NRDPA: Review-Aware Neural Recommendation with Dynamic Personalized Attention

Review-based recommendation can utilize user and item features extracted from review text to alleviate the problems of data sparsity and poor interpretability. However, most existing methods focus on static modeling of user personality and item attributes while ignoring the dynamic changes of user and item features. Therefore, this paper proposes a neural recommendation method with dynamic personalized attention (NRDPA). First, this method captures the changes in user behavior at the word level and review level and models the personalized features of users and items by dynamically highlighting key words and important reviews. Second, the method considers information interaction in the process of user and item modeling and adjusts the feature representations of the interacting parties according to the user’s preferences for different items. Finally, experiments on five public datasets from Amazon demonstrate that the proposed NRDPA model has superior performance, with improvements of up to 10% in MSE and 6.3% in MAE compared to state-of-the-art models.