Kernelized design of experiments

Abstract

This paper describes an approach for selecting instances in regression problems in the cases where observations x are readily available, but obtaining labels y is hard. Given a database of observations, an algorithm inspired by statistical design of experiments and kernel methods is presented that selects a set of k instances to be chosen in order to maximize the prediction performance of a support vector machine. It is shown that the algorithm significantly outperforms related approaches on a number of real-world datasets.