Fraud Detection Using ML Algorithms

The increase in digital transactions has created an urgent need for reliable fraud detection. This study evaluates the performance of various machine learning models—Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Classifier (SVC), Decision Trees, Random Forests, and Neural Networks—alongside a Graph Neural Network (GNN) to detect fraudulent transactions. Recognizing the inherent class imbalance in fraud detection datasets, we applied Random Undersampling and the Synthetic Minority Over-sampling Technique (SMOTE) to create balanced training sets and optimize model performance. Our results revealed that Neural Networks achieved the highest accuracy, reaching 98.22% with under-sampling and 100.00% with SMOTE. Logistic Regression also demonstrated strong performance with a 98.67% accuracy on SMOTE-balanced data, followed by KNN (97.46%) and SVC (98.51%). Although the GNN model leverages advanced graph-based representation learning, it achieved an accuracy of 70.38%, suggesting limitations in Graph Neural Networks for this specific task. We evaluated model performance using accuracy, precision, recall, and F1-score metrics, which provided comprehensive insights through confusion matrices. The findings emphasize the critical role of data handling techniques, as balancing the dataset significantly enhanced model sensitivity and accuracy. This study underscores the efficacy of machine learning in fraud detection and highlights areas for improvement, including exploring optimized architectures and real-time data analysis for GNNs. In conclusion, selecting appropriate models and data balancing techniques is essential in developing effective fraud detection solutions, with Neural Networks and SMOTE proving particularly successful in identifying fraudulent transaction.