{36 () Selective Sampling for Nearest Neighbor Classiiers *

Most existing inductive learning algorithms assume the availability of a training set of labeled examples. In many domains, however, labeling the examples is a costly process that requires either intensive computation or manual labor. In such cases, it may be beneecial for the learner to be active by intelligent selection of examples for labeling with the goal of reducing the labeling cost. In this paper we propose a lookahead algorithm for selective sampling of examples for nearest-neighbors classiiers. The algorithm attempts to nd the example with the highest utility considering its eeect on the resulting classiier. Computing the expected utility of an example requires estimating the probability of the possible labels. We propose to use the random eld model for this estimation. The proposed selective sampling algorithm was evaluated empirically on real and artiicial data sets. The experiments show that the proposed algorithm outperforms other methods.