How RAS‑AUCC Eliminates Offline‑Online Gaps in Multi‑Treatment Uplift Modeling

Background

Intelligent marketing aims to meet subsidy‑rate constraints while maximizing GMV. In ride‑hailing platforms, different scenarios have vastly different supply‑demand, making blanket subsidies infeasible; the problem is modeled as “satisfy subsidy‑rate constraint & maximize GMV”.

Uplift Model

Uplift modeling estimates the incremental effect (ΔGMV/ΔCost) of a treatment. Unlike conventional response models, uplift models must predict label differences under multiple treatments, requiring counterfactual reasoning.

Characteristics of the Problem

Discount decisions at the scenario granularity.

Objective: maximize GMV or completed orders under a subsidy‑rate constraint.

Multi‑treatment discount decision problem.

Evaluation Metrics

Radcliffe proposes Boldness (ranking ability) and Accuracy (bias). Common metrics include AUUC (Area Under Uplift Curve) and Qini coefficient, which measure the area between model and random lines after normalizing axes.

AUUC focuses on ranking; Qini handles imbalanced treatment/control sizes and is generally more stable.

Both metrics ignore absolute ROI, leading to gaps when budget is limited.

AUCC

AUCC (Area Under Cost Curve) incorporates both incremental revenue and incremental cost, using cost as the x‑axis and revenue as the y‑axis, aligning evaluation with ROI.

MT‑AUCC

MT‑AUCC extends AUCC to multi‑treatment scenarios by aggregating per‑treatment AUCCs or weighting them according to online discount distribution.

RAS‑AUCC

To close the offline‑online gap without assuming ROI decreases with deeper discounts, we propose RAS‑AUCC (Area Under Cost Curve for Resource Allocation Simulation). It sorts samples by marginal ROI across all treatments, uses RCT data, and draws curves that match the optimal policy’s decision space.

X‑axis: Cost or subsidy‑rate.

Y‑axis: scale metrics such as call volume, completed orders, GMV, ROI.

Implementation steps include computing per‑sample cost and scale estimates for each discount, retaining only convex‑hull points, sorting by a derived ranking metric, selecting cut‑off points, and plotting the resulting curve.

Summary and Outlook

Our RAS‑AUCC metric reproduces the optimal discount‑allocation strategy, works with multi‑treatment settings, does not rely on ROI‑decreasing assumptions, and has been validated on online experiments, showing higher offline scores correlate with real‑world gains. Future work includes adding bias‑free checks, ATE error metrics, and exploring observation‑data‑only evaluation.