Optimisation and Resampling methods for Handling Imbalanced Datasets in Credit Card Fraud Detection using Artificial Neural Networks

Abstract

In recent years, there has been a significant increase in fraud cases, causing financial losses for many companies. Detecting credit card fraud is a challenge due to the infinite and growing number of ways for fraudsters to commit fraud. Machine learning techniques have been proposed to detect fraudulent transactions, but the extremely low rate of fraud cases versus non-fraudulent transactions produces imbalanced data, making it difficult to train machine learning models. This study presents a comparative experimental approach to address the imbalance classification problem by employing several optimization and resampling methods to deal with imbalanced datasets. We examined methods such as binary cross-entropy loss minimization, minimization using weights to adjust for class imbalance, under-sampling, and over-sampling using the SMOTE technique, in combination with artificial neural networks. We applied the experiment to 284,807 anonymous transactions to compare the performance metrics of the listed approaches. It is crucial to consider the impact of class imbalance on model performance and choose appropriate techniques to mitigate its effects, as failure to do so can lead to unreliable models.