DADF: A Distribution-Aware Debiasing Framework for Watch-Time Regression in Recommender Systems

Abstract

Watch-time prediction is a central regression task in short-video recommender systems, where labels are highly long-tailed and residual errors vary systematically across observed watch-time regions. In practice, a model may appear globally calibrated while still overestimating short views and underestimating long views, because opposite errors cancel out in aggregate. Existing methods mainly improve the first-stage watch-time predictor, but often leave such residual distributional bias insufficiently corrected. We propose DADF, a distribution-aware debiasing framework for watch-time regression. Instead of replacing a deployed predictor, DADF performs second-stage multiplicative residual correction on top of it. DADF combines three complementary designs: a dynamic distribution-aware transformation for stabilizing long-tailed correction targets, a debias-factor-aware module for modeling heterogeneous residual patterns using inference-time observable factors, especially video duration, and a multi-label-aware module that exploits auxiliary prediction signals from engagement heads. We evaluate DADF on public short-video benchmarks and a large-scale industrial ranking system. DADF consistently improves both pointwise accuracy and ranking quality across datasets and backbones. In the industrial setting, it achieves an aggregated 2.07 percentage-point ranking-quality gain over the production baseline, consistently reduces MAE, and yields statistically significant online lifts of 0.649% in average time spent per device and 0.656% in total app time. These results demonstrate that DADF effectively mitigates local calibration bias and provides a practical plug-in solution for debiasing long-tailed continuous targets. The source code is available at https://github.com/liuzhao09/DADF.