Study On Multi-class Classification Method Of Support Vector Machine Based On Decision Tree

Problems of Multi-class classification widely exist in real life applications. In 1990 s, one new kind of pattern recognition method—Support Vector Machine was built on the basis of rigorous statistical theory by Vapnik etc. It exhibits many unique advantages when solving the problem of pattern recognition with small-scaled samples, nonlinear and high dimension. Although the SVM was first raised for classification problems of two classes, it has already been used in solving problems of multi-class classification now. In connection of disadvantages of currently existing multi-class classification algorithms such as low accuracies or speeds, the thesis raised an improved multi-class classification algorithm of SVM based on decision tree which efficiently improved the accuracy of data set. This thesis mainly studies:(1)In connection of the process of multi-class classification of SVM based on decision tree, analyze the root of error accumulation, build partial binary decision tree with a suitable measurement of class separation; Moreover, ensuring the classification accuracy of negative classes and correcting those possibly misclassified positive samples, this two things should be done when the method is applied in the support vector training and recognition process based on the imbalance training data sets on tree node, then an improved multi-class classification algorithm is raised. Simulation experiments illustrate that while its running time is still feasible, new algorithm can improve the classification accuracy.(2)Realize handwritten digit recognition with the improved SVM multi-class classification algorithm based on decision tree, then good results are got.