Study On Feature Selection Based On Semi-supervised And Unsupervised Learning

Feature selection is a key issue in machine learning and an efficient means to solve the"curse of dimensionality".It is widely applied in many scientific domains such as computer vision and biometrics.Therefore,research on feature selection algorithms is of great significance theoretically and practically.Recently,unsupervised and semi-supervised feature selection is becoming hot as the dimensions of samples explode dramatically.Unsupervised and semi-supervised feature selection algorithms can guarantee the generalization performance and reduce the labeling cost simultaneously.In this thesis,we focus on the study of semi-supervised and unsupervised feature selection.The main contents of this thesis include:?1?This thesis elaborates the feature selection based on sparse norm,the semi-supervised learning based on manifold learning and anchor graph constructed on massive data;?2?Based on the theory of sparse norm,we propose a multi-view feature selection algorithm in semi-supervised scenarios with the joint use of l_2,1 norm and l_G1 norm.The experimental results show the proposed algorithm can improve the classification accuracy of multi-view feature selection in semi-supervised scenarios;?3?On the basis of anchor graph,we use anchor graph to construct graph Laplacian and constrain the projection matrix as feature selection matrix so as to reduce the time complexity.The experimental results demonstrate the proposed algorithm is effective to reduce time complexity.