Research On Multi-View Feature Selection And Semi-Supervised Support Vector Machine

With the rapid development of society and the continuous advancement of science and technology,people's information interaction through the Internet is becoming more and more frequent,followed by the explosive growth of data.How to accurately and efficiently mine effective information in big data has gradually become a focus of attention.Increasing data dimensions and scales have brought about problems such as "dimension disaster",high computational complexity,data redundancy,and the expensive cost of label acquisition.Therefore,feature selection,semi-supervised learning and unsupervised learning gradually enter people's field of vision.And analyzing problems from a single view is often limited.If you can observe the same thing from multiple views and give a comprehensive assessment,it can greatly improve the effectiveness of data mining.Taking the above problems as the starting point,this dissertation focuses on the multi-faceted nature of the things and the subspace structure information of the data,and makes the following work and contributions:1.In the third chapter,this dissertation proposes a new multi-view unsupervised feature selection algorithm.This method can learn the global sparse solution of the projection matrix,and adjust the regression coefficients of the least squares regression by using a set of scale factors for evaluating the importance of the features.Finally,the scale factor is embedded into the projection matrix,thereby extending the least squares regression model.The introduction of the scale factor gives us a definition and provides a theoretical explanation for the rank of features using the projection matrix.With reference to the embedding of the feature weight scale factor,the view weight can be embedded into the projection matrix to adjust the regression coefficients to achieve a measure of the importance of data from different views,so that the feature weight and view weight can be obtained automatically while optimizing the projection matrix,thereby reducing model complexity and improving model stability.Because evaluating samples from any view does not change the category information between samples,this dissertation uses the form of sample self-representation and uses the sample self-representation matrix to construct the connection between views in unsupervised multi-view learning.In order to optimize the new model,this dissertation proposes a simple and effective convergence algorithm,and obtaining the number k of adjacent samples in the solution process avoids the trouble of setting k in advance.Through experiments to compare with a number of current excellent feature selection algorithms,we can see the superiority of the multi-view unsupervised feature selection algorithm proposed in this dissertation.2.In the fourth chapter of this dissertation,a novel semi-supervised adaptive Laplacian support vector machine method is proposed,and its primal solution is given because relevant research has shown that the primal solution of a support vector machine has more advantages than its dual solution.Compared with other Laplacian support vector machine algorithms,this dissertation also considers the hinge loss of unlabeled samples to maximize the distance between different classes of unlabeled samples to improve the classification performance and generalization ability of support vector machine models.At the same time,the method proposed in this dissertation trains the Laplacian matrix and support vector machine at the same time to improve the adaptability and accuracy of the Laplacian matrix,and introduces the primal solver to solve the adaptive Laplacian support vector machine problem.After comparing the experimental results with several state-of-art semi-supervised support vector machine algorithms,it can be seen that the semi-supervised support vector machine algorithm proposed in this dissertation has a good classification effect.