Algorithms for Classification based on k-NN
Manuel Laguía, Juan Luis Castro · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2007
In this paper we focus on methods that solve classification tasks based on distances, and we introduce some variants of the basic k-NN method adding up to 3 characteristics. The experiments reveal a relationship between the accuracy of 1-NN (distances) and the accuracy of the methods based on those distances. We propose a heuristics according to this observation and test its correctness. We study the usefulness of the proposed methods epsilon-ball, epsilon-ball^{k-NN} and epsilon-ball^{1-NN}, and make an exhaustive comparison using six different distance functions and 68 data sets, including UCI--Repository and artificial data sets. The proposed methods are useful and significantly outperform k-NN frequently. We have also found some evidence about the weakness of k-NN when the optimal value of $k$ varies in different regions along the space.