Research Of Support Vector Machine Based On Sample Cluster: Application To Multi-class SVM Based On Binary Tree

As the support vector machine algorithm (SVM) has a good generalization and learningability, it have been widely studied. The standard SVM algorithm has been already designed as aclassification problem and demonstrated the good performance in dealing with two-classificationproblems, but we usually encounter a lot of multi-classification problems in practice. Thereforehow effectively the two-classification SVM is extended to the multi-classification SVM iscurrently a hot research. At present, the Binary tree based on multi-classification SVM algorithm(BT-SVM) in the several existed multi-classification SVM algorithms has demonstrated thebetter classification performance and generalization. In addition, the general machine learningalgorithms including the SVM directly face the samples in mining or learning. But with moresamples and their properties, how the datas are reduced become a valuable research.Firstly, we detailedly elaborated on the SVM, study the inverse problem of the SVM thatmeans the SVM about supervised learning be extended to the unsupervised learning, and lead tothe maximum margin clustering algorithm (MMC).Secondly, we firstly propose a new concept of"Sample Cluster", study the learning ability of the SVM based on sample cluster in the large data,and formulate the SVM based on hypersphere algorithm (HS-SVM) and the MMC based onhypersphere algorithm (HS-MMC). Finally, we have a depth analysis of important influence ofbinary tree structure in generalization and classification accuracy. Secondly, we roundly discussthe influence of structure of binary tree in the BT-SVM to the generalization and accuracy in themulti-classification. To take the HS-MMC as a binary tree structure of the optimal generationstrategy, we design the BT-SVM based on maximum generalization algorithm (MgBT-SVM)starting from the largest generalization.The experiments show that our algorithm reflects the better performance in the training time,the test time and the classification accuracy.