Algorithm Study Of Graph-based SVM Classification Under Semi-supervised Learning Framework

In the field of machine learning,Support vector machine(SVM)algorithm is an early supervised learning algorithm,which solves the problem of over-fitting and "dimension disaster" in early neural networks,and has played a very good application in many fields.Semi-supervised learning can effectively use labeled samples and unlabeled samples to fully exploit the clustering information of the whole samples.Compared to the supervised learning,the number of labeled samples is not high,and the performance is better.Among them,Graph based semi-supervised learning is the most popular semi-supervised algorithm.In this paper,The algorithm study of graph-based SVM classification under semi-supervised framework is proposed to incorporate the feature information of unlabeled samples into the training process and improve the classification accuracy of SVM algorithm.Firstly,the pseudo-label is given to the unlabeled samples by using the graph model,and then the pseudo-labeled samples and the labeled samples are input into the SVM algorithm.In the process of producing pseudo-labeled samples,there may be noise samples.We should first denoise the pseudo-labeled samples to avoid the positive effect of the noise samples to reduce the set of extended training samples.In addition,the higher the accuracy of the pseudo-labeled samples,the less the noise samples,the more valuable the sample information,the workload will be reduced.Therefore,we attempt to select a graph with better classification and better performance in the pre-processing phase of the extended training samples,and complete the experiment with the SVM algorithm.The main work of this paper is as follows:(1)The first stage,.For the UCI data set and the USPS handwritten data,the exponential weighting(EW),the k-nearestneighbors graph(kNN),the1? norm(LN)and the low rank representation(LRR)were experimented and analyzed.Finally,the low rank representation(LRR)is used as the pre-processing process of the sample.Different graphs are combined with Gaussian random field and harmonic function(GHF)propagation algorithm to complete the classification experiment.(2)The second stage,The low-rank representation(LRR)model is given a pseudo-labeled sample using the k-nearest neighbors graph algorithm to remove the noise samples by comparing the label values.The experiment results show that proposed algorithm can fully exploit the overall samples information and it has higher classification accuracy in the case of fewer labeled samples in the end.