When the distributions of 2 class samples are different

the classification results will be biased by using least square classifiers

such as proximal support vector machine (PSVM).Inevitably

this decision bias will cause non-minimal classification error rates.In the present paper

based on equivalent generalized eigenvalue decomposition model of PSVM

a novel optimizing samples distribution PSVM model is proposed

which can improve original PSVM decision.The model is constructed as a generalized eigenvalue decomposition model and contains an optimal samples distribution regularization item.It can maximize distances between correctly classified samples and decision boundary and minimize distances between misclassified samples and decision boundary.Experimental results under artificial datasets and 10 data subsets from UCI datasets show that using this novel model can adjust decision effectively and achieve better classification effects.