Layered Behavior Modeling for Recommendations

Sequential recommendation, a critical aspect of recommendation systems, focuses on predicting users’ next preferred items based on their behavioral history. Existing advanced models, including transformer-based and dynamic graph-based approaches, often prioritize macroscopic event sequences or node relationships, overlooking fine-grained information within user and item attributes. This gap poses challenges in constructing stable models, particularly in sparse data scenarios. To address this, we propose an approach that seamlessly integrates side information while maintaining computational efficiency. Our novel method, named Micro Event Graph (MiEG), employs layered computations to project fine-grained attribute information onto a temporal event graph. Subsequently, final node representations are generated through graph neural network processing. Extensive experiments on real-world datasets showcase MiEG’s superiority over the state-of-the-art baselines, underscoring its exceptional information mining capabilities. Degeneration experiments further reveal MiEG’s competitiveness even in nonsequential recommendation scenarios.