Generating custom classification datasets by targeting the instance space
Mario Andrés Muñoz, Kate Smith‐Miles · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2017
While machine learning has evolved at a fast pace in the last decades, the testing procedure of new methods may be not keeping pace. It often relies on well-studied collections of classification datasets such as the UCI repository. However, a meta-analysis through features has showed that most datasets from UCI are not sufficiently challenging to expose unique weaknesses of algorithms. In this paper we present a method to generate datasets with continuous, binary and categorical attributes, through the fitting of a Gaussian Mixture Model and a set of generalized Bernoulli distributions. By targeting empty areas of the instance space, this method has the potential to generate datasets with more diverse feature values.