Research On Sample And Feature Subspace Collaborative Ensemble Algorithm And Application

Sample learning and feature learning are two important aspects of machine learning.Although deep learning has excellent feature learning capabilities,it relies heavily on a large number of samples and is not suitable for small sample learning areas such as medical,military and online learning.Sample and feature subspace collaborative learning is a solution that has emerged in recent years.This idea learns from the perspectives of samples and features at the same time.It has the ability to combine the advantages of both sample learning and feature learning,and subspace learning can improve the small sample learning with the help of ensemble learning.However,there are still some problems with this method,including how sample and feature subspaces are sampled collaboratively,how to combine subspace learning with feature learning,how to combine the advantage of sample learning with feature extraction,and how to combine classifiers.Based on this,the thesis is for small sample applications,and specifically studies the sample and feature subspace collaborative ensemble algorithms,as well as the key issues,including sample and feature subspace sampling and collaborative learning,feature extraction,classifiers ensemble,and so on.The following work is done in this thesis:The thesis studies a collaborative ensemble learning method based on hierarchical subspace switching for hyperspectral image classification.Firstly,the spatial features of the image are extracted by fast bilateral filtering,and combined with the spectral features to form spectral-spatial joint features,and then the sample learning is performed by iterative mean clustering to form hierarchical subspaces.In each hierarchical subspace,a sample feature subset is formed by the random subspace sampling method,semisupervised sampling is performed on each subset,and then feature extraction is performed by micro noise linear dimension reduction.Then ensemble multiple kernels support vector machine is trained on each hierarchical subspace and voting strategy 1 is executed to get the subspaces ensemble result.Then,through voting strategy 2,the ensemble result of multiple hierarchical subspaces is obtained as the final ensemble model.By testing on datasets such as Indian Pines,the method shows better classification performance than most other recent methods.The thesis designs a collaborative learning method of sample subspace and correlation feature grouping.Firstly,sample feature subsets are formed by random sample subspace sampling and feature grouping.Then,the edge locality preserving discriminant projection method is used to transform each subset.After training multiple base classifiers,the ensemble of multiple classifiers is performed by L1 regularization sparse constraint weighted ensemble algorithm.This method outperforms comparison methods on multiple public datasets.This thesis also analyzes the parameters of edge locality preserving discriminant part of the method.In the thesis,two methods are designed to combine sample learning and feature learning to realize the collaborative ensemble of sample and feature subspaces and improve classification performance.