Research On Recommendation Models Based On Multi-Granular Attention Networks

Due to the explosive growth of information,Recommendation Systems emerged as a technology to solve the problem of information overload.Recommendation Systems not only effectively mine potential user interests but also bring huge commercial value to the enterprise.Unlike traditional recommendation models that are limited by problems such as sparse rating data,the review-based recommendation model additionally considers valuable user review information,and has thus received widespread attention in recent years.However,most existing methods only extract the semantic features of reviews at the word granularity,regardless of the context,ignoring the enormous noisy review data that is irrelevant to the prediction task,and failing to capture the potential user preferences changes.As a result,it is difficult for existing methods to model user preferences fully and accurately.In order to solve the above problems,this paper proposes a Dynamic Multi-Granularity Attention Network Recommendation Model(D-MG-Attn).Specifically,the model simulates the order of human understanding,following the order of word granularity,sentence granularity,and review granularity to learn the semantic features of user and item reviews in a step-by-step manner.While constructing the semantic features,the model incorporates attention mechanism at each granularity to learn the importance of words,sentences,and reviews.Furthermore,based on the importance,the multi-granularity features are weighted and added step by step to obtain the user/item features,effectively filtering out the noise data in the reviews.Subsequently,user features are input into the gated recurrent unit in time sequence to model dynamic user features.Finally,the item features and dynamic user features are input into the factorization machine to model the high-order relationship among the features and predict the rating score.Based on the above improvements,the performance of D-MG-Attn on rating prediction can be effectively improved.In addition,from the perspective of interpretability,the importance of multi-granularity learned by D-MG-Attn can highlight the most important words,sentences,and reviews,which is helpful for attribution analysis of recommendation results.This paper designs and conducts extensive comparison experiments based on 24 datasets published by Yelp and Amazon.The experimental results show that the model proposed by this paper outperforms state-of-the-art review-based recommendation models on the mean square error for the rating prediction problem.It verifies the effectiveness and necessary of the multigranular attention mechanism and the consideration of dynamic user preferences.In addition,this paper also uses a case analysis method to qualitatively evaluate the importance of reviews that affect the prediction results.The analysis results show that the method proposed by this paper has good interpretability.