Enhanced Fraud Detection Using Multi-Autoencoder Feature Extraction and Optimized SVM

This study employs multiple autoencoder to learn hierarchical feature representations from transaction data. Furthermore, a Genetic Algorithm is utilized to fine-tune the hyperparameters of the SVM classifier in order to enhance its ability to detect fraudulent activities. The proposed approach was evaluated using a large-scale e-commerce transaction dataset containing 1,472,952 records, where 5.01% of the transactions were labeled as fraudulent. The model’s effectiveness was assessed using a combination of classification metrics, including accuracy, precision, and recall, along with the F1 metric and curve-based evaluations such as ROC-AUC and PR-AUC. Experimental results show that the proposed model achieved 0.9683 accuracy, 0.6679 precision, 0.7664 recall, 0.7137 F1-score, and 0.9793 AUC-ROC, outperforming baseline models and several previous studies using the same dataset. Ablation study results further demonstrate that the integration of multi-autoencoder feature extraction and evolutionary hyperparameter optimization significantly improves fraud detection performance. These findings suggest that the proposed framework is capable of effectively identifying fraudulent transactions even when dealing with highly skewed e-commerce data. In addition, the approach shows meaningful promise for being applied in practical fraud detection contexts.