Robust Fraud Detection Through Boosted Classifier Consensus

Abstract

Fraud detection in financial transaction systems must be robust enough to handle severely imbalanced datasets where fraudulent transactions constitute less than 1% of all transactions, and must continuously adapt to evolving fraud patterns that change as fraudsters develop new techniques to circumvent existing detection mechanisms. Individual classifiers often fail to capture the diverse range of fraud patterns present in real-world financial data, and their predictions may be unreliable on edge cases that fall near decision boundaries. This paper introduces a robust fraud detection approach based on boosted classifier consensus, where multiple weak classifiers are systematically enhanced using two complementary boosting techniques—AdaBoost and Gradient Boosting—to improve their individual learning capabilities and then combined through an optimized weighted consensus voting strategy to produce more reliable final detection decisions. AdaBoost operates by iteratively training weak classifiers and increasing the weights of misclassified samples in each round, forcing subsequent classifiers to focus on difficult-to-classify transactions. Gradient Boosting complements this by training classifiers to correct the residual errors of the ensemble, optimizing prediction accuracy through gradient descent in function space. The predictions from both boosted ensembles are combined through a weighted consensus mechanism where optimal combination weights are learned from validation data. Evaluation on a financial transaction dataset containing 500,000 transactions with 0.5% fraud rate demonstrates that the boosted consensus approach achieves 95.8% accuracy and 0.94 F1-score, reducing false positives by 62% and false negatives from 14.2% to 4.8% compared to single classifier approaches, providing a stable and effective solution for robust fraud detection in real-world financial applications.