Spectral clustering method for high dimensional data based on K-SVD

Abstract

Aimed at solving the problem that traditional clustering methods are vulnerable to the sparsity feature of the high dimensional data, a spectral clustering algorithm is proposed based on K-SVD dictionary learning. The algorithm firstly learns a dictionary by K-SVD and obtains sparse representation coefficients of all data samples in the dictionary by l1 sparse optimization. Then the similarity matrix between data samples is constructed through standardization and symmetrization of the solution to coefficients matrix. At last, we cluster the high dimensional data using spectral clustering algorithm with the similarity matrix as input. Empirical tests show that the algorithm proposed outperforms the spectral clustering algorithm based on sparse representation and traditional k-means in clustering accuracy, false alarm rate and detection rate.