Isolation Forests with Local Outlier Factor and Mahalanobis Distance for Real‐Time Fraud Detection

This chapter introduces a complex real-time fraud detection model that encompasses the Isolation Forests, local outlier factor (LOF) and Mahalanobis distance to detect fraud transactions in the high-dimensional streaming payment data. The fraud detection issue was to classify the transactions as legitimate or fraudulent according to the features calculated due to transaction history and contextual information. Isolation Forest groups anomalies by using recursive random partitioning of the feature space where anomalies are grouped in fewer partitions compared to the standard observations. Mahalanobis distance gives principled statistic anomaly detection based upon feature correlations and variance structure. Future studies will build the framework to support explainability mechanisms that would give fraud investigators clear rationales on how they can detect fraud, active learning methods that would allow human-in-the-loop feedback to quicken model adjustments, graph-based anomaly detection that would capture the payment network structure and money flow, and contextual information which would help in making finer fraud risk indications.