An adaptive category-aware recommender based on dual knowledge graphs

Combining the knowledge graph (KG) with the personalized item recommendation has become an important method to improve user experience. In the personalized item recommendation, users have their preferences on categories that influence their choices of items. In order to fully use category information, we explicitly focus on their impact on user preference and run through the whole recommendation process. We construct two dual knowledge graphs (KG-UI and -UC). Based on them, we propose KG-CICEF, a recommendation system based on knowledge graph aggregation and user preference modeling. Our model effectively captures user preferences for explored and unexplored item categories by aggregating information from two types of knowledge graphs. We convert user preference over unexplored item categories to the crossitem-category exploration factor (CEF). We utilize CEF to build a category-wise loss function for the item recommendation. For consistency, we also propose a category-based negative sampling mechanism to optimize this loss function. Experimental results on three benchmark datasets demonstrate that KG-CICEF achieves significant improvements over the state-of-the-art methods, and the case study validates the effectiveness of CEF in item recommendations.