Kernel-based clustering algorithms for spectral pattern recognition

Unsupervised clustering algorithms are an important part of spectral pattern recognition. These spectral pattern clusters may be difficult to be formed by traditional unsupervised clustering algorithms due to nonlinear boundaries lie in between patterns in the original pattern space. The kernel method can extend unsupervised clustering algorithms to a kernel space which will enable the discovery of clusters. We apply the kernel method to the Generalized Possibilistic Clustering Algorithms (GPCA) which was recently developed. The proposed algorithm will be called kernel-based GPCA (KGPCA). Three well known functions including the Gaussian, exponential and uniform were used as kernel functions. The proposed algorithm was tested on some pattern classes including his and Glass to show the effectiveness of the proposed algorithm.