Sampling with Confidence: Using k-NN Confidence Measures in Active Learning

Active learning is a process through which classifiers can be built from collections of unlabelled examples through the cooperation of a human oracle who can label a small number of examples selected as most informative. Typically the most informative examples are selected through uncertainty sampling based on classification scores. However, previous work has shown that, contrary to expectations, there is not a direct relationship between classification scores and classification confidence. Fortunately, there exists a collection of particularly effective techniques for building measures of classification confidence from the similarity information generated by k-NN classifiers. This paper investigates using these confidence measures in a new active learning sampling selection strategy, and shows how the performance of this strategy is better than one based on uncertainty sampling using classification scores.