Credit Card Fraud Detection with Autoencoder and Probabilistic Random Forest

Abstract

This paper proposes a method, called autoencoder with probabilistic random forest (AE-PRF), for detecting credit card frauds. The proposed AE-PRF method first utilizes the autoencoder to extract features of low-dimensionality from credit card transaction data features of high-dimensionality. It then relies on the random forest, an ensemble learning mechanism using the bootstrap aggregating (bagging) concept, with probabilistic classification to classify data as fraudulent or normal. The credit card fraud detection (CCFD) dataset is applied to AE-PRF for performance evaluation and comparison. The CCFD dataset contains large numbers of credit card transactions of European cardholders; it is highly imbalanced since its normal transactions far outnumber fraudulent transactions. Data resampling schemes like the synthetic minority oversampling technique (SMOTE), adaptive synthetic (ADASYN), and Tomek link (T-Link) are applied to the CCFD dataset to balance the numbers of normal and fraudulent transactions for improving AE-PRF performance. Experimental results show that the performance of AE-PRF does not vary much whether resampling schemes are applied to the dataset or not. This indicates that AE-PRF is naturally suitable for dealing with imbalanced datasets. When compared with related methods, AE-PRF has relatively excellent performance in terms of accuracy, the true positive rate, the true negative rate, the Matthews correlation coefficient, and the area under the receiver operating characteristic curve.

Keywords

• autoencoder
• credit card
• deep learning
• fraud detection
• data imbalance
• random forest