Study On Review-aware Recommendation Method Using Mutual Attention Mechanism

The rapid development of Internet technology has sped the explosive growth of information,which brought about the problem of information overload.The recommendation system can recommend content that may be of interest to users,thereby alleviating the problem of information overload.The traditional collaborative filtering technology in the recommendation system only uses the rating matrix,and the integration of side information can provide more information for the recommendation system,so the fusion of multiple side information and rating matrices has become a popular trend.Among the side information,the amount of review text is relatively rich,which can effectively reflect the user's interest preferences and the attributes of the item.Therefore,the review-aware recommendation methods have become the current research hotspot.Existing researches show that deep models can better learn the characteristics of users and items from reviews.The attention mechanism is widely used in the two-tower model because it can recognize important text in reviews.However,the existing works based on the attention mechanism often train item nets and user nets independently,and only perform feature fusion in the last layer,which weakens the correlation between users and items.In concern with the above problems,this paper proposes a review-aware recommendation algorithm based on mutual attention mechanism.Firstly,a mutual attention module is utilized in the algorithm that combines review features and nonreview hidden features.It uses the non-review hidden features of users and items to bridge the two networks through a mutual attention mechanism,thereby strengthening the mutual influence of the training process via the two networks.Taking the user network as an example,in the process of modeling user review text,we use the personalized attention mechanism and the mutual attention mechanism to select the two important tokens that better reflect the user's personalized preferences and the influence of items on the user.The item network modeling process is similar.It is worth noting that we also consider the impact of different granular texts on representation learning,so we use hierarchical attention mechanisms to perform word-level and review-level modeling respectively.Secondly,a mutual attention module that integrates user-side reviews and item-side reviews is built in the algorithm,and the mutual influence of reviews on both sides is considered in the modeling to mark important reviews on the user-side and itemside.Finally,we use the factorization machine model in the output layer to fuse the features output by these two modules and perform rating prediction.In order to verify the effectiveness of the method proposed in this paper,we conduct experiments on three subsets of Amazon's product datasets and compare them with the existing state-of-the-art comment perception methods.The experimental results show that this method has the best recommendation performance.Further degradation experiments show that the mutual attention mechanism of bridging the two networks plays an important role in improving the model performance which strengthens the interaction between users and items,and helps to further refine the feature expression of users and items.