Attribute Reduction And Classification Decision Based On Support Vector Data Description

Classification widely exists in human activities.People distinguish different objects from their varied features and attributes.Nevertheless,there are too many attributes to make sense in the actual productive activity.In this way,reasearchers have started to introduce artificial techniques to classification procedure of computers,and made a certain achievement.However,numerous attributes are required to exactly describe objects and improve classification performance,which would cause the issues of space and time complexity.Under such situation,attribute reduction becomes one of important subjects.Approaches for reducing attribute including feature selection,feature elimination,and feature extraction have been used to simplify feature infomation,and obtain core features from the high-dimensional data with least space as soon as possible.Compared to utlizing all original features,it is more effective to get attribute subsets by adopting attribute reduction methods,which balances classfication accuracy and computation complexity.This paper focuses on support vector data description(SVDD),and combines it with attribute reduction to make classification and decision.Detailed work is described as follow.A dual feature elimination method based on multiple SVDD models is put forward.SVDD solves a max-dual quadratic programming using only target samples and gives a unique optimal solution.The dual feature elimination method achieves the dual objective function for each type of data and removes the worst feature according to the dual ranking criterion.Experiments validate that the method can effiectively reduce attribues and improve the classfication speed.A radius-recursive feature elimination method based on multiple SVDD models is proposed.In the new method,SVDD builds the closed boundary surrounding with each class data and rejects novel data outside the boundary.Each hyper-sphere model is established with center and radius.For multi-classfication,the new method ranks each feature according to radius of hyperspheres.The feature with the smallest ranking value would be eliminated.By doing so,we can get the critical feature subsets.Experiments on several datasets validate that radius-recursive method can enspeed the classification procedure.Compared with the dual feature elimination method,this method has a better classfication performance in the case of low-dimensional data.An orientation-distance feature extraction method based on SVDD is presented.As mentioned,attribute reduction approaches trim data down by removing bad features.However,there are no clear differences between good and bad features in practice.Attributes may have complicated connections each other.After removing some features,it is possible to loss the correlation between features.The orientation-distance feature extraction method based on SVDD devotes to reserve all attributes of original data,and calculates the distances between the data from different classes and hyper-sphere models to create new features.The proposed method reduces the dimensionality of data and speeds up the process of classification decision,while it keeps useful information of the original data.