Support vector machines based on positive feedback are put forward with the analysis of both advantages and disadvantages of current support vector machines based on Gaussian kernel function(including support vector machines based on K-nearest neighbors) and the essence why support vector machines is capable of describing data sets'distribution characteristics, thus the rigor constraint is overcome that maintains "corresponding parameters ofkernel function support vectors should be equal". The learning algorithm of support vector machines based on positive feedback is given. Simulation experiments of artificialand real data proves that support vector machines based on positive feedback is be obviously superior to current ones in its generation capabilities. Though, only the support vector machines based on positive feedback for pattern recognition is discussed, the idea included in support vector machines based on positive feedback is using kernel functions with different corresponding parameters to construct support vector machines, and is adaptive to other types support vector machines.