Research map / Learning theory and algorithms / Prediction and decision losses

Adaptive Merchant-Centric Risk Control via Unbiased Decision-Making and Dynamic Optimization in E-Commerce

AAAI 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Adaptive testing and experiments · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes an automated merchant risk-control framework that uses linUCB-style contextual bandit estimates of PRO/PCO from randomly explored data, with delay-aware learning rates and primal-dual optimization over exploration, risk, and experience modes, deployed at Ant Group.

Problem

Allocating merchant-level risk control strategies in e-commerce balances risk exposure (PRO) against merchant complaints (PCO), but feedback is delayed by days to a month, exposure bias limits unbiased evaluation of unchosen actions, and manual allocation is costly.

Approach

Treat PRO and PCO prediction per merchant group as a contextual bandit with linUCB-style updates whose learning rate decays with feedback delay, use random exploration traffic for unbiased data, and allocate actions by a constrained optimization (online primal-dual) that minimizes one KRI while bounding the other.

Key results

• PRO and PCO prediction reaches AUC-ROC 0.87 and 0.84 respectively.
• Risk mode reduces average PRO by 23% (0.44 to 0.34) with PCO unchanged; experience mode reduces PCO by 55% (0.20 to 0.09) with PRO unchanged versus manual strategy.
• In deployment, risk exposure rate decreased by 33% and merchant complaints dropped by 69%; decision time was one-fifth that of the Pareto front method.

Measured effects reported in the paper

Surprising findings

• Early deployment showed the risk mode failed to control risk exposure and favored lenient strategies, traced to over-representation of weak control actions in exploration data; increasing exploration traffic fixed it.

Limitations

• Relies on extensive, diverse, and accurate data; may not apply in data-constrained or biased settings.
• Only PRO and PCO considered; no multi-dimensional risk factors or explainability mechanisms yet.
• Merchant grouping based on past business experience may introduce bias.

Experimental setup

Datasets: Ant Group merchant risk control feedback logs

Benchmarks: Ant Group merchant risk control online deployment (A/B split)

Metrics: AUC-ROC average PRO average PCO relative PRO reduction relative PCO reduction decision time

Baselines: Manual Strategy Single (PRO Minimization) Single (PCO Minimization) Pareto Front

Builds on: linUCB online primal-dual algorithm Pareto front optimization

Closest papers by meaning

• Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.81)
• Adaptive Market Making with Inventory Constraints via Online Learning (AAAI 2025 · similarity 0.80)
• Performative Risk Control: Calibrating Models for Reliable Deployment under Performativity (NeurIPS 2025 · similarity 0.79)
• Wait-Less Offline Tuning and Re-solving for Online Decision Making (ICML 2025 · similarity 0.79)
• Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data (NeurIPS 2025 · similarity 0.79)
• Decoupling Learning and Decision-Making: Breaking the $\mathcal{O}(\sqrt{T})$ Barrier in Online Resource Allocation with First-Order Methods (ICML 2024 · similarity 0.79)