Live streaming recommendation based on multiple types of repeated behaviors
Mengxiao Zhu, Qi Shu, Shuanghong Shen, Li Feng, Jiancan Wu, Zhenya Huang · Expert Systems with Applications · 2025
In recent years, live streaming develops rapidly, attracting an increasing number of users. Providing personalized live streaming recommendations is crucial for platform improvement in enhancing user experience and increasing profitability. In live streaming scenarios, users often enter the same live streaming room multiple times, send chat messages and give virtual gifts repeatedly. However, existing recommendation models fail to effectively model the complex multiple types of repeated behaviors of users in live streaming scenarios, thus failing to achieve optimal recommendation results. To address this issue, we propose a novel live recommendation model called MRB4LS based on multiple types of repeated behaviors data. Specifically, we first construct three bipartite graphs to better capture the effects of multiple types of users’ repeated behaviors, including enter, chat, and gift. Second, we introduce a graph attention network named RepGAT, which explicitly learns from users’ repeated behaviors. RepGAT incorporates the number of repeated interactions between nodes when computing normalized attention coefficients, enabling a deeper exploration of users’ preferences and the heterogeneous strength of the interaction relationship between users and live streaming rooms. Then, we design two embedding fusion strategies, namely concatenation-based and attention-based methods, to integrate node representations generated by different repeated behaviors. Finally, we adopt a multi-task learning approach to enhance the prediction effectiveness of gift behavior by leveraging predictions from enter behavior and chat behavior. To validate our approach, we construct two live streaming datasets from a large-scale game live streaming platform. Extensive experiments on two real-world datasets with different scales show that our method can significantly outperform various baseline approaches.