Answers are grounded in this paper and cite their sources.
Severe class imbalance hinders the detection of card fraud, and applying resampling before partitioning the data inflates performance estimates due to information leakage. Mahalanobis distance has been used previously for outlier detection and in resampling procedures, but the behavior of two geometrically distinct selection rules, one local and one global, under the same metric for fraud detection has not been studied. In this manuscript, we propose and evaluate two subsampling strategies that retain only real transactions: Majority-class Extreme-aware UnderSampling (MEUS) matches each minority observation with a close majority candidate according to the Mahalanobis metric, and Furthest-point Extreme UnderSampling (FEUS) retains observations farther from the global centroid of the training set. Both use a global covariance matrix estimated from the training set, but apply distinct observation-selection rules. We compare them with five data-level resampling techniques on the European Credit Card Dataset (ECCD) and the IEEE-CIS Fraud Detection dataset using repeated stratified cross-validation, with resampling restricted to the training folds. Under severe imbalance, FEUS offered the most favorable compromise between precision and recall in ECCD and intermediate performance in IEEE-CIS, with low computational cost; MEUS was competitive in ECCD but lost precision in IEEE-CIS. A Euclidean ablation showed that incorporating covariance changes the balance between metrics without being uniformly superior. The effect of covariance undersampling depends on the classifier and the imbalance structure of the data, and we did not observe a method that dominated under all conditions.
No takes yet. Share an insight, caveat, or question.
Saavedra-Garrido et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: