Modeling user interactions by feature-augmented graph neural networks for recommendation

Xinzhou Dong, Beihong Jin, Wei Zhuo, Beibei Li, Taofeng Xue, Jiageng Song

Abstract: Analyzing user behaviors is a conventional approach to accomplish personalized recommendation. However, due to the intrinsic complexity of user behaviors, modeling the user behaviors in an accurate and comprehensive way to achieve effective recommendation is still a challenge. In the paper, we focus on user interaction behaviors and propose a model named Sirius, which designs graph neural networks to model the collaborative relations implied in the interactions and capture the dynamics of sequence features (including time and attribute features). Sirius constructs two kinds of graphs from interaction sequences and then builds two kinds of graph neural networks to jointly mine the implied relations in interactions. In particular, the sequence time feature and sequence attribute feature are fused into the generation of item and sequence embeddings. At last, Sirius gives the item recommendation by next item prediction. We conduct extensive experiments on multiple real-world datasets. The experimental results show that Sirius outperforms several state-of-the-art models in terms of recall and mean reciprocal rank (MRR). Moreover, Sirius has been deployed in MX Player, one of India’s largest streaming platforms, and achieves the improvement on online unique click-through rates (CTRs), which demonstrates the effectiveness and feasibility of Sirius in a real-world production environment.