Research On Sparsity Based Machine Learning Methods

With the rapid development of information technology, machine learning has developed as one of hottest fields in artificial intelligence research. Sparsity has been successfully used to develop more efficient learning machines. The kernel learning and spectral graph theory can be understand easily and have good generalization prformance, which are hot issues in the field of machine learning problem. This paper introduces the sparsity into the construction of kernel function and graph matrix to raise three classification algorithms, and test their performance in some open data sets in machine learning. The contents mainly include the following aspects:Firstly, based on the existing sparse coding classification algorithm, this paper proposes a sparse center kernel based coding classification algorithm. This method firstly constructs a center samples matrix in the dictionary by selecting k-nearest neighbors, and then maps the dictionary and the test sample into the feature space by the sparse kernel function, which is derived from the center samples matrix. Finally, in the feature space, we can have the sparse coding of the new testing samples on the new dictionary, and then utilize the sparse coefficients and error discriminant function to identify the label of samples with unknown labels. Because mapping by the sparse kernel function, our algorithm can enhance the separability of data and reduce the iteration times to calculate kernel function simultaneously. As a result, the classification accuracy can be improved without much time lost. Some experiments are taken on commonly used facial image databases and handwritten digit databases to test the performance of our proposed method.Secondly, this paper proposes a kernel l1, graph based semi-supervised classification algorithm. Based on the existing sparse l1, graph method, we introduce the idea of kernel learning, which can make nonlinear separated data more seperately. Combine the kernel l1, graph with semi-supervised framework to construct the algorithm of kernel l1, graph based semi-supervised classification. The kernel l1, graph can enhance the similarity of samples in the same class and the difference of samples between different classes, which can achieve higher classification accuracy. Some experiments on artificial helix data, standard facial image databases and handwritten digit databases show that the proposed method can achieve higher classification accuracy than existing algorithms.Thirdly, we propose a kernel low-rank based subspace segmentation method. Based on the traditional method of low-rank based segmentation, our method utilize the kernel function to map samples into the feature space firstly, then solve the low-rank coefficients matrix jointly to constuct the undirected graph, which can be used in the spectral clustering to finish the subspace segmentation. Because the kernel mapping can improve the seperability of data, the segmentation accuracy can be increased. Some experiments on the artificial data, the ORL facial image database and a handwritten digit database are taken to investigate the performance of our proposed algorithm. Compared with some available methods, our method can achieve higher segmentation accuracy.The research is supported by NSFC(61072108,60601029,60971112,61173090), new century excellent talents item(NCET-10-0668), Higher school subject innovation engineering plan (111plan), No. B0704and central university basic scientific research business expenses.