Fraud Transaction Detection Using Machine Learning on Financial Database

Financial fraud poses a significant threat to the digital economy, with credit card fraud being a prevalent challenge. This study evaluates the performance of a Voting Classifier Ensemble, Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), and AutoEncoder with Gradient Boosting models in detecting fraudulent transactions using financial datasets. The study uses practical data from 284,807 transactions, but only 492 are fraudulent; the imbalanced class issue is solved using the Synthetic Minority Oversampling Technique (SMOTE). Our findings show that Voting Classifier Ensemble is the best with 99.93% accuracy, 100% precision, 86% recall and 90% F1 Score. XGBoost yielded an accuracy of 99.96%, precision of 95.11%, recall of 79.61%, and F1 score of 86.61%, while for Logistic Regression, the corresponding percentage was 99.92%, 88.1%, 60.5%, and 71.7%. AutoEncoder with Gradient Boosting had an accuracy of 99.98%, precision 99.81%, recall 64%, F1 Score 69%. The AUC statistic of 0.98 for XGBoost and Voting Classifier Ensemble against 0.97 for LR, and 0.46 for AutoEncoder wjth Gradient Boosting classified the models as having better discriminant power. The results show that XGBoost is more suitable for real-time fraud detection. However, computational limitations and explainability issues should be considered. For future work, it is suggested that semi-supervised and supervised learning approaches be investigated and work with larger datasets to improve fraud detection in financial systems.