Improvements To The Subspace Clustering Algorithm

Subspace clustering is a kind of clustering method that transforms high-dimensional space into low-dimensional subspace under the assumption that the data can be linear repressed by the other data in the same subspace as that data.Using the linear self-representation of data,the method clusters the data that can represent each other linearly into the same category.The method assumes that the data in different clusters lying on the different subspace.Subspace clustering method can solve the problem of high-dimensional clustering efficiently and it has been widely used on face recognition,image processing and trajectory recognition.But for the traditional subspace clustering algorithms,the clustering accuracy is not high enough;the applicability of the method is too narrow due to the strong assumptions about data structures;the clustering results on nonlinear high-dimensional are poor.(1)In order to improve the clustering accuracy,we proposed a new sparse subspace clustering algorithm.Firstly,we constructing the subspace optimization goal,which is more efficient and better robustness.Secondly,we adjust the coefficient matrix differently with other subspace clustering algorithm based on the property of the matrix to satisfy the definition of the affinity matrix in spectral clustering algorithm.The new matrix can strengthen the contact of the data in the same cluster.Finally,we use spectral clustering algorithm to get the clustering results.The experimental results show that the accuracy of the clustering results obtained by the new method is higher than that of the original subspace clustering algorithm,and the difference is more obvious when the data noise is large.(2)In order to solve the problem that the assumption of subspace clustering algorithm is too strong on data distribution,a new subspace clustering algorithm based on KNN method is proposed.By combining traditional algorithm KNN to the subspace clustering algorithm,the new subspace algorithm can get good clustering results on the nonlinear highdimensional dataset.Firstly,we constructed an affinity matrix based on the theory of KNN.Then,the matrix is integrated into the subspace clustering algorithm as a constraint term.In this way,the subspace clustering algorithm is inclined to choose the coefficients which contain more K-nearest neighborhood of the data.Finally,the spectral clustering algorithm is used.The experimental results show that the new subspace clustering algorithm based on KNN can cluster well both on the linear high-dimensional dataset and nonlinear highdimensional ones.And the robustness is better than other subspace clustering algorithms.