Gene selection of multiple cancer types via huberized multi-class support vector machine

This paper proposes a new type of regularization in the context of multi-class support vector machine for simultaneous classification and gene selection. By combining the huberized hinge loss function and the elastic net penalty, the proposed support vector machine can do automatic gene selection and further encourage a grouping effect in the process of building classifiers, thus leading a sparse multi-classifiers with enhanced interpretability. Furthermore, a reasonable correlation between the two regularization parameters is proposed and an efficient solution path algorithm is developed. Experiments of microarray classification are performed on the leukaemia data set to verify the obtained results.