Researches On Hierarchical Granular Support Vector Machine Method

Support Vector Machine (SVM) as a kind of learning machine, is concerned broadly, and it is based on statistical learning theory of VC dimension theory and structural risk minimization principle. With concise mathematical forms, standard training algorithms and excellent generalization performance, SVM has been widely applied in data mining issues like pattern recognition and time series prediction. Because the learning efficiency of SVM often depends on the size of the given data, SVM has not achieved the efficiency as expected, especially in dealing with the large-scale data sets for the practical problems. Therefore, how to improve the learning efficiency of SVM and guarantee the generalization performance at the same time becomes the one of focus in SVM researches, and it is also the purpose of this thesis.Support vector machine learning algorithms is prone to produce redundancy samples in the process of training. Different samples have different contribution to the training, the samples with higher contributions are easily to wrongly classified, and samples that are always classified correctly have lower contributions. Therefore, selecting samples with higher contribution to classification hyperplane to form hierarchical support vector machine is one of the methods to increase efficiency of training process.This thesis integrates the hierarchical classification and granular computing theory with the traditional SVM method effectively, and builds the hierarchical granular support vector machine learning mechanism. By defining a factor to measure the importance of the granule to pick out the valuable granules, the redundant granules are removed, using some representative point in the reserved granules formed support vector machine training data. Because of the learning training set only include important classification information, the classification speed is much faster than the SVM, Furthermore, the training can be executed in different levels of granules. In doing so it can obtain satisfactory generalization ability which is superior to the traditional granular support vector machine.The main works are concluded as follows:(1) Detailed the structure and principle of the traditional SVM model, pointed out the main problems of SVM algorithms in dealing with classification problem. The thesis also explained the advantages and disadvantages of granular SVM algorithm in details, exported and researched to solve these issues.(2) Put forward a hierarchical granular support vector machine learning method for large datasets classification, which defines a new data confidence to pick out the valuable data sample (mean the sample which is bigger contribution the decision boundary), it can automatically divide granules in each layer of the training according to the distribution, and hope to get a better generalization ability.(3) Proposed a dynamic granular support vector machine algorithm which is used to dealing with distribution uniform data set. According to the different distribution of the granules, some granules will be divided automatically and SVM training will performed in different level of granule space. The model not only can effectively overcome the low training efficiency of the traditional SVM when dealing with large-scale data set, but also can achieve a better generalization performance.(4) Conducted a series of experiments on standard UCI datasets to verify the proposed hierarchical granular support vector machine learning algorithm, and achieved expected effect. Experimental results show that the proposed hierarchical granular SVM algorithm and dynamic granular SVM algorithm are very efficient Comparing with the classical SVM and the tradition granular SVM.The proposed hierarchical granule support vector machine learning mechanism can effectively solve the low efficiency classification problem that existing in large dataset and distribution uniform dataset. The research achievement results can not only enrich the application of SVM, but also provide a beneficial exploration of machine learning method in practical application based on the granular cognition.