Time-Decay Aware Incremental Feature Extraction for Real-Time Transaction Fraud Detection

Abstract

Real-time fraud detection in high-frequency transaction environments demands both high accuracy and low latency. Feature extraction is a critical bottleneck, often accounting for 60-80% of the end-to-end processing time. This paper proposes a time-decay aware incremental feature extraction method that achieves amortized O (1) computational complexity for per-transaction state updates of core statistical features (decay-weighted sum, count, mean, and variance).In contrast, sketch-derived distributional features (entropy, heavy-hitter indicators) require O(w) operations where w denotes sketch width. The proposed approach incorporates an exponential decay weighting mechanism that automatically reduces the influence of historical transactions, enabling the algorithm to capture temporal dynamics without recomputing entire historical windows. A lightweight sketch-based structure is designed to maintain compact behavioral representations with constant per-user memory consumption. Experiments on the Kaggle Credit Card Fraud Detection dataset demonstrate that the proposed method reduces feature-extraction latency by 73.2% relative to traditional sliding-window approaches while maintaining comparable detection accuracy (AUC-ROC: 0.9847 vs. 0.9862). The throughput reaches 44,843 transactions per second on a single CPU core, indicating potential for high-volume transaction processing.