Investigation of Credit Card Fraud detection under the lens of Comparative Machine Learning models

Fraudulent use of credit cards is common and may result in significant financial losses. Using Trojans or Phishing, criminals may get the credit card numbers of innocent victims for their criminal acts. As a result, it is vital to have fraud detection technology that can catch a thief in the act of purchasing using a stolen credit card. One solution is to utilize machine learning algorithms to construct normal/fraud behavior characteristics from all historical transaction data, including legitimate and fraudulent ones, and then use these features to identify whether a transaction is fraudulent or not. In this paper, SVC, Multinomial Naïve Bayes, K-Neighbors Classifier, Logistic Regression, Random Forest, Bernoulli Naïve Bayes, and SVM algorithms were used for credit card fraud detection. The detailed results of all seven algorithms were studied. Then a comparison between these seven algorithms with their precise accuracy, ROC curve, train-test, validation scores, and the learning curve was discussed. Logistic Regression, Random Forest, Bernoulli Naïve Bayes, and SVM showed 100% accuracy in training dataset. While Random Forest algorithm outdid every other algorithm by scoring 100% for both train and test data, the result of the remaining algorithms for the train and test data were also satisfactorily ranging from 75.3% to 99.6%. Logistic Regression, Random Forest, Bernoulli Naïve Bayes, and SVM all 4 scored 100% in both train and test data, and Random Forest scored 100% leaving behind all the algorithms. Logistic Regression, Bernoulli Naïve Bayes and SVM scored 99.6% in test. The best-performing algorithm was selected according to its performance and discussed its AUC score and confusion matrix.