Classification for high-dimension low-sample size data

High-dimension and low-sample-size (HDLSS) data sets have posed great challenges to many machine learning methods. To deal with practical HDLSS problems, development of new classification techniques is highly desired. After the cause over-fitting phenomenon identified, a criterion for sets, termed tolerance similarity, proposed emphasize maximization within-class variance on premise class separability. Leveraging this criterion, novel linear binary classifier, No-separated Data Maximum Dispersion classifier (NPDMD), designed. The main idea NPDMD spread samples two classes in large interval respective positive or negative space along projecting direction when distance between projection means enough. salient features are: (1) operates well sets; (2) solves objective function entire feature avoid data-piling phenomenon. (3) leverages low-rank property covariance matrix accelerate computation speed. (4) suitable different real-word applications. (5) can be implemented readily using Quadratic Programming. Not only theoretical properties been derived, but also series evaluations conducted one simulated six real-world benchmark including face mRNA classification. Experimental results comprehensive studies demonstrate superiority terms correct rate, mean within-group rate area under ROC curve.