UniMVT: Joint modeling of debiased CTR and coupon uplift¶
Core public-data uplift mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Kuaishou Technology. |
| First public release | 2026-02-13 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | unimvt |
| Local reproduction code | src/auto_research/reproductions/unimvt/ |
Paper method and evidence¶
Separate base click propensity from coupon-intensity sensitivity with deconfounded representations. Dual prediction towers and a counterfactual X-Network constrain factual and counterfactual outcomes. Unit uplift puts coupon tiers on one scale.
flowchart LR
A[Separate propensity and sensitivity] --> B[Predict factual and counterfactual outcomes]
B[Predict factual and counterfactual outcomes] --> C[Normalize unit uplift]Original paper figure¶
Figure shows debiasing and counterfactual uplift. From the original paper; copyright remains with the authors.
Core equation¶
\[
\hat\tau(x,t)=\hat p_t(x,t)-\hat p_0(x),\qquad \mathcal L=\mathcal L_{base}+\mathcal L_{treat}+\lambda\mathcal L_{cf}.
\]
Reported results¶
Kuaishou on 10% traffic reports coupon revenue +18.14%, ROI +8.80, and coupon PCOC error reduced by 86.51%.
Local reproduction¶
MovieLens-100K full catalog: NDCG@10 0.05401→0.03955 (-26.76%). Three-seed metrics, mean, SD, and 95% CI. No private logs, production models, bucketing or serving. Public proxies cannot establish online lift.