Learning Distance Functions in k-Nearest Neighbors

Abstract

Normally the distance function used in classification in the k-Nearest Neighbors algorithm is the euclidean distance. This distance function is simple and has been shown to work on many different datasets. We propose a approach where we use multiple distance functions, one for each class, to classify the input data. To learn multiple distance functions we propose a new distance function with two learning algorithms. We show by experiments that the distance functions that we learn yields better classification accuracy than the euclidean distance, and that multiple distance functions can classify better than one.