An Active Learning Method Based on Most Possible Misclassification Sampling Using Committee

Jun Long, Jianping Yin, En Zhu

Abstract: By selecting and asking the user to label only the most informative instances, active learners can significantly reduce the number of labeled training instances to learn a classification function. We focus here on how to select the most informative instances for labeling. In this paper we make three contributions. First, in contrast to the leading sampling strategy of halving the volume of version space, we present the sampling strategy of reducing the volume of version space by more than half with the assumption of target function being chosen from nonuniform distribution over version space. Second, via Halving model, we propose the idea of sampling the instances that would be most possibly misclassified. Third, we present a sampling method named CBMPMS (Committee Based Most Possible Misclassification Sampling) which samples the instances that have the largest probability to be misclassified by the current classifier. Comparing the proposed CBMPMS method with the existing active learning methods, when the classifiers achieve the same accuracy, the former method will sample fewer times than the latter ones. The experiments show that the proposed method outperforms the traditional sampling methods on most selected datasets.