Credit Card Fraud Detection Based on Support Vector Machine

Abstract

Due to the increasing popularity cashless transactions, credit card fraud has become one of the most common frauds and caused huge harm to the financial institutions and individuals in real life. In this academic paper, the algorithm Support Vector Machine (SVM) is used to build models to deal with the credit card fraud detection problem with the performance metrics AUC and F1-score. The experiment dataset is named Credit Card Transactions Fraud Detection Dataset from the Kaggle website. After the step of preprocessing, the dataset is split into the training, testing and validation dataset with 11 numerical features and a label feature called “is_fraud”. The inner parameter “class_weight” of the SVM algorithm in Python is set as “balanced” to deal with the imbalanced datasets. The main method to find the optimized models is using the GridSearchCV function in Python library sklearn. After tuning the hyperparameters and handling the overfitting phenomenon, the optimized models for the two metrics are found. The parameter values of the best model for AUC are C=10, class_weight= “balanced”, g =0.01, kernel = “rbf”. The training AUC is 0.87 and testing AUC is 0.90. The parameter values of the final optimized model for F1-score are C=0.8, class_weight= “balanced”, g =0.06, kernel = “rbf”. The final training F-score is 0.305 and testing F-score is 0.260.