A Context-aware Interest Drift Network for Session-based News Recommendations
Lingkang Meng, Chongyang Shi · 2020
Session-based news recommendation systems aim to provide users with personalized reading suggestions based on their short-term sessions. In the news domain, users' interests change rapidly and are easily affected by the environment and breaking events, and this in turn affects users' next click. However, most existing approaches only capture a single dynamic of users' interests and consider little or no external influences. In this paper, we propose a context-aware interest drift network (CaIDN), a deep context-rich session-based news recommendation framework, that contains environment, breaking news, and article content information. The key component of CaIDN is a bidirectional attention recurrent network that effectively catches the drift of users' reading interests from various aspects. Additionally, to alleviate the cold-start problems, we perform CNN with multi-height filters on textual content and additional news information to generate valid article content features. Extensive experimental results on two real-world news datasets demonstrate that CaIDN outperforms state-of-the-art session-based news recommendation methods.