Continuous Temporal Line Graphs for News Recommendation: A Novel Approach to Modeling User Interest Drift

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In personalized news recommendations, users' interests frequently shift dynamically, a challenge that conventional node-based graph models find difficult to address. Many existing approaches depend heavily on similarity between static embeddings of users and items, but they do not effectively capture the causal relationships and influence between successive interactions. In this paper, we introduce a new framework, CTLGNN, which effectively captures long-term user preferences using a dynamic self-attention network while also tracking short-term interest drift with a continuous temporal line graph. Our approach converts user-news interactions into edge-nodes, allowing it to clearly capture the changing influence of adjacent interactions on user behavior. Evaluations conducted with real-world datasets demonstrate that CTLGNN surpasses other baseline models in terms of F1, AUC, and Recall metrics.