Methods Of Multiclass Support Vector Data Description Based On Extreme Learning Machine

Support vector data description is a supervised support vector machine based one-class classification algorithm, which is effective for data description and thus has great practical value in abnormal behavior detection, image classification, machine fault detection, etc. However, support vector data description exists some problems, such as when data distribution is not balanced or feature dimension is too low, the obtained training hypersphere will be large and the accuracy will not be satisfied, support vector data description cannot deal with the multiclass classification problem directly. In order to address the problems above, this thesis mainly studies the methods of multiclass support vector data description based on the extreme learning machine feature space.Firstly, the main ideas, merits and drawbacks of support vector machine, support vector data description and extreme learning machine are discussed in detail. Specially, extreme learning machine is a novel fast learning algorithm. Extreme learning machine derives good classification as long as the number of hidden layer nodes is large enough, which offers foundation of improving support vector data description algorithm.Secondly, for the low accuracy problem of support vector data description, the support vector data description based on extreme learning machine feature space is proposed. Specifically, the algorithm maps the original low dimensional input data to the high dimensional extreme learning machine feature space at first, and then the support vector data description is utilized for the mapped data to get the corresponding hypersphere. The experimental results show that the proposed method can improve the accuracy of support vector data description efficiently and reduce hypersphere radius, when choosing proper output function and parameters of hidden nodes.Finally, in order to extend support vector data description to solve the multiclass classification problems and remain the ability of outlier detection, a multiclass support vector data description algorithm is proposed. At first, support vector data description is applied to the data in each category and the hypersphere of each category is obtained, and then we can identify test sample by comparing test sample with each hypersphere center distance and hypersphere radius, including correct class, error class, uncertain class or outlier class, etc. In order to further improve the accuracy of the multiclass data description algorithm and reduce the error rate, the outlier rate and the uncertain rate, the input data is mapped into feature space, the uncertain samples is classified again and the training hypersphere radius is expended. The experimental results show that with the proper parameters, the proposed multiclass classification support vector data description algorithm can solve the multiclass classification problems better.