Inference for class-imbalanced and nonseparable SVM data

Extend the proportional-asymptotic inferential framework for the L1-penalized support vector machine to class-imbalanced designs and to the nonseparable perceptron regime.

Background

The theory developed in the paper assumes balanced classes and a linearly separable Gaussian-mixture classification model. The authors explicitly identify both class imbalance and nonseparable data as settings not covered by the current framework, despite related statistical-physics analyses of these regimes.