An Empirical Re-Examination of Weighted Voting for k-NN
Jakub Zavrel, Walter Daelemans, P. Flach, A Bosch · Research portal (Tilburg University) · 1997
For some applications of k-nearest neighbor classifiers, the best results are obtained at a relatively large value of k. With the majority voting method, these results can be suboptimal. In this paper the performance of various weighted voting methods is tested on a number of machine learning datasets. The results show that weighted voting is often superior to majority voting, and that the linear weighting function proposed by Dudani [5] often yields slightly better results than the inverse distance function that has commonly been used in more recent work. 1 Introduction Classification algorithms from the family of k-nearest neighbors (k-NN) [6] [4] or Instance Based Learning [1] [13] [12] are based on the idea that similar instances of a problem tend to have similar solutions. The basic algorithm stores a set of classified cases, represented by feature-value vectors, in memory. When a new case -- the query -- is to be classified, the k vectors with the smallest distance to it are se...