SPECTRAL CLUSTERING WITH A NEW SIMILARITY MEASURE

Spectral clustering has been receiving more and more concerns in recent years. The performance of spectral clustering algorithm depends heavily on similarity measure. By analyzing the characteristics of spectral clustering and global features of clustering structure, we propose a new similarity measure method based on Gaussian kernel function. It is relatively insensitive to the nuclear parameter and can handle multi-scale clustering issues. Experiment in the synthetic data sets and USPS handwritten datasets demonstrates the proposed algorithm is superior to the traditional one.