Research And Application About Algorithm Of Subspace Clustering Based On Data Representation

In the field of machine learning and pattern recognition,clustering as an unsupervised learning method,has been widely used in data analysis and processing.Nowadays,as we enter the era of big data,high-dimensional data exists in all kinds of fields.The traditional clustering algorithm measure the similarity between data based on distance.However,in large-scale high-dimensional data,it exits the phenomenon that the distance between samples and clusters is almost equal,which makes the traditional distance-based clustering algorithm invalid.In recent years,the representation-based subspace clustering algorithm has been widely used because of its excellent clustering performance in high-dimensional large data.In this paper,the subspace clustering algorithm is also deeply analyzed and discussed,and the related improvement methods are proposed to improve the performance of subspace clustering algorithm.The main work of this paper is as follows:1?Different from sparse representation subspace clustering using sparse technology and low rank representation subspace clustering using low rank technology,this paper proposes a cooperative representation of which the subspace representation coefficient matrix using Frobenius norm.Frobenius norm has grouping effect: the greater the correlation between data,the closer and the larger the coefficients corresponding to them in the coefficient matrix and vice versa.When the data belong to the same cluster,the correlation between them is relatively high,and the corresponding value of the representation coefficient is larger.When the data belong to different categories,the correlation between them is relatively low and the value is small.Therefore,the cooperative representation proposed in this paper has the effect of making the corresponding data representation coefficients of the same cluster very large and the representation coefficients between clusters very small.This effect satisfies the sparsity requirement of representation coefficient matrix in subspace clustering,and further improves the quality of sparsity and reduce the computational complexity of the model.2? There are two separate processes for constructing representation coefficient matrices and graph segmentation in the subspace clustering algorithm based on representation.These separate links cause that the algorithm can't get the complete internal relationship between coefficient matrix construction and graph segmentation.The quality of graph segmentation depends entirely on the quality of the representation coefficient matrix.The segmentation results can't be fed back to the process of constructing the coefficient matrix,which can't form a closed-loop feedback.In this paper,spectral enhancement is proposed.Regularize graph segmentation and then introduce it into the construction of representation coefficient matrix.The indicator matrix representing the segmentation results of the graph promote the block diagonal sparsity of the coefficient matrix.Image clustering experiments and motion segmentation experiments prove the effectiveness of the algorithm.3?The existing subspace clustering algorithms consider more about the global structures of the dataset and easily ignore the potential local manifold structures of high-dimensional dataset,which are very important for clustering.In this paper,the manifold structures are approximately equivalent to their local tangent space structure.Firstly,the base matrix of tangent space is obtained,and then reconstructed linearly to approximate the tangent space.Then the sparse technology is used for feature selection,and the principal eigenvectors of the base matrix are selected to obtain the sparse reconstruction coefficient matrix.Finally,the sparse reconstruction coefficient matrix is used to construct the nearest neighbor graph.The similarity relation of neighborhood graph is used to guide the construction of coefficient matrix and semi-supervised learning.The priori label information of the data is transferred on the nearest neighbor graph through regularization constraints,and complete semi-supervised learning.The results of face recognition experiments show that this model greatly improves the performance of subspace clustering.