Neighborhood-Based Local Sensitivity

  • Paul N. Bennett

Proceedings of ECML 2007 |

Published by Springer Verlag

We introduce a nonparametric model for sensitivity estimation which relies on generating points similar to the prediction point using its k nearest neighbors. Unlike most previous work, the sampled points differ simultaneously in multiple dimensions from the prediction point in a manner dependent on the local density. Our approach is based on an intuitive idea of locality which uses the Voronoi cell around the prediction point, i.e. all points whose nearest neighbor is the prediction point. We demonstrate how an implicit density over this neighborhood can be used in order to compute relative estimates of the local sensitivity. The resulting estimates demonstrate improved performance when used in classifier combination and classifier recalibration as well as being potentially useful in active learning and a variety of other problems.