earticle

영어

Spectral clustering is a clustering method based on algebraic graph theory. It has solid theoretical foundation and good performance of clustering. However, during the process of nonlinear low rank approximation, the traditional spectral clustering algorithm can’t effectively remove redundant features leading to the phenomenon that the local area can not distinguish. It also suffers from the high computational complexity of eigen-decomposition when dealing with the high dimensional data. In order to resolve the aforementioned problems, in this paper a novel Spectral clustering algorithm called LF-SC is proposed. Firstly, based on the nonlinear low dimensional embedding feature selection, we realize dimension reduction. The multi clustering structure of the data is captured, the potential manifold structure is fully discovered, and the geometry structure of the low dimensional manifold clustering is well maintained. Secondly, utilizing the SVD instead of EVD to obtain the eigenvectors reduces the computational complexity and maintain the local structure of the data as well as low dimensional manifold. Extensive experiments show the effectiveness and efficiency of our approach.