Knowledge-aware recommendation model with dynamic co-attention and attribute regularize

As important information provided by recommender systems, knowledge graphs are widely applied in computer science and many other fields. The recommender system performance can be significantly improved by leveraging the knowledge graph between the user and item. Various recommendation approaches have been proposed based on the knowledge graph in recent years; however, most of the existing models only apply item-level user representations or attention mechanisms to users and items in the same way and ignore the fact that user and item attributes are significantly different. Hence, these models are not an effectively exploited attribute information and circumscribe the further improvement of recommender performance. In this paper, a novel approach of dynamic co-attention with an attribute regularizer (DCAR) for a knowledge-aware recommender system is proposed to explore the latent connections between the user level and item level. The model dynamically adjusts the dynamic co-attention mechanism through the attribute similarity between the target user and the candidate item. Specifically, an attribute regularizer between user and item is designed to improve the quality of attribute embedding. Experimental results on two realistic datasets show that our proposed model can significantly improve recommender system effectiveness and represents an advancement beyond the compared deep models.