RIFF Algorithm for Fraud Detection from Decision Trees

This technical report presents RIFF, a rule induction algorithm designed to generate low false positive rate (FPR) rule sets for fraud detection by leveraging decision trees. The document outlines the challenges faced in traditional fraud detection systems, which often rely on manually crafted rules that require significant expert input and may not perform as well as modern machine learning models. RIFF aims to automate the rule creation process, thereby reducing the complexity of rule sets while maintaining or improving performance. The report details the methodology of RIFF, which includes inducing candidate rules from decision trees and selecting the best-performing rules based on their precision and FPR. The evaluation of RIFF is conducted using both publicly available and private datasets, comparing its performance against existing decision tree algorithms and expert-tuned rules. The findings indicate that RIFF can effectively produce rule sets that outperform those created by human experts, making it a valuable tool for enhancing fraud detection systems.