real-world datasets demonstrate that our model can not only significantly outperform the state-of-the-art methods, but also provide intuitive explanations to the recommendations.