Joint inter-word and inter-sentence multi-relation modeling for summary-based recommender system

Review is an essential piece of information that influences users' decisions, but excessively long reviews not only degrade the user experience but also affect the accuracy of the recommender system. Therefore, Joint Inter-Word and Inter-Sentence Multi-Relation Modeling for the Summary-based Recommender System (MRSR) is proposed in this paper. In MRSR, the concise summary information serves as representation data, and a multi-relation modeling module is designed to construct user and item characteristics from two levels. Specifically, the inter-word relation modeling module, which consists of the Transformer and the pooling layer, is used to learn the long dependencies of summaries and extract word-level features by calculating the relative weights between words within sentences. The inter-sentence relation modeling module is used to enrich the sentence-level features of users and items, where an attention mechanism is employed to perceive the relative weights between different summary sentences. Finally, the fusion layer based on multi-head attention and the prediction layer based on attentional factorization machine are implemented to conduct the shallow and deep interactions between user and item features, based on which MRSR completes the final rating prediction. Extensive experimental results on five publicly available datasets demonstrate that MRSR achieves a 5.94% improvement in RMSE metrics compared to state-of-the-art methods. Furthermore, the accuracy of most existing models is improved by 1%similar to 2% while the inference time is reduced by 10% by utilizing summaries as representation data. It proves the efficiency and effectiveness of our proposed approach, which has promising application prospects.