A new improved FSVM algorithm based on SVDD

摘要

SVM (Support Vector Machine) is a popular machine-based learning method, with wide applicabilityand excellent generalization performance in a large number of real-time classificationproblems.However,SVMis sensitive to the noise and outliers.As an improved algorithm based onSVM, FSVM (Fuzzy Support Vector Machine) gives the training samples different fuzzy membershipvalues in order to reduce the interference of the noise and outliers.However, like the normalSVM algorithm, the FSVM algorithm still needs to solve the problems such as how to improve theaccuracy of classification and how to accurately recognize the noise and outliers. In this paper, animprovedFSVMof data classification algorithm (IFSVM) hasbeen proposed.Firstly, this algorithmdeletes the outliers or noise based on the average density algorithm and removes them from thesamples, thus avoiding the influence of the noise to the classification accuracy. Secondly, the centersand radiuses of the two minimum hyperspheres are extracted based on the SVDD algorithm.Finally, this algorithm sets the membership function values by comparing the distance betweeneach sample and the center of the sample's hypersphere, the distance between the sample and theopposite hypersphere center, and the distance between the two hypersphere centers. In thisway,this algorithm highlights the importance of the boundary vectors,which could be support vectorsfor classification,andimproves the classification accuracy.The experimental resultsshowthat thisalgorithm has improved the anti-noise ability and the accuracy of classification when comparedwith the normal SVM algorithm and three other FSVM algorithms.