Waleed, Muhammad, Iqbal, Nasir and Alsafi, Tariq ORCID: https://orcid.org/0000-0002-9244-9352 (2026) Bayesian-Optimized Gradient Boosting for Credit Card Fraud Detection: A Cost-Sensitive Evaluation Under Hybrid Resampling. In: Arai, K. and Lorenz, P., (eds.) Intelligent Computing. CC 2026. Lecture Notes in Networks and Systems (2064). Springer, pp. 226-240

Full text not available from this repository.

Abstract

Credit card fraud detection under extreme class imbalance demands classifiers that are not only statistically accurate but also operationally viable when deployed at scale. This study presents a comparative evaluation of four gradient boosting algorithms, which are XGBoost, LightGBM, CatBoost, and HistGradientBoosting, against four non-boosting baselines on the benchmark European cardholder dataset (284,807 transactions, 0.17% fraud). A two-stage hybrid resampling pipeline combining random under-sampling and random over-sampling is applied exclusively within training folds to preserve evaluation integrity. Hyperparameters for all boosting models are tuned via Bayesian optimization using Optuna, with area under the precision-recall curve (AUPRC) as the optimization objective. Decision thresholds are calibrated by maximizing F1-score on training predictions, eliminating test-set leakage. Statistical significance is established through repeated stratified 5 × 5 cross-validation with Friedman omnibus testing and pairwise Wilcoxon signed-rank comparisons under Holm-Bonferroni correction. Results confirm that optimized boosting models achieve significantly higher AUPRC (0.88–0.89), F1-scores (0.81–0.88), and MCC (0.83–0.89) compared to non-boosting baselines. A cost-sensitive financial impact analysis grounded in the LexisNexis study ($5.75 per dollar lost) reveals that the primary deployment advantage of boosting models lies in operational efficiency as HistGradientBoosting achieves the highest benefit-per-alert ($609) with the lowest alert rate (0.13%). SHAP-based interpretability analysis identifies V14, V4, and transaction timing as the dominant fraud predictors. These findings establish that optimized gradient boosting algorithms deliver superior fraud detection not merely through marginal metric gains but through their capacity to minimize false alarms while sustaining high fraud recovery, reducing the cost of each investigation in resource-constrained environments.

Item Type:	Book Section
Status:	Published
DOI:	10.1007/978-3-032-31755-1_16
Subjects:	T Technology > T Technology (General)
School/Department:	London Campus
URI:	https://ray.yorksj.ac.uk/id/eprint/15760

University Staff: Request a correction | RaY Editors: Update this record