Active Kernel Learning

Identifying the appropriate kernel function/matrix for a given dataset is essential to all kernel-based learning techniques. In the past, a number of kernel learning algorithms have been proposed to learn kernel functions or matrices from side information, in the form of labeled examples or pairwise constraints. However, most previous studies are limited to the "passive" kernel learning in which the side information is provided beforehand. In this paper we present a framework of "Active Kernel Learning" (AKL) that is able to actively identify the most informative pairwise constraints for kernel learning. The key challenge of active kernel learning is how to measure the informativeness of each example pair given its class label is unknown. To this end, we propose a min-max approach for active kernel learning that selects the example pairs that will lead to the largest classification margin even when the class assignments to the selected pairs are incorrect. We furthermore approximate the related optimization problem into a convex programming problem. We evaluate the effectiveness of the proposed active kernel learning algorithm by comparing it with two other implementations of active kernel learning. Empirical study with nine datasets on data clustering shows that the proposed algorithm is considerably more effective than its competitors.