Reducing the computational demands for Nearest Centroid Neighborhood classifiers

The k Nearest Centroid Neighbor (k-NCN) is a relatively new powerful decision rule based on the concept of so-called surrounding neighborhood. Its main drawback is however slow classification, with complexity O(nk) per sample. In this work, we try to alleviate this disadvantage of k-NCN by limiting the set of the candidates for NCN neighbors for a given sample. It is based on an intuition that in most cases the NCN neighbors are located relatively close to the given sample. During the learning phase we estimate the fraction of the training set which should be examined only to approximate the "real" k-NCN rule. Similar modifications are applied also to ensemble of NCN classifiers, called voting k-NCN. Experimental results indicate that the accuracy of the original k-NCN and voting k-NCN may be preserved while the classification costs significantly reduced.