Explainable Recommendation Merging Sentiment Analysis and Multi-Scale Collaborative Filtering
Teng Chang, Bohui Li, Zhixia Zhang, Xingjuan Cai · 2024
Recommender systems (RSs) based on collaborative filtering (CF) are often perceived as black-box models incapable of offering plausible explanations for recommended items. CF that combines methods such as deep learning mitigates this problem but explains only in terms of feature similarity, which neglects the user-item correlation and users’ real sentiment. In order to improve the explainability of CF models and the rationality of explanation, an explainable recommendation model merging sentiment analysis and multi-scale collaborative filtering (ERSA-MCF) is proposed in this paper. Specifically, we extend the scale of CF to implement filtering in three aspects: user acceptance, user preference, and item similarity, to increase the association between users and items. Simultaneously, the sentiment score of conformity ratings is introduced to mirror the user’s genuine receptiveness. In order to balance the recommended lists generated by the three filtering scales with accuracy, diversity, and explainability, a multi-objective particle swarm optimization algorithm based on a global optimal quality centroid (MPSO-GOC) is proposed for optimizing the candidate lists and improving the overall efficiency of the model. Experimental results show that the model improves diversity and explainability while ensuring accuracy.