Asian Research Journal of Current Science

Credit card fraud detection is a highly imbalanced classification problem in which overall accuracy can remain high even when a model fails to detect a substantial proportion of fraudulent transactions. This study evaluates artificial neural network (ANN), linear support vector machine (SVM), non-linear SVM, and a weighted ANN–non-linear SVM ensemble on the widely used public European credit-card transaction dataset containing 284,807 transactions, of which 492 are fraudulent. The dataset was partitioned using a 70:30 training–test design. Feature standardisation was based on the training data, and Synthetic Minority Over-sampling Technique (SMOTE) was confined to the training data to limit information leakage. The hybrid model combines the calibrated probability outputs of the ANN and non-linear SVM using weights of 0.6 and 0.4, respectively. Performance was assessed using precision, recall, F1-score, accuracy, ROC-AUC, and computational time, with particular emphasis on minority-class detection. The recorded aggregate results show that the hybrid model achieved precision of 0.89, recall of 0.75, F1-score of 0.81, accuracy of 0.99, and ROC-AUC of 0.94. The ANN achieved 0.88 precision, 0.71 recall, 0.78 F1-score, and ROC-AUC of 0.92; the non-linear SVM achieved 0.84, 0.66, 0.74, and 0.90, respectively; and the linear SVM produced the weakest minority-class performance. These results indicate a modest descriptive advantage for the weighted hybrid on this dataset and split, particularly for recall and F1-score. Because the dataset is historical, anonymised, and highly imbalanced, and because the evaluation is based on aggregate metrics without paired row-level predictions, the findings should be interpreted as benchmark evidence rather than proof of deployment-level superiority. Repeated validation, precision–recall analysis, temporal evaluation, and testing on contemporary institutional data are priorities for future work.