How k-nearest neighbor parameters affect its performance

The k-Nearest Neighbor is one of the simplest Machine Learning algorithms. Besides its simplicity, k-Nearest Neighbor is a widely used technique, being successfully applied in a large number of domains. In k-Nearest Neighbor, a database is searched for the most similar elements to a given query element, with similarity defined by a distance function. In this work, we are most interested in the application of k-Nearest Neighbor as a classification algorithm, i.e., each database element has a label (class) associated, and the main goal of the algorithm is to decide the class of a new case based on the classes of the k most similar database elements. This work provides a discussion and presents empirical evidence of how the main parameters of k-Nearest Neighbor influence its performance. The parameters investigated are the number of nearest neighbors, distance function and weighting function. The most popular parameters choices were evaluated, including nine values for k, three popular distance measures and three well-known weighting functions. Our experiments were performed over thirty-one benchmark and “real-world” data sets. We recommend the use of the inverse weighting function and k = 5 for HEOM and HMOM distance functions and k = 11 to HVDM distance function.