Dynamic Hierarchical Attention Network for news recommendation

Existing news recommendation methods often rely on static user-news interactions that fail to account for the evolving nature of users’ preferences over time. To address the prevalent challenges, which frequently struggle to accurately capture and reflect the dynamic nature of users’ interests, this paper introduces a groundbreaking approach through a dynamic attention mechanism. Our proposed model alleviates this limitation by incorporating continuous time information into a hierarchical attention framework. This framework is designed to capture the nuances of news content at multiple levels of granularity, including news text, element, and clicked news sequence, thereby offering a more comprehensive representation of news articles. To further enhance the model’s predictive accuracy and relevance of recommendations, we integrate a dynamic negative sampling technique. This technique is aimed at refining the model’s understanding of users’ implicit feedback by distinguishing between genuinely uninteresting content and articles that have not yet been encountered by the user. Our model is evaluated through three real-world datasets, encompassing diverse time information and news articles. The results from these experiments demonstrate that by effectively integrating time dynamics and a multi-level representation of news, our model demonstrates an improvement of up to 4.54 points over second-best baselines.