Spectral Clustering of High-Dimensional Data via k-Nearest Neighbor Based Sparse Representation Coefficients

Recently, subspace clustering has achieved promising clustering quality by performing spectral clustering over an affinity graph. It is a key to construct a robust affinity matrix in graph-oriented subspace clustering. Sparse representation can represent each object as a sparse linear combination of other objects and has been used to cluster high-dimensional data. However, all the coefficients are trusted blindly to construct the affinity matrix which may suffer from noise and decrease the clustering performance. We propose to construct the affinity matrix via k-nearest neighbor (KNN) based sparse representation coefficient vectors for clustering high-dimensional data. For each data object, the sparse representation coefficient vector is computed by sparse representation theory and KNN algorithm is used to find the k nearest neighbors. Instead of using all the coefficients to construct the affinity matrix directly, we update each coefficient vector by remaining the k coefficients of the k neighbors unchanged and set the other coefficients to zero. Experiments on six gene expression profiling (GEP) datasets prove that the proposed algorithm can construct better affinity matrices and result in higher performance for clustering high-dimensional data.