DADF: A Distribution-Aware Debiasing Framework for Watch-Time Regression in Recommender Systems¶
Core public-data debiasing mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Kuaishou Technology. |
| First public release | 2026-05-18 (arXiv v1) |
| Author code | liuzhao09/DADF. |
| Local adapter or method key | dadf |
| Local reproduction code | src/auto_research/reproductions/dadf/ |
Paper method and evidence¶
Freeze a mature first-stage watch-time predictor. Learn a distribution-conditioned multiplicative residual without changing the scalar serving interface.
flowchart LR
A[Freeze first-stage predictor] --> B[Learn conditional residual]
B[Learn conditional residual] --> C[Keep scalar serving interface]Original paper figure¶
The original figure shows the conditional residual stage. From the original paper; copyright remains with the authors.
Core equation¶
\[
\hat y=\hat y_0\,b_\phi(x,\hat y_0),\quad b^*=y/\max(\hat y_0,\epsilon)
\]
Reported results¶
Seven-day production A/B then full rollout: average time per device +0.65%. Not comparable to MovieLens.
Local reproduction¶
Matched MovieLens-100K 220 users/360 items, full catalog, seed 42: shared transition+content baseline and DADF both NDCG@10 0.0540 (+0.00%) and Hit@10 0.1091. Blend chosen on validation only. Fixed metrics. No private logs, production models, live traffic, or serving; public-data measures do not establish online lift.