Research On Multi-class Classification Method Of Improved SVM Based On FBT

Support Vector Machine(SVM), as a new machine learning method based on statistical learning theory, has become a hot research issue in the field of machine learning, because it can overcome the defect of neural network method such as unstable solution and bad generalization. Traditional SVM is developed for binary classification problems, but most of actual problems are multi-class classification. It is one of the important challenges how to effectively solve multi-class classification problem in recent year. The SVM multi-class classification algorithm based on binary tree has been widely adopted, but the different binary tree structure has a great impact on classifier's performance.Aiming at the shortage in time complexity and classification precision of SVM algorithm based on binary tree, an multi-class classification algorithm of improved sphere-structured SVM based on Full Binary Tree(FBT) is put forward. This algorithm can enhance the classification precision via using improved sphere-structured SVM, because it considered the situation from unbalanced samples, at the same time, a FBT with reasonable structure is established, the multi-class classification problem can be transformed into a series of binary classification problems, an internal node represents a classifier in binary tree, the classifiers can work simultaneously in the same layer, so it can advance the speed of training and classification. Experiment result shows that compared with other multi-class classification algorithms, the algorithm proposed can obtain better classification result.With the increase of sample numbers, traditional SVM needs to train all training samples, because it don't support incremental learning, so its training speed reduced significantly. For the shortage of sphere-structured SVM incremental learning algorithm in training time and classification precision, an multi-class classification incremental learning algorithm of improved sphere-structured SVM is proposed. According to multi-class classification algorithm of improved sphere-structured SVM based on FBT, this thesis analyses the KKT condition of sphere-structured SVM classifier and the influence of new increased samples on support vector set, the part of samples in incremental sample set, support vector in original training set and some samples within certain range of sphere were combined as new training sample set to reconstruct SVM classifier. It is tested on UCI standard data sets and compared with sphere-structured SVM incremental learning algorithm, the result shows that the algorithm can obtain shorter training time and higher classification precision.