Enhancing Risk Management In E-Commerce Using Deep Reinforcement Learning And Markov Chains

Abstract

The swift rise of e-commerce has raised the chances of fraudulent online payments, irregular transactions, and cyberattacks, giving rise to the significant problems of intelligent financial risk management. Static machine learning models used in traditional fraud detection methods are prone to failing to leverage changes in fraud patterns and fraud behaviors across sequential transactions. In this paper, a hybrid Deep Reinforcement Learning (DRL) and Markov Chain-based framework to manage risks in e-commerce in an adaptive manner is proposed by using the IEEE-CIS Fraud Detection Dataset. To enhance the fraud detection and perform dynamic risk mitigation, the proposed methodology combines the probabilistic transaction risk state modeling with Deep Q-Network (DQN)-based autonomous decision optimization. The DRL agent learned fraud response actions by rewards in a dynamic manner, and Markov Chain modeling was employed to analyze the state transition for transactions between safe, low-risk, suspicious, and fraudulent states. The experimental assessment resulted in the better performance of the proposed framework than with traditional machine learning. The proposed model has demonstrated a 97.6% accuracy, 96.8% precision, 95.9% recall, and 96.3% F1-score, which are significantly better than Logistic Regression, Random Forest, and XGBoost models. The false-positive rate was lowered to 2.1%, and the fraud detection rate was raised to 96.7%. The DRL agent converged stably at 320 training episodes with a cumulative reward score of 0.91. The results show that the combination of deep reinforcement learning and Markov chain probabilistic modeling methods noticeably improves the adaptive fraud detection, intelligent decision-making, and efficiency of real-time e-commerce risk management.