Abstract: To resolve the unclassifiable regions and to elevate generalization ability of multi-category support vector machines at present，an improved multi-category SVM based on Binary Tree（BT-SVM） is proposed.The improved algorithm defines the decision-radius which comes from the category-hypersphere-radius that is divided into kernel-radius and max-radius based on Pareto principle.The final BT’s structure is created by the decision-radius that can make the category which has wide-distribution and small-scatter on the top of BT，therefore the measurement errors are avoided.Numerical experiments on standard dataset demonstrate that the improved BT-SVM can better deal with practical multi-category classification problems with more adaptive ability.