A greedy algorithm for optimizing the kernel alignment and the performance of kernel machines

Kernel-target alignment has recently been proposed as a criterion for measuring the degree of agreement between a reproducing kernel and a learning task. It makes possible to find a powerful kernel for a given classification problem without designing any classifier. In this paper, we present an alternating optimization strategy, based on a greedy algorithm for maximizing the alignment over linear combinations of kernels, and a gradient descent to adjust the free parameters of each kernel. Experimental results show an improvement in the classification performance of support vector machines, and a drastic reduction in the training time.