EFFECTS OF IMBALANCED DATA ON BANKING FRAUD DETECTION MODELS

Abstract

The quality of data is crucial for the effective performance of machine learning (ML) predictive models. To create a high-quality dataset, input data should be thoroughly processed to remove missing values, entry errors, outliers, and data imbalance. In this study, various techniques to address imbalanced datasets, namely, random under-sampling (RUS), random oversampling (ROS), Synthetic Minority Oversampling Technique (SMOTE), and Adaptive Synthetic (ADASYN), were tested using a highly imbalanced credit card transaction dataset. A total of 284807 credit card transactions were downloaded from previously published sources, with 28315 legitimate records and 492 frauds identified. In the first phase, the performances of four commonly used ML predictive models, Logistic Regression (LR), Random Forest (RF), Gaussian Naive Bayes (GNB), and Extra Trees (ET), were evaluated using a balanced subset of 984 records to identify the best-performing model. In the second phase, the selected ML model was used to assess the effectiveness of each of the data-modification techniques using three metrics: Precision, Recall, and F_1score. Results showed that the ET model performed best with a dataset modified using the SMOTE technique, achieving Precision, Recall, and F_1score values of 0.8787, 0.8613, and 0.8699, respectively. In contrast, the dataset modified with RUS yielded poorer results, with a Precision, Recall, and F_1score of 0.0788, 0.9208, and 0.1452, respectively. Finally, an important analysis was conducted, and the ET model configuration utilizing eight key features was recommended for this dataset.