Fraud Detection in Financial Transactions Using Advanced Neural Network Techniques
Hye Jin Kim, Rhee Jung Soo · International Journal of Uncertainty Fuzziness and Knowledge-Based Systems · 2025
This paper examines a binary classification issue in fraud detection using several neural network methodologies, including the Synthetic Minority Oversampling Technique (SMOTE) and SMOTE combined with Edited Nearest Neighbours (ENN). This work underscores the pressing need for improved detection approaches due to the rising incidence of fraud in financial institutions. Several configurations were tested on a Feedforward Neural Network (FNN) to tackle class imbalance with class weighting, undersampling, oversampling (SMOTE), and [Formula: see text]. The results demonstrated that [Formula: see text] successfully addressed the imbalance problem, leading to further investigation of models that integrate [Formula: see text] with Long Short-Term Memory (LSTM) networks and Random Forest (RF). Finally, a hybrid model combining the balancing techniques of [Formula: see text] with the neural network architectures of [Formula: see text] was developed to harness the strengths of both approaches. The [Formula: see text] achieved an impressive accuracy of 99.99% with an ROC-AUC score of 0.9678, showcasing its strong performance in the classification task. In contrast, the LSTM model reached an accuracy of 99.91% and a ROC-AUC score of 0.9424. While LSTM performed well, it faced challenges in effectively distinguishing the minority class.