Abstract:
k-Nearest Neighbours (k-NN) is a well understood and widely-used approach to classification and regression problems. In many cases, such applications of k-NN employ the standard Euclidean distance metric for the determination of the set of nearest neighbours to a particular test data sample. This paper investigates the use of a data-driven evolutionary approach, named KTree, for the automatic construction of a kernel-based distance metric as an alternative to Euclidean distance. The key idea behind this approach is that a different distance metric is generated for a particular data domain. The performance of k-NN with the standard Euclidean distance measure is compared with that of k-NN based on a kernel-based distance metric evolved by KTree. This comparison is based on experiments on both synthetic and real-world datasets.
