MSCA: a multi-scale context-aware recommender system leveraging reviews and user interactions

Purpose Reviews, serving as authentic user feedback, encompass rich semantic information, including descriptions of user preferences and item characteristics. Introducing reviews as auxiliary information into recommendation systems can enhance the modeling of user and item feature representations. However, existing methods insufficiently account for the semantic variations of review words across different contextual scales, often leading to suboptimal representations of review information. In addition, they fail to effectively integrate review features with interaction features, resulting in a semantic gap between the two modalities. To address these issues, this paper aims to propose a novel review-based recommendation model that incorporates a multi-scale context-aware network and a cross-attention mechanism (MSCA). Design/methodology/approach MSCA uses a multi-scale convolutional network to capture the semantic representations of words in reviews across different contextual scales. It then uses a self-attention layer to learn the associations among reviews, enhancing the representation of review features. Subsequently, MSCA deploys a cross-attention mechanism to emulate personalized attention across different users and items, effectively fusing interaction features with the semantic features of reviews. Findings Experiments on five Amazon review data sets demonstrate that the proposed model outperforms baseline methods in terms of evaluation metrics. Originality/value The authors propose MSCA, a novel recommendation model that more effectively extracts review features using a multi-scale context-aware network. It also uses a cross-attention mechanism to fuse semantic information with interaction features from two modalities, thereby improving the performance of recommendation systems.