Study Of Multi-view Dimensionality Reduction Algorithms Based On Graph Embedding

Faced with the problem of information overload and dimensionality increase caused by massive data,feature dimensionality reduction in the fields of computer vision and data mining has become increasingly important and is now a research hotspot in the current academic community.The purpose of dimension reduction is to remove a large amount of redundant information from the data and to find useful essential features.However,in practice,the features from a single view do not fully reflect the information of the original samples.On the contrary,fusion of multi-view features can reflect the different attributes and complementary information of a particular sample.Based on the graph embedding framework,this thesis deals with the problem of dimensionality reduction from both the single-view and multi-view settings.The research findings obtained are as follows:Firstly,for single-view setting,we study the classical subspace projection algorithms from a unified graph embedding framework,and point out their two major drawbacks.To be specific,the gap between the ideal embedded subspace and the actual embedded subspace is not considered in the traditional graph embedding methods.Besides,the intrinsic and penalty graphs are constructed in the high-dimensional space.Therefore,we first constructed a new graph embedding framework with structure-preserved characteristics.This proposed Actual Graph Embedding(AGE)method achieves comparable classification performance to the traditional algorithms;Furthermore,to design the ideal subspace,we introduce label information to construct the label graph.The errors between the actual embedded subspace and the ideal embedded subspace are modeled by ridge regression,thus the closed-form solutions is available.The proposed Actual-Ideal Graph Embedding(AIGE)method can make the actual projected subspace to approach the ideal subspace to the utmost.The experimental results on the standard image databases show that the proposed algorithms not only enjoy the computational efficiency as conventional graph embedding based methods,but also have better recognition performance.Secondly,as the existing multi-view methods fail to exploit local discriminant information between different classes in the same view,we propose a multi-view learning method(MGRDA)based on graph regularization.It considers intra-class information,local inter-class information,and global discriminant information by constructing three connection graphs,ie,intrinsic graph,local penalty graph,and global penalty graph,respectively.These relationships of samples between each pair of view are preserved through discriminative canonical correlation analysis(DCCA).Furthermore,we also present the kernel expansion of MGRDA algorithm(KMGRDA).Experiments on multi-feature data of images and texts show that both MGRDA and KMGRDA effectively integrate complementary information from different pair of views,and have better recognition performance than several classic methods for multi-view dimensionality reduction.Finally,we apply the proposed single-view method(AIGE)and the multi-view method(MGRDA)to the reduction and fusion of radar emitter features,respectively.Experimental results based on real-measured data show that the proposed algorithms can significantly improve the discriminative ability of radar emitter features,by means of single-view dimension reduction and multi-view information fusion.