Authors: Mrs.N.Zabina
Abstract: With growing adoption of e-commerce platforms, the problem of identification of fraudulent financial transactions in a real-time manner has increased manifold. Current approaches of rules-based approaches and stand-alone machine learning models have faced difficulty adapting and dealing with novel types of fraud cases when there is heavy class imbalance and only few examples of frauds out of the total volumes. In this research, we propose the novel hybrid system of fraud detection for e-commerce using techniques of isolation forest, representation learning via autoencoder, and supervised learning using xgboost for classification. Unsupervised learning using anomaly detection is used to identify statistical anomalies as well as reconstruction-based outliers and supervised gradient boosting approach is used to improve the decision boundaries on labeled data. Evaluation experiments performed on large-scale anonymous transaction dataset show better discriminative performance of our hybrid approach as compared to the current benchmark models with respect to both the F1-score and AUC score. An explainable approach using SHAP values is provided as part of our framework, complying to regulations on explainability.
DOI:
