Research On Feature Representation Learning Based On Graph Embedding

With the rapid development of information technology,a great deal of information comes into being.In order to efficiently find the information that people need,researchers have performed a lot of work on computer vision and machine learning,in which classification and clustering have attracted much attention.But in order to represent an object more comprehensively before clustering or classifying,we need to extract multi-view features of objects,which often are high-dimensional.However,the noise and redundant information in these data will have a great impact on the clustering or classifying task.Therefore,learning feature representation is one of the effective ways.In this paper,we did some research on feature representation learning based on graph embedding.Firstly problem: We proposed a robust affinity graph learning framework to deal with multi-view clustering problem.First,an Robust Hypergraph Laplacians Construction(RHLC)algorithm is used to select the significant features from the original feature space such that more robust graph Laplacians for each view are obtained.Second,we model hypergraph Laplacians as points on a Grassmann manifold and propose a Consistent Affinity Graph Learning(CAGL)algorithm to fuse these views.Experiments on five publicly available datasets demonstrate that our proposed method achieves convergence within a small number of iterations and obtains promising results compared with state-of-the-art methods.Second problem: We proposed a transductive Tensor-Driven Low-Rank Discriminant Analysis(TLRDA)model for image set classification,in which image sets represented by subspace objects,i.e.,points on a Grassmann manifold,are modeled as a higher-order tensor to model relations within vectors or matrices.By exploiting the advantages of the Low-Rank Representation(LRR)on noise or corrupted data,our approach seeks the lowest-rank representation of subspace embedded in tensorial data.We also show that the geometrical structure embedded in data can be preserved well by characterizing the compactness of intra-class and separability of inter-class under the framework of graph embedding.Experiments on six publicly available datasets demonstrate that our proposed method is guaranteed to converge.