In-sample Model Selection for Trimmed Hinge Loss Support Vector Machine

摘要

In this letter, we target the problem of model selection for support vector classifiers through in-sample methods, which are particularly appealing in the small-sample regime. In particular, we describe the application of a trimmed hinge loss function to the Rademacher complexity and maximal discrepancy-based in-sample approaches and show that the selected classifiers outperform the ones obtained with other in-sample model selection techniques, which exploit a soft loss function, in classifying microarray data.