Abstract:A supervised classification method having a local learning ability, called weighted projection twin support vector machine ( WPTSVM) , is proposed. This method aims to improve upon a defect that projection twin support vector machines ( PTSVMs) have, namely, that PTSVMs do not take account of the local structure and local infor?mation of a sample space in the training process. Compared with PTSVM, WPTSVM improves its local learning a?bility to some extent by attaching different weights for each sample according to the within?class neighborhood graph and replacing the standard mean with a weighted mean. Moreover, to reduce computational complexity, WPTSVM chooses a small number of boundary points in the contrary?class based on the between-class neighborhood graph to construct constraints of the original optimization problems. The method inherits the merits of PTSVM and can be re?garded as an improved version of PTSVM. Experimental results on artificial and real datasets indicate the effective?ness of the WPTSVM method.

An improved projection twin support vector machine