Visualization of individual ensemble classifier contributions

Abstract

Ensembles of classifiers are usually considered a valuable approach in different scenarios. A broad range of methods to deal with the construction, diversity and combination of multiple predictive models have been extensively studied. While the focus is often to obtain more accurate and robust predictions than single models seldom the individual contribution of classifiers which could contribute to a better understanding of the uncertainty associated with ensembles’ outputs is taken into account. In this work we look into this issue and focus on evaluating the individual ensemble classifier contributions using several scenarios. We propose a visual web model that allows for the evaluation of both individual contributions as well as their interactions. We apply the proposed approach on a benchmark dataset and show how it can visually be used to better understand the uncertainty associated with the construction of ensembles, presenting some insight on the individual contributions and interactions.