Research On Multi-classification Method Based On Support Vector Machine

Support Vector Machine(SVM) is one of the machine learning method based on VC dimension theory and structural risk minimization principle(SRM). It shows many unique advantages in solving high-dimensional pattern recognition, small sample and nonlinear problems, and largely overcomes the issues of "curse of dimensionality", "over-learning" and partial minimum in traditional machine learning. SVM is a hot research topic in the field of machine learning, it shows apparent advantages in the classification and makes a large number of research results, it is widely used in many fields. The traditional SVM is essentially used to solve the two classification problems,but practical applications are mostly multi-classification problems. How to extend the binary classification ideology of SVM to multi-classification, and effectively solve the multi-classification problem are the important content of SVM research in recent years.The DT-SVM multi-classification method is widely used in practical applications, but different binary tree structures have big influence on the performance of SVM classifier.Euclidean distance is the simplest and the most common distance measure method,but there is a big problem when it is directly used to measure inter-class similarity. It only considers the Euclidean distance between the centers of two sample classes, when the Euclidean distance is equal among several sample classes, it can’t determine the similarity between different classes; For the existing sphere structure method, when all kinds of sample data imbalance, the classifier has tendency to predict, so it will reduce the accuracy of classification.At the same time,the building process of sphere structure is very complex. By improving the Euclidean distance, this text redefines a new inter-class separability measure function which uses the advantage of Euclidean distance’s simple calculation and solves the defect of sphere structure method.The multi-classification method of decision tree SVM exists "error accumulation" phenomenon, has a serious impact on the classification accuracy, and its classification time is not very satisfactory. For the defect of decision tree SVM classification, using the inter-class separability measure function, this text proposes a established method about full binary tree, studies its training process and classification process, and analyzes the time of training and classification. By building the full binary tree, and using each non-leaf node to represent a classifier to judge the category of the unknownsample. This method implements that by using a set of two classifiers to solve the multi-classification problems, and the classifiers in the same layer can work simultaneously, so it improves the speed of training and classification.Through theoretical analysis and example verification, this text compared the inter-class separability measure function and SVM multi-classification method with the traditional SVM multi-classification. Theoretical analysis and experimental results showed that the proposed algorithm in this text not only had faster training and classification speed, but also improved classification accuracy.