Behavior-Driven Model Design: A Deep Learning Recommendation Model Jointing Users and Products Reviews

Le Wang, Xi Zhao, Zhihao Shen

Abstract: Description Data-driven is widely mentioned, but the data is generated by user behavior. Our work aims to utilize a behavior-driven model design pattern to improve accuracy and provide explanations in review-based recommendations. Review-based recommendation introduces review text to overcome the sparseness and unexplainably of rating or scores-based model. Driven by users rating behavior and human cognitive abilities, we proposed a deep learning recommendation model jointing users and products reviews (DLRM-UPR) to learn user preferences and product characteristics adaptively. The DLRM-UPR consists of word, text, and context co-attention layers considering the interaction between each user-product-context pair. Extensive experiments on real datasets demonstrate that DLRM-UPR outperforms existing state-of-the-art models. In addition, the relevant information in the reviews and the suggestion for improving the user experience can be highlighted to explain the recommendation results.