Reconstructing Fraudulent Strategies in Online Payments via Interpretable Inverse Reinforcement Learning (arXiv (Cornell University))

Summary

• The dataset from Ant International covers approximately 1.29 million users and 26.28 million distinct payment events across 3.46 million labeled detection examples.p.5
• Target payments span March 2025 to January 2026.p.5
• On GraphSAGE, the Recall@2 gain reaches 4.561 points, exceeding the strongest generation baseline by 1.186 points.p.6
• A reward-guided analysis finds a card-cycling pattern in 3.95% (35/885) of fraud-target trajectories versus 0.25% (4/1,604) of legitimate-target ones.p.7

Key insights

1. PaySeqGen reconstructs fraud behavior policies from observed fraud trajectories using interpretable inverse reinforcement learning.p.1
2. On the RT detector the full PaySeqGen model gains +0.393 ΔAUC and +1.621 ΔRecall@2.p.6
3. Removing the interaction component turns the RT ΔAUC to -0.257.p.6
4. On the public MS-FFSD dataset, PaySeqGen gives +2.194 ΔAUC on GraphSAGE.p.12
5. The pattern analysis requires a minimum amount reduction of 0.2.p.12
6. Existing generation methods do not model the action-feedback loop through which payment outcomes guide later decisions.p.8