ACLEGR-TADD: Adaptive Continual Learning for Financial Fraud Detection under Extreme Class Imbalance

Chen Yanyang, Yilin Su, Jing Zhu

Keywords: continual learning, financial fraud detection, extreme class imbalance, concept drift, temporal attention, wavelet analysis, catastrophic forgetting, differential privacy

TL;DR: We introduce ACLEGR-TADD, a continual learning framework that combines attention-based and wavelet drift detection to achieve 94.7% PR-AUC on financial fraud detection despite extreme class imbalance (0.2% fraud rate) and adversarial concept drift.

Abstract: Financial fraud detection systems face catastrophic performance degradation under adversarial concept drift and extreme class imbalance, where fraud comprises less than 0.2% of transactions. Existing continual learning methods fail as they assume balanced classes and static distributions. We propose ACLEGR-TADD, a novel framework that integrates Temporal Attention-based Drift Detection (TADD) with multi-resolution wavelet analysis, achieving a 4-fold reduction in detection delay (from 4.8h to 1.2h). Our method incorporates a Fraud-Aware Variational Memory Network (FA-VMN) that leverages class-specific variance disparities and Information-Theoretic Adaptive Consolidation (ITAC) using PAC-Bayes bounds. We provide the first catastrophic forgetting bound under extreme imbalance, proving that forgetting scales with the square root of the fraud rate over sample size $\\mathcal{O}\\left(\\sqrt{\\frac{\\rho}{n}}\\right)$. Experiments on five datasets comprising over 10 million transactions demonstrate that ACLEGR-TADD achieves 94.7% PR-AUC with sub-10ms CPU inference latency, significantly outperforming DER++ (75.6%) and FT-Transformer (78.1%). The framework satisfies differential privacy with formal guarantees while reducing false positives by 64% in production deployment.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 15593