Credit Card Fraud Detection Using One-Class Support Vector Machines

Abstract

The dissertation explores feature engineering and feature selection using one-class support vector machines (OC-SVMs) and demonstrates how this approach can improve fraud-detection models. Feature selection for fraud detection is challenging due to severe class imbalance. The main objective is to use an OC-SVM to explore the available features and develop a feature selection model based on OC-SVM. A dataset was downloaded from the Kaggle website, containing 31 features (columns) and 284807 rows. Missing the account balance and the datetime feature, which are necessary to improve the amount-per-transaction feature, greatly affect results. An OC-SVM with an RBF kernel was used to explore the nature of the features and to improve the model’s usability and create an understanding of the features. An aggregation strategy based on transaction count was applied to the amount spent, with the hope that the model would capture subtle differences in performance. The model’s optimisation was limited, so optimising the dataset was key. Using the confusion matrix to build performance measures, the OC-SVM performed well in selecting high-return key variables. Shorter aggregation periods (3 to 10) improved the base models more than longer ones (> 20).