Subset selection classifier (SSC): A training set reduction method

Instance-based learning algorithms are often required to choose which instances to store for use during classification. Keeping too many instances usually results in more storage and processing time requirements during classification. Many attempts have been made to reduce the size of the training set. The major drawback of majority of these attempts is their expensive learning process that limits their application in practical domains. In this paper, we propose a new training set reduction algorithm called Subset Selection Classifier (SSC), which chooses a minimal subset by performing an incremental search in the training set. SSC extends the nearest neighbor concept by constructing several circular regions in the training sample and building a model by collecting the central instance of each circular region along its radius. A test instance is classified by the selected instances if it falls within the radius of any selected instance. Experimental evaluation against 12 existing techniques on 11 benchmark datasets show that SSC has the best accuracy as well as the best reduction of the size of the training set in the average case.