Optimizing Transaction Fraud Detection: A Comparative Study of Nature-Inspired Algorithms for Feature Selection

As credit card transactions continue to rise, reaching over 900 billion annually, so does the incidence of transaction fraud, necessitating advanced data mining techniques for effective detection. Globally, credit card fraud was estimated to be around ${\$}$33 billion in 2022, based on the report published by Nilson, and it is estimated to reach ${\$}$400 billion in the next decade [1]. This study investigates the efficacy of five nature-inspired algorithms-the Firefly Algorithm, the Cuckoo Search, the Bat Algorithm, the Flower Pollination Algorithm, and the Grey Wolf Optimizer-in the feature selection process for binary classification models. Our comparative analysis reveals that although all tested algorithms contribute to reducing model complexity, the Grey Wolf Optimizer (GWO) stands out by achieving a $1.4 \%$ improvement in ROC AUC score over traditional feature selection methods while utilizing only 53 features - less than half compared to other approaches. These findings demonstrate GWO’s superior ability to identify optimal feature subsets in fraud detection scenarios, offering both improved performance and significantly reduced computational requirements.