MDIASE-Autoencoder: A Novel Anomaly Detection Method for Increasing The Performance of Credit Card Fraud Detection Models
Aji Gautama Putrada, Nur Ghaniaviyanto Ramadhan
Abstract
Machine learning has presented diverse solutions for identifying credit card fraud by analyzing datasets with anonymous transactions. However, an unexplored avenue lies in employing autoencoders for anomaly detection in credit card fraud identification. Our research objective is to apply and assess autoencoders to enhance credit card fraud detection through machine learning-based classification methods. The study utilizes the credit card fraud dataset available on Kaggle, applying various preprocessing techniques such as standardization and over-sampling using the Synthetic Minority Oversampling Technique (SMOTE). Subsequently, a 5-layer autoencoder is applied to the modified dataset. A novel metric for evaluating reconstruction errors, Mean Decrease in the Impurity of Absolute and Squared Error (MDIASE), is employed for feature selection in the new dataset, combining mean reduction in impurity (MDI), square error, and absolute error concepts. Finally, we assess the influence of the autoencoder on four classification methods: decision tree, KNN, logistic regression, and MLP. The evaluation metrics used for comparison are AUC maximization and geometric mean (g-mean). The test results demonstrate that the autoencoder enhances the performance of four state-of-the-art models in credit card fraud classification. Mainly, MLP exhibits the most significant improvement, with its AUC increasing from 0.558 without the autoencoder to 0.999 after its implementation.