Multi-granularity label-aware user interest modeling for news recommendation

Jianxing Zheng, Min Li, Suge Wang, Jian Liao, Xiaoya Wan

Abstract: The primary method for news recommendations revolves around leveraging the user’s browsing history to gauge their interests. Existing models prioritize analyzing news content to infer user interests, ignoring the role of label category information in fine-grained interest modeling. In this paper, we introduce a special approach for personalized news recommendation through a multi-granularity label-aware user interest modeling technique. First, the user’s initialized interest representation is modeled by analyzing their interaction with clicked news articles. Then, considering different attractiveness of category labels of news content to the user, a multi-granularity label-aware user interest representation is designed. Meanwhile, by fusing the representations of the coarse-grained labels, fine-grained labels, and news content, we modeled the semantic representation of candidate news. Through aligning the semantic features of candidate news with the label-aware user profile, we predict the user’s interest score for the given candidate news item. Experimental findings demonstrate the superiority of the proposed technique over current methods across public datasets, as evidenced by improvements in AUC, MRR, and NDCG metrics.