Joint Graph Embedding And Label Propagation For Semi-supervised Subspace Clustering

As we slip further into the Internet era,the traditional social mode and consumption pattern have been influenced imperceptibly by the wave of the Internet,all walks of life are turning to provide people with convenient and high-quality services through massive data interactions,for the data clustering analysis can accelerate the upgrading of industrial services and bring more convenience to people's lives in all aspects.In the face of a large amount of data,unsupervised clustering algorithms do not use prior information,and the performance of the model often fails to meet our requirements.Semi-supervised clustering algorithms can use a small amount of prior information to improve the performance of the model while increasing the model's generality of the data.However,these data are often high-dimensional,and contain many irrelevant attributes.The clustering algorithm based on the original feature space is no longer applicable now.Researchers found that high-dimensional data generally has a certain low-dimensional subspace structure,so semi-supervised subspace clustering algorithms came into being.When facing problems about a large number of high-dimensional data clustering,the semi-supervised subspace clustering algorithm has got increasing attention due to its excellent performance.This thesis takes semi-supervised subspace clustering as the research direction,and proposes two innovative semi-supervised subspace clustering algorithms.The main contents of the thesis are detailed as follows:(1)The learning process of the traditional semi-supervised subspace clustering algorithm is divided into two stages: learning similarity matrix and subspace clustering,which cannot make full use of the relationship between similarity matrix and data labels.In addition,some semi-supervised subspace clustering algorithms based on low-rank representation can aptures the global mixture of subspaces structure,but they will ignore the local information of the data.In response to these problems,this thesis proposes a semi-supervised subspace clustering algorithm based on the combination of MFA(Marginal Fisher Analysis)and label propagation.The algorithm adds low-rank representation,label propagation,and graph embedding framework to the unified framework for optimization,unifying the two stages of semi-supervised clustering algorithms,and making the similarity matrix and the label matrix guide each other in learning.In addition,the Marginal Fisher Analysis Algorithm(MFA)in the Graph Embedding Framework is used to constrain the labeled data,so that it has the local structure characteristics of intra-class compactness and inter-class separability,so that the algorithm model can simultaneously learn global and local information of the data.This thesis proposes an effective optimization algorithm for this algorithm.Experiments show that this algorithm has certain advantages in clustering and is robust to the included noise.(2)In semi-supervised subspace clustering algorithm based on the combination of MFA and label propagation.The MFA algorithm only uses the nearest neighbor relationship to constrain the labeled data,and can not make all data have the local structure characteristics of intra-class compactness and inter-class separability.In terms of the above problems,this thesis proposes a semi-supervised subspace clustering algorithm based on the combination of adaptive graph embedding and label propagation.The algorithm uses the label matrix and the similarity matrix of all data to jointly guide the construction of the graph embedding framework,so that they can establish a close relationship,perform adaptive learning,and appropriately apply this local constraint to all data.This thesis proposes an effective optimized algorithm for this algorithm and it would prove that the algorithm has a better clustering performance and is robust to the included noise.