k-Nearest Neighbor Classification Using Dissimilarity Increments

Abstract

In this paper we propose a classification method that generalizes the k-nearest neighbor (k-NN) rule in a maximum a posteriori (MAP) approach, using an additional characterization of the datasets. That characterization consists of a high order dissimilarity called dissimilarity increment; this dissimilarity measure uses information from three points at a time, unlike typical distances which are pairwise measures. In practice, in this model, the likelihood of a point not only depends of its direct k neighbors, but also of the nearest neighbor of each one of its k neighbors. Experimental results show that the proposed classifier outperforms more traditional and simple classifiers like Naive Bayes and k-nearest neighbor classifiers. This improved performance is especially noticeable relative to k-NN when k is poorly chosen.