Application of Q-learning for Online Fraud Detection

Abstract

The rapid expansion of online transactions has led to a concurrent increase in the sophistication of fraudulent activities, necessitating the development of advanced and adaptive fraud detection systems. This study proposes the application of Q-learning, a reinforcement learning technique, for enhancing online fraud detection capabilities. The central idea revolves around the creation of a dynamic decision-making model that learns optimal actions in response to evolving patterns of fraudulent behaviour. The core of the proposed approach lies in the Q-learning algorithm, where a Q-table is initialized to store the learned values associated with state-action pairs. The model is trained using real-world transaction data, with the Q-values being iteratively updated according to the observed rewards and transitions between states. Striking a balance between exploration and exploitation during training ensures that the system learns optimal strategies while adapting to emerging fraud patterns. The culmination of this research is a real-time fraud detection system that utilizes the learned Q-values to make informed decisions regarding the approval, flagging, or blocking of online transactions.