Auto-CVE: a coevolutionary approach to evolve ensembles in automated machine learning

ABSTRACTAutomated Machine Learning (Auto-ML) is a growing field receiving a lot of attention. Several techniques are being developed to address the question of how to automate the process of defining machine learning pipelines, using diverse types of approaches and with relative success, but still this problem is far from being solved. Ensembles are frequently employed in machine learning given their better performance, when compared to the use of a single model, and higher robustness. However, until now, not much attention has been given to them in the Auto-ML field. In this sense, this work presents Auto-CVE (Automated Coevolutionary Voting Ensemble) a new approach to Auto-ML. Based on a coevolutionary model, it uses two populations (one of ensembles and another for components) to actively search for voting ensembles. When compared to the popular algorithm TPOT, Auto-CVE shows competitive results in both accuracy and computing time.