In this talk, we present a regression method, called delta support vector regression (deltasvr), that replaces a regression problem with binary classification problems which are solved by support vector machines (svm). The results indicate that deltasvr achieves comparable to epsilon insensitive support vector regression generalization error, fewer support vectors, and smaller generalization error over different values of epsilon and delta. We present also a method called phi support vector classification for incorporating knowledge about margin of an example for classification and regression problems and two applications: decreasing the generalization error of reduced models, and incorporating the nonlinear constraint to the svm optimization problem. Finally, we present a concept of the balanced support vector classifier which balances distances from points to the decision boundary.