Deep fusion explainable recommendation based on sentiment analysis and feature weighting

Recommendation systems are crucial in helping users search for personalized information. However, the original ratings used by existing recommendation systems can't accurately represent users' true rating intentions, ignore sentiment attitudes, and lack explainability. To address this issue, the article proposes the recommendation of Sentiment Analysis and Feature Weighting (SAFW). This method uses BERT pre-trained review text to generate comment word vectors. Using Bidirectional Recurrent Neural Network (BiRNN) to extract sentiment features of users towards items and generate sentiment classification labels to reduce the deviation between user ratings and their true sentiment tendencies. Decompose the sentiment score matrix and the original rating matrix into vectors, perform deep fusion of multi-level feature vectors, and model high-order interactions between users and items to improve the accuracy of recommendations. Calculate the weight of users' interest preferences for item features based on review frequency and sentiment score. On the premise of high recommendation accuracy, using the three features with the highest interest preference weight as the explanation for the recommendation makes it more persuasive.