Learning the Kernel with Hyperkernels

摘要: This paper addresses the problem of choosing a kernel suitable for estimation with support vector machine, hence further automating machine learning. goal is achieved by defining reproducing Hilbert space on kernels itself. Such formulation leads to statistical similar minimizing regularized risk functional.We state equivalent representer theorem choice and present semidefinite programming resulting optimization problem. Several recipes constructing hyperkernels are provided, as well details common learning problems. Experimental results classification, regression novelty detection UCI data show feasibility our approach.