A novel Multi-surface Proximal Support Vector Machine Classification model incorporating feature selection

The currently proposed Multi-surface Proximal Support Vector Machine Classification via Generalized Eigenvalues (GEPSVM) is an effective method on 2-class problem, which only needs to proximally solve two not parallel planes corresponding to each of two data sets, and the planes can be easily obtained by solving generalized eigenvalues. However, this approach can not effectively constrain the effect of those irrelevant or redundant features. To overcome this drawback, in this paper, we introduce a novel multi-surface proximal support machine classification model incorporating feature selection, which simultaneously implements classification and feature selection for improving the classification performance. Based on this model, we propose a linear multi-surface classification algorithm by a greedy nonexhaustive search strategy(called GEPSVMFS). Further, we develop a non-linear classifier by using kernel trick (called KGEPSVMFS). Experiments show that two algorithms of this paper have better or comparable classification performance as compared to GEPSVM on almost all benchmark data sets.