Support VectorMachines have been a dominant learning technique for almost ten years, moreover they have been applied to supervised learning problems. Recently twoclass unsupervised and semi-supervised classification problems based on Bounded C-Support Vector Machines are relaxed to semi-definite programming[7]. In this paper we will present another version to two-class unsupervised and semi-supervised classification problems based on Bounded ?-Support Vector Machines, which trained by convex relaxation of the training criterion: find a labeling that yield a maximum margin on the training data. But the problems have difficulty to compute, we will find their semidefinite relaxations that can approximate them well. Experimental results show that our new unsupervised and semisupervised classification algorithms often obtain more accurate results than other unsupervised and semi-supervised methods.