<![CDATA[Conventional Mahalanobis metric learning (MML) algorithms exhibit significant sensitivity to outliers and noise in training data, leading to biased distance metrics with poor generalization performance on unseen data, to address this limitation, we propose a systematic framework integrating tunable regularization with K-fold cross-validation for robust metric learning. Specifically, we augment standard MML objectives with a Frobenius norm regularization term λ‖M‖²_F to penalize solution complexity and control overfitting. Crucially, we employ K-fold cross-validation as a data-driven mechanism to automatically determine the optimal regularization hyperparameter λ* that maximizes generalization potential, the resulting learned metric M* demonstrates enhanced resistance to noise and superior generalization capability. Empirical evaluation across 12 benchmark datasets (including real-world noisy data like Food-101N and CheXpert) confirms that our approach significantly outperforms non-regularized baselines and manually tuned alternatives: It reduces overfitting to noisy training constraints by 13.8–22.4% and improves test accuracy on distance-based tasks (k-NN classification, clustering) by 10.3–17.2% under severe noise conditions (40% label flips, 30% feature corruption), these results establish that the synergistic combination of mathematical regularization and cross-validated hyperparameter selection provides a principled, effective solution for learning reliable Mahalanobis metrics in noisy real-world environments]]>
Copyrights © 2025
