Unsupervised Kernel Learning for Correlation Based Clustering

Successful clustering of multiple objects using kernels, heavily relies on the proper selection of kernel parameters. This can be a computationally complex process and may necessitate prior knowledge of label information. In this paper, a novel method has been introduced that is computationally efficient and requires no prior information. The method relies on the eigenvalues of each kernel matrix to determine a proper linear combination of kernels among a dictionary of kernels that results in good clustering. A difference of convex functions formulation is proposed and solved via an algorithmically simple method which is extremely cost-effective in implementation. Comparisons using various forms of real-world data with two popular supervised methods show its superior performance.