Authors: Mollah Mohammad Mahabubuzzaman
Abstract: Credit card fraud is a growing threat to the global financial system, resulting in huge financial losses and eroding trust in digital payments. Traditional fraud detection methods – including legacy rule-based systems and traditional machine learning models – have serious limitations in adapting to new fraud tactics, reasoning under uncertainty and optimizing sequential decision making. This paper introduces a new integrated framework that combines Dynamic Decision Networks (DDN), Bayesian inference and Reinforcement Learning (RL) to address these challenges. I reformulate fraud detection as a sequential decision making problem under uncertainty where each transaction is evaluated in its temporal context using an adaptive policy that maximizes long term expected utility. Mathematical foundations are rigorous while experimental results on real world datasets show 99.99% detection rate with minimal false positives – far better than state of the art. This work sets a new standard for financial security systems that balance protection and customer experience
DOI: http://doi.org/10.5281/zenodo.17083615
