Fast yet Accurate Learning: A Novel Joint Data Stream and Model Framework for Staytime Prediction in Live Streaming Recommendation

Abstract

Accurate staytime prediction is crucial for live streaming recommendation, directly influencing user experience and platform revenue. However, due to the extremely short lifespan of live streams, this task faces a fundamental conflict: models must make real-time predictions shortly after a live stream begins (timeliness), while high-value watching behaviors require longer observation to be modeled accurately (accuracy). Existing solutions, such as delayed feedback modeling and fixed multi-window approaches, struggle to reconcile this trade-off. To this end, we propose a novel joint data stream and model framework. First, we design a multi-slice data sending mechanism to replace fixed windows. It uses denser slices to enhance timeliness, employs a Balanced Elastic-Window (BEW) strategy to prevent data loss from reporting delays, and introduces a Cascaded Subspace Multi-Slice (CSMS) strategy to maintain accuracy and unbiasedness without full-data repetition. Second, based on this data stream, we propose a Multi-Slice Incremental Learning (MSIL) paradigm. Its architecture combines shared and slice-specific parameters to alleviate sparsity of high-staytime samples and update conflicts arising from multi-slice data, improving long-tail behavior modeling while preserving timeliness. Furthermore, we identify and mitigate the End Live Bias. This bias occurs when a live stream ends unexpectedly, truncating the staytime and misleading the model. We correct it by applying gradient masking to such samples. Extensive experiments on an industrial dataset show our framework effectively balances timeliness and accuracy, outperforming state-of-the-art baselines. The solution has been fully deployed in TikTok's live streaming recommendation system, and online A/B tests show significant business gains.