Mean Shift Spectral Clustering using Kernel Entropy Component Analysis

We present a promising clustering algorithm which combines mean shift (MS) clustering and spectral clustering (SC). A novel feature of the method is the use of two bandwidths, one for the mean shift algorithm in the first stage and another for the spectral clustering in the second. The first bandwidth should describe the local details, while the second captures the global structure of the dataset. Compared to traditional spectral clustering, our method may handle larger data sets, and the proposed MSSC procedure is shown to provide good clustering results in general when following some basic principles for selecting parameters.