New Methods for Spectral Clustering.

摘要: Analyzing the affinity matrix spectrum is an increasingly popular data clustering method. We propose three new algorithmic components which are appropriate for improving performance of spectral clustering. First, observing eigenvectors suggests to use a K-lines algorithm instead commonly applied K-means. Second, works best if has clear block structure, can be achieved by computing conductivity matrix. Third, many problems inhomogeneous or asymmetric in sense that some clusters concentrated while others dispersed. In this case, context-dependent calculation helps. This method also turns out allow robust automatic determination kernel radius σ.