Selective Sampling for Nearest Neighbor Classifiers

In the passive, traditional, approachto learning, the information available to the learner is a set of classified examples, whichare randomlydrawnfrom the instance space. In many applications, however, the initial classification of the training set is a costly process, andan intelligently selection of training examplesfromunlabeled data is doneby an active learner. This paper proposesa loolmheadalgorithm for example selection and addresses the problemof active learning in the context of nearest neighborclassifiers. Theproposedapproachrelies on using a random field modelfor the examplelabeling, whichimplies a dynamicchange of the label estimates during the samplingprocess. The proposedselective samplingalgorithm wasevaluated empirically on artificial andreal data sets. The experiments showthat the proposed method outperforms other methodsin most cases.

Authors

• Michael Lindenbaum
• Shaul Markovich
• Dmitry Rusakov
• Technion - Israel Institute of Technology

Keywords

No keywords are indexed for this paper.