An analysis of order dependence in k-NN

Abstract

In classification based on k-NN with majority voting, the class assigned to a given problem is the one that occurs most frequently in the k most similar cases (or instances) in the dataset. However, different versions of k-NN may use different strategies to select the cases on which the solution is based when there are ties for the kth most similar case. One strategy is to break ties for the kth most similar case based on the ordering of cases in the dataset. We present an analysis of the order dependence introduced by this strategy and its effects on the algorithm's performance. © 2010 Springer-Verlag.