Cost-Sensitive Learning, Simulated PU Learning, and One-Class Autoencoding for Extreme-Imbalance Credit Card Fraud Detection

Abstract

Extreme class imbalance makes fraud detection evaluation sensitive to both the ranking metric and the chosen operating point. This revised study presents a single-split benchmark on the Credit Card Fraud Detection dataset (284,807 transactions; 492 frauds) comparing cost-sensitive gradient boosting, cost-aware neural networks, simulated positive-unlabeled (PU) training, and a one-class autoencoder baseline. Two methodological corrections are explicit: the benchmark itself is fully labeled, so PU learning is implemented by hiding negative labels during training; and the autoencoder is treated as a one-class reconstruction baseline rather than as strictly self-supervised learning. Models are evaluated with average precision (AP), ROC-AUC, recall at fixed false-positive rates, and a simple monetary cost model in which each reviewed alert costs one unit and each missed fraud costs the transaction amount. LightGBM attains the best ranking performance (AP 0.824), while XGBoost attains the greatest monetary savings after validation-based threshold selection, reducing test cost from 8483.36 to 3891.01 cost units (54.1% savings). Logistic regression remains competitive by ROC-AUC, which highlights why ROC-based comparisons can be misleading in this setting. Among neural models, nnPU gives the best ranking performance, whereas focal loss yields the best savings among the deep baselines. The one-class autoencoder is materially weaker than supervised models but remains useful as a label-light reference detector. The central conclusion is that ranking quality alone does not determine the preferred fraud-screening policy; business utility and threshold choice must be specified jointly.

; ; ; ; ; ; ;