Abstract: Many pattern classification methods,such as support vector machine and L₂ kernel classification,often use the kernel methods and are formulated as quadratic programming problems,but computing kernel matrix would require O(m²)computation,and solving QP may take up to O(m3),which limits these methods to train on large datasets．In this paper,a new classification method called difference of similarity support vector machine(DSSVM)was proposed.DSSVM pursues a best linear representation of a total similarity between any sample and a particular class．According to the sparsity of the linear representation and the max margin of the difference of similarity,a new optimization problem is obtained．Meanwhile,the difference of similarity support vector machine can be equivalently formulated as the center constrained minimal enclosing ball,and thus difference of similarity support vector machine can be extended to diffe-rence of similarity core vector machine(DSCVM)by introducing fast learning theory of minimal enclosing ball,to solve the classification for large datasets．This is confirmed in the experimental studies.

Key words: Difference of similarity,Sparsity,Core set,Minimum enclosing ball,Support vector machine