EMRM: Enhanced Multi-source Review-Based Model for Rating Prediction

Xiaochen Wang, Tingsong Xiao, Jie Shao

Abstract: Rating prediction, whose goal is to predict user preference for unconsumed items, has become one of the core tasks in recommendation systems. Recently, many deep learning-based methods have been applied to the field of recommendation systems and have achieved great performance, especially when user reviews are available. User reviews usually contain rich semantic information and can reflect the preferences of users. However, user reviews are usually sparse. To alleviate this problem, we propose a method called EMRM, which stands for Enhanced Multi-source Review-based Model for rating prediction, to collect multi-source auxiliary reviews for each user. EMRM not only collects multi-source auxiliary reviews from nearest neighbors but also from farthest neighbors who have dissimilar consuming behaviors and historical rating records, so it can improve both the accuracy and diversity of recommendations. Our method extracts useful semantic information from user reviews and multi-source auxiliary reviews by applying Ordered-Neurons Long Short-Term Memory (ON-LSTM). Experimental results demonstrate that EMRM achieves better rating prediction accuracy than other baselines on three real-world datasets.