Rating Prediction Based on Merge-CNN and Concise Attention Review Mining

Abstract

Online review websites provide an open platform for users to write reviews or give ratings on items (business services) as well as share their consumption experience. However, the volume of reviews is large, while the rating scores provide users with a quick picture of the items without reading all reviews. Recommendation systems can help users find items of interest by predicting user’s ratings on unrated items. Review contents contain more personalized preference features than simply user ratings. Therefore, it is important to consider both ratings and review contents when making rating predictions. This research proposes a novel approach that combines deep learning and review mining with attention mechanism for rating predictions. Review mining with attention mechanism is adopted to extract concise attention reviews with important words and sentences. A merge convolutional neural network (merge-CNN) model is proposed to consider both the target user’s preference features and performance features of target business for rating prediction. This method extracts quality business performance features from the quality reviews written by elite (credible) users. Moreover, the proposed method uses the concise attention reviews of target user’s neighbors to simulate target users’ reviews on unrated target business. Experiments were conducted on Yelp data sets to evaluate our proposed methods. The results show that the proposed method, i.e. considering concise attention reviews and quality reviews written by elite users, outperforms traditional methods in improving prediction accuracy. The experiment result also shows that our review simulation methods can well simulate target user’s reviews on unrated target business.

Keywords

• Recommender system
• rating prediction
• matrix factorization
• review mining
• deep learning