Credit Card Fraud Detection using AdaBoost Algorithm in Comparison with Various Machine Learning Algorithms to Measure Accuracy, Sensitivity, Specificity, Precision and F-score

Abstract

Credit card fraud detection is a critical problem for any credit card issuing banks. The AdaBoost classifier is used in this study to identify fraudulent transactions. By comparing the proposed algorithm with Naive Bayes, logistic regression, ANN and decision tree algorithms the efficiency of the algorithm is evaluated. A total of 2,84,807 transactions are divided into two subsets: a training dataset [n=2,27,845 (80%)] and a test dataset [n=56,962 (20%)] (0.8 g power). Out of 2,84,S07 transactions in the dataset, 492 transactions are fraud transactions. To detect the credit card frauds Adaboost algorithm is used and various machine learning algorithms are compared with it for performance evaluation. To determine the performance of algorithms, metrics such as accuracy, sensitivity, specificity, precision, and f-score are estimated. The detection accuracies of AdaBoost, Naive Bayes, logistic regression, ANN and decision tree algorithms are 99.43%, 90.93%, 95.35%, 94.81% and 94.81% respectively. The AdaBoost algorithm obtained an f-score of 99.48% with significance value p<0.05. From the qualitative analysis, it is observed that the proposed AdaBoost algorithm performed significantly better than the Naive Bayes, logistic regression, ANN and decision tree algorithms in detecting credit card frauds.