Reproduce “LLM-Enhanced Self-Evolving Reinforcement Learning for Multi-Step E-Commerce Payment Fraud Risk Detection” on OpenResearch

About this paper

Bo Qu, Zhurong Wang, Daisuke Yagi, Zhen Xu, Yang Zhao, Yinan Shan, Frank Zahradnik

eBay researchers developed an LLM-enhanced self-evolving reinforcement learning framework for multi-step e-commerce payment fraud detection. This framework consistently outperformed traditional supervised learning models and human-designed RL reward functions, achieving up to 73.74% dollar-wise precision at 80% recall, and demonstrated long-term robustness over a six-month evaluation period.

Key insights

• Modeling multi-step payment fraud as a constrained Markov Decision Process (MDP) enables integrated sequential decision-making that optimizes cumulative business metrics over isolated stage-wise classifications.
• Large Language Models (LLMs) can effectively automate the complex task of designing and refining RL reward functions, often outperforming human-crafted versions.
• A self-improving feedback loop, where LLMs generate and refine reward functions based on RL agent performance, facilitates continuous adaptation and robustness in dynamic environments.

Summary by alphaXiv.