Cluster Ensemble Kernel for Kernel-based Classification

This paper presents a method for some semi-supervised and supervised classification problems based on properties of the averaged co-Association matrix obtained with a cluster ensemble. The ensemble clustering is performed as a preliminary step of data processing. The main property states that the matrix is a valid kernel matrix, thus it can be used in different classification methods that use kernels such as Kernel Nearest Neighbor, SVM, Kernel Fisher Discriminant. Some properties of the suggested method connected with its convergence to optimal classifier are studied. Numerical experiments show that the accuracy of the proposed algorithms is often higher than other state-of-The-Art methods, especially under the presence of complex data structures and noise distortions.