Matrix-pattern-oriented classifier with boundary projection discrimination

The matrix-pattern-oriented Ho-Kashyap classifier (MatMHKS), utilizing two-sided weight vectors to constrain the matrix-based pattern, extends the representation of sample from vector to matrix. To further improve the classification ability of MatMHKS, we introduce a new regularization term into MatMHKS to form a new algorithm named BPDMatMHKS. In detail, we first divide the samples into three types including noise sample, fuzzy sample and boundary sample. Then, we combine the projection discrimination with these boundary samples, thus proposing the regularization term which concerns the priori structural information of the boundary samples. By doing so, the classification ability of MatMHKS has been further improved. Experiments validate the effectiveness and efficiency of the proposed BPDMatMHKS.