CAFE-RL: Counterfactual Augmented Reinforcement Learning for Mechanism-Aware Onboarding Fraud Detection in E-Commerce

Buyer Onboarding Fraud Detection (BOFD), defined as fraud detection at the early stages of user registration and first-transaction screening, is a strategic decision-making problem that defines a platform intervention mechanism and shapes user participation, adversarial behavior, and long-term outcomes such as Gross Merchandise Value (GMV). Existing systems typically optimize these checkpoints independently, leading to suboptimal trade-offs between fraud prevention and legitimate user retention. In this study, we model BOFD as a two-stage Markov Decision Process that captures sequential platform decisions and their long-term effects. Build on this, we propose CAFE-RL, a counterfactual augmented offline reinforcement learning framework that learns a unified decision policy from historical logs. To address the strategic and statistical bias induced by deployed mechanisms, CAFE-RL introduces a counterfactual augmentation scheme that constructs complementary transitions, ensuring full state–action coverage. In addition, we propose a hybrid reinforcement–contrastive objective that combines conservative Q-learning with supervised contrastive losses over factual–counterfactual pairs, providing stronger supervision and stabilizing policy convergence. Experiments on large-scale eBay data demonstrate that jointly optimizing early-stage interventions significantly outperforms deployed checkpoint-wise mechanisms, reducing fraud while preserving legitimate transactions.