NRSA: Neural Recommendation with Summary-Aware Attention

Abstract

Reviews are widely used in recommendation systems to handle the sparsity problem of rating matrix. However, learning the representations of users and items only from reviews would be challenging since there are less meaningful words and reviews while modeling users or items. In fact, in addition to reviews there are rich off-the-shelf summaries written by users along with reviews, but existing recommendation methods ignore this useful information. The summary of a review is to describe the review with shorter sentences, and can be seen as a high-level abstraction of the review. Thus the summary can play a guidance role to indicate the important parts in a review and the informativeness of the review. Hence, we propose a neural recommendation method to learn summary-aware representations of users and items from reviews. We firstly apply a summary encoder to learn representations of text summary, which will be used as the guidance indicator. We design a summary-aware review encoder to learn representations of reviews from raw words, and another summary-aware user/item encoder to learn representations of users or items from reviews. To be specific, we propose a hierarchical attention model with summary representations as attention vectors under word- and review-level to select important words and reviews for users/items respectively. We conduct extensive experiments on real-world benchmark datasets and the results demonstrate that our approach can effectively improve the performance of neural recommendation.