This work proposes a kernel amplification method with non-stationary characteristics for binary classification of non-noisy imbalanced datasets. Our methodology features two key innovations, including that a derived non-stationary kernel construction enables adaptive exploration of minority class regions, and a Riemannian metric–guided kernel amplification mechanism effectively induces minority class migration in feature space, tightening the spatial distance inner minority class instances. Experimental validation across ten UCI benchmark datasets with class imbalance demonstrate the superior performance of our proposed method. The method achieves statistically significant superiority over all six baseline approaches on five highly imbalanced datasets (with imbalance ratios (IR) > 10:1), notably achieving 0.883 F1-score on datasets with 40.22:1 imbalance ratio and 0.800 sensitivity to the minority class. Furthermore, our approach maintains competitive advantages on the remaining five moderately imbalanced datasets (IR < 10:1), outperforming a subset of the baseline methods across all evaluation metrics. Furthermore, the kernel amplification mechanism boosts the sensitivity to perception minority classes by a maximum 6.35-fold enhancement on highly imbalanced datasets, and by a maximum 2.17-fold enhancement on moderately imbalanced datasets. The derived amplification factor exhibits dimension-dependent characteristics, showing independence from both sample size and imbalanced ratio——a critical advantage for high-dimensional imbalanced classification.
Zheng et al. (Tue,) studied this question.