Performance evaluation of class balancing techniques for credit card fraud detection

The number of online transactions has unraveled in large proportions with each passing day. Credit card transactions constitute a huge portion of these transactions. The financial losses have also increased analogously along with the credit card fraud transactions. Therefore, fraud detection systems have acquired great importance for banks and financial institutions. As the occurrence of fraud is unlikely in comparison to normally occurring transactions, we are posed with the class imbalance problem and to handle this imbalance problem we use resampling techniques in this paper. We applied oversampling (SMOTE, SMOTE ENN, SAFE SMOTE, ROS, SMOTE TL). On the resampled data, we applied cost sensitive (CSVM, C4.5) and ensemble classifier (Adaboost, Bagging) to evaluate the performances using sensitivity, specificity, G-mean, Area under ROC. We observed that the SMOTE ENN method detects the fraud in a better way than other classifiers in the set of oversampling techniques considered, and TL works better on the set of undersampling techniques taken.