Discovering and Alleviating Data Leakage in Staytime Prediction for Live Streaming Recommendation

Staytime prediction plays a central role in live-streaming recommendation. To satisfy the high timeliness requirements of live scenarios, previous methods mainly employ multi-window streaming, splitting each watching record into multiple samples with different temporal horizons. However, our empirical results reveal that such multi-window streaming inherently introduces data leakage issue (i.e., short-window training updates embeddings that long-window serving-time prediction cannot access). This causes a training–serving mismatch and significantly degrades long-horizon staytime prediction. To better analyze this problem, we construct the first benchmark for data leakage detection in live-streaming recommendation. We reveal that this issue mainly stems from shared bottom embeddings rather than task-specific heads, and identify a cutoff time beyond which the timeliness gains are outweighed by leakage-induced losses. In light of this analysis, we propose a simple but effective approach, called DAST, which decouples embeddings across different temporal windows to block leakage paths while preserving the advantages of multi-window supervision. Experiments demonstrate consistent improvements on staytime prediction, and our method can be seamlessly integrated into existing frameworks, making it a general and model-agnostic solution for leakage-aware live-streaming recommendation.