UNEG: A Description-Enhanced Graph-based News Recommendation Method

Yipu Chen, Jingkun Wang, Yongtao Jiang, Liqiang Wen, Wen Zhao

Abstract: Modeling news and users accurately became more and more crucial in personalized news recommendation. Exploring rich relational information with graph-based representation learning methods is challenging. Existing graph-based methods learn the representations of news items and user interests with external knowledge graphs or user-news bipartite graphs. However, they rarely link two news through the entities they refer to in common and fail to consider their correlations. In this paper, we propose a novel User-News-Entity Graph, named UNEG, which associates news items with their mentioned entities and users with the news articles they browsed. In our practice, we use category-aware graph attention networks to model representations of high-order relationships between nodes in UNEG. Furthermore, we introduce entity descriptions as auxiliary information to enrich the initial representations of entity nodes, which provide additional text information on entity labels. We conduct experiments on a real-world dataset, and the results validate the effectiveness of our approach.