Gaussian Naive Bayes for Real-World Credit Card Fraud

Abstract

Researchers explored how well a simple classification method spots fraudulent credit card use. They utilized a real-world collection of nearly 300,000 transactions - spanning just two days - where fewer than 500 were actually cases of fraud. Because so few transactions were fraudulent, they tried different techniques to even things out before training a particular version of this classifier, then tested its accuracy. Because data leans heavily toward one outcome, we focused less on overall correctness yet more on how well the system identifies all actual positive cases while minimizing expenses. It turns out that even a straightforward method - Naive Bayes - works surprisingly well when data is prepared carefully and imbalances are addressed; its ability to pinpoint relevant items alongside reasonable precision levels rivals those of complex systems. We also cover real-world deployment details, where this falls short, along with ideas for improvement.