Comparison with Parametric Optimization in Credit Card Fraud Detection

We apply five classification methods, neural nets (NN), Bayesian nets (BN), naive Bayes (NB), artificial immune systems (AIS) and decision trees (DT), to credit card fraud detection. For a fair comparison, we fine adjust the parameters for each method either through exhaustive search, or through genetic algorithm (GA). Furthermore, we compare these classification methods in two training modes: a cost sensitive training mode where different costs for false positives and false negatives are considered in the training phase; and a plain training mode. The exploration of possible cost-sensitive metaheuristics to be applied is not in the scope of this work and all executions are run using Weka, a publicly available software. Although NN is claimed to be widely used in the market today, the evaluated implementation of NN in plain training leads to quite poor results. Our experiments are consistent with the early result of Maes et al. (2002) which concludes that BN is better than NN. Cost sensitive training substantially improves the performance of all classification methods apart from NB and, independently of the training mode, DT and AIS with, optimized parameters, are the best methods in our experiments.