AI-Powered E-commerce Fraud Detection Using Modern Neural Network Architectures
Dr. C. Krishna Priya
Assistant Professor
Department of Artificial Intelligence and Data Science Central University of Andhra Pradesh, Ananthapuramu, India Email: krishnapriyarams@cuap.edu.in
Mantri Vamsikrishna
Department of Artificial Intelligence and Data Science Central University of Andhra Pradesh, Ananthapuramu, India Email: vamsikrishnamantri@gmail.com
Abstract—This study conducts a comprehensive evaluation of six machine learning models—Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), Autoencoder, Random For- est, XGBoost, and Logistic Regression—for detecting fraudulent transactions in e-commerce, addressing escalating fraud losses projected to reach $60 billion by 2027 and false positive costs of $50 billion annually. A balanced dataset of 19,008 trans- actions (9,504 fraudulent, 9,504 legitimate) was preprocessed using SMOTENC, normalization, one-hot encoding, and noise reduction techniques (ENN, Tomek Links). Models were assessed using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrices, and SHAP for interpretability. Random Forest achieved superior performance (84% accuracy, 0.90 fraud recall, 0.85 F1- score), detecting 8,554 fraudulent transactions with 2,186 false positives, followed by XGBoost (77% accuracy, 0.85 fraud recall). Neural networks (MLP, LSTM, Autoencoder) underperformed due to the tabular dataset’s lack of sequential or distinct anomaly patterns, with LSTM failing entirely (0.00 fraud recall). SHAP analysis identified transaction amount, shipping distance, and time of day as critical predictors. Random Forest and XGBoost are recommended for real-time API deployment, offering scala- bility and GDPR-compliant interpretability. Limitations include the balanced dataset’s mismatch with real-world fraud sparsity (1–2%), high computational costs for neural networks, and the need for threshold optimization to reduce false positives. Future work should explore imbalanced datasets, hybrid models, and federated learning for privacy-preserving fraud detection.
Index Terms—E-commerce, Fraud Detection, Deep Learn- ing, Neural Networks, Random Forest, XGBoost, SHAP, Inter- pretability
