<![CDATA[The rise of fintech and digital payment systems has increased efficiency but also escalated the risk of online transaction fraud, particularly under imbalanced data conditions where fraudulent cases are rare. This study addresses the limitations of traditional rule-based and machine learning models in such scenarios by proposing the use of Extreme Learning Machine (ELM) with hyperparameter tuning as a novel and efficient solution for fraud detection. Unlike most prior studies relying on default settings or data resampling, this research focuses on enhancing ELM performance purely through parameter optimization using the Optuna framework. A dataset of 20,000 real-world online transactions was used to evaluate model performance before and after tuning. In its default configuration, ELM yielded high overall accuracy (96.80%) but failed to detect fraudulent cases (0% recall and F1-score). After tuning key parameters such as the number of hidden neurons and activation function, the model achieved a significantly better balance between accuracy and fraud detection performance, with 99.53% accuracy, 98.20% precision, 86.51% recall, and a 91.98% F1-score. These results demonstrate that hyperparameter tuning alone, without resampling, can substantially improve ELM’s sensitivity to minority class detection. The findings suggest that optimized ELM offers a promising alternative for real-time fraud detection in imbalanced financial datasets, contributing to more adaptive and reliable security systems in the digital finance landscape.]]>
