Data characteristics that determine classifier performance

Abstract

We study the relationship between the distribution of data, on the one hand, and classifier performance, on the other, for non-parametric classifiers. It is shown that predictable factors such as the available amount of training data (relative to the dimensionality of the feature space), the spatial variability of the effective average distance between data samples, and the type and amount of noise in the data set influence such classifiers to a significant degree. The methods developed here can be used to gain a detailed understanding of classifier design and selection.