Support Kernel Machine-Based Active Learning to Find Labels and a Proper Kernel Simultaneously

Abstract

SVM-based active learning has been successfully applied when a large number of unlabeled samples are available but getting their labels is costly. However, the kernel used in SVM should be fixed properly before the active learning process. If the pre-selected kernel is inadequate for the target data, the learned SVM has poor performance. So, new active learning methods are required which effectively find an adequate kernel for the target data as well as the labels of unknown samples.

In this paper, we propose a two-phased SKM-based active learning method and a sampling strategy for the purpose. By experiments, we show that the proposed SKM-based active learning method has quick response suited to interaction with human experts and can find an appropriate kernel among linear combinations of given multiple kernels. We also show that with the proposed sampling strategy, it converges earlier to the proper combination of kernels than with the popular sampling strategy MARGIN.