HRFA: Don’t Ignore Strangers with Different Views

Review-based recommender suffers from the sparsity of reviews: only a few users leave substantial comments in the real world. As a result, some recent methods resort to supplementary reviews written by similar users, which only leverage homogeneous preferences. However, users holding different views could also supply valuable information with heterogeneous preferences. In this paper, we propose a recommendation model for rating prediction, named Heterogeneous Review-based Recommendation via Four-way Attention (HRFA). To take advantage of the heterogeneous preferences, the supplementary reviews in HRFA are redefined as reviews from all users with common purchase history, no matter whether they give similar ratings. Specially, we integrate the heterogeneous preferences into the one semantic space via introducing a similarity projection based on rating difference. Experiments conducted on five datasets demonstrate that our model achieves higher rating prediction accuracy than other baselines.