Tiered Financial Fraud Detection Utilizing Precision Stratified Random Forest Assembly
Charles Gardner, Dan Chia-Tien Lo, Johng-Chern Chern, P. Paschos, Chung Ng · 2019
This study focuses on financial fraud detection via the creation of a three-tiered anomaly detection system. The system is constructed by tuning multiple random forest classifiers, each with different fitness functions. The process is done using a randomized grid search that optimizes the random forest parameters to match the fitness function. Once complete, the models are compared to form three-tiers of detected frauds with each tier containing a different level of precision. Separating detected frauds into different tiers allows for both high precision and high recall values. With this strategy, 96% of frauds are classified correctly while still maintaining a high precision of over 90% for 85% of detected frauds detected. Our studies show that the tiered random forest outperforms other algorithms such as SVM and logistic regression with a precision at 85% and a recall of 72%.