An Interpretable and Drift-Aware AI Framework for Real-Time Financial Fraud Detection in Large-Scale Transaction Systems

Abstract

Real-time fraud detection in payment and banking infrastructures is constrained as much by operating conditions as by model capacity. Effective systems must separate rare fraudulent activity from a dominant legitimate population, remain reliable as adversaries adapt (concept drift), and deliver decisions within strict latency budgets. This paper presents a deployable fraud detection framework for large-scale transaction streams that couples a low-latency gradient- boosted decision tree (GBDT) scorer with graph-derived relational signals, and embeds the resulting model within an explainability and governance layer designed for auditability.
We describe the end-to-end pipeline—stream ingestion, feature computation with online/offline parity, model training and calibration, online serving, and continuous monitoring—and evalu- ate the approach on anonymized, benchmark-style transaction data using time-sliced splits to approximate production drift. The empirical results show consistent, incremental gains over representative baselines in AUC and in false positive rate at fixed recall, while preserving deci- sion evidence suitable for operational review. Practical implications for transaction trust, loss mitigation, and the resilience of digital financial infrastructure are discussed.