The Semi-supervised SVM Classification Algorithm Rsearch And Application Based On Graph Model

Support vector machine algorithm is developed on the structural risk minimization. Itavoids the over-fitting, easy fall into local minima and dimension disaster problem in previousneural network learning. It has been widely used in many areas. The classical SVM algorihtm isa early developed algorithm in supervised classification algorihtm.But it relys heavely on thenumber of labeled samples. While only minority samples have been labeled and other samplesare unlabeled in practice.Now a new semi-supervised leaning algorithm is becoming the current hot research ofmachine learning area.Because it can combine the cluster information provided by labeled andunlabeled samples,and can solve practical prlblems better than traditional classificationalgorithm.The current improving algorithm on semi-supervised SVM focus mainly on adding theunlabeled sample imformation to the optimazition function.But such simulations are sensitive tothe nosiy samples and encounter optimazition problem.In order to solve these problem,this paperproposes the semi-supervised support vector machine classification algorithm based ontwo-stage-learnin. To provide initial pseudo labels for the unlabeled samples, the graph model isused. And k-nearest graph is applied to distinguishing and removing the possible noisy samples.Then the denoised samples are inputted into the support vector machine (SVM) as labeledsamples, so that the global information of the whole samples can be utilized by SVM when it isused in the training to improve the classification accuracy.Meanwhile,we propose thesemi-supervised support vector machine classification algorithm based on gauss mixture modelkernel.The unlabeled data information is incorporated into the gauss mixture model kernel inorder to improve the SVM algorithm’s classification performance under few labeded samples.This is achieved by not only learning the labeled sample,but also considering the clusteringinformation of the whole training samples. The experimental results show that the proposedmethod can not only improve classification performance of SVM in few training samples, butalso increase the overall classification robust performance.This paper firstly proofs the feasibility of models from theory,then use the synthesized andUCI sample sets to test the classification performance of the algorithm,and compare with the existing other semi-supervised support vector machine algorithm.The results verify theadvantage of the models.At last,we utilize the algorithm to test the bearing fault detectionperformance and the effcet of the important parameters on the algorithm’s performance.Theexperiments demonstrate that the proposed models’ detection performance are better than otheralgorithms.