Delay-Aware Sequential Recommendation with Dynamic Graphs and Variate-Temporal Decomposition

초록

Sequential recommendation systems may operate on partially observed histories when interaction logs become available only after non-trivial ingestion latency. Most existing methods assume immediately observed feedback, which can degrade when the model-visible interaction stream is temporally misaligned with the underlying action sequence. We propose DVGD, a delay-aware sequential recommendation framework that learns lag-conditioned message passing over censored histories induced by delayed log availability. Our method represents observed user-item interactions as a dynamic graph, decouples feature-wise and temporal dependencies via variate-temporal decomposition, and learns a time-lag adjacency tensor that routes information across multiple temporal lags. Experiments on benchmark datasets with controlled simulated delays show that the proposed method consistently outperforms strong baselines, with larger gains under longer and more heterogeneous delays. These results support DGVD as an effective approach for delayed-observation robustness in sequential recommendation, while further validation on real delayed logs remains future work.

키워드