A systematic literature review on AutoML for multi-target learning tasks

Abstract

Automated machine learning (AutoML) aims to automate machine learning (ML) tasks, eliminating human intervention from the learning process as much as possible. However, most studies on AutoML are related to unique targets. This article aimed to identify and analyze studies on AutoML applied to multi-label classification and multi-target regression through a systematic literature review (SLR). Initially, we defined the research questions, the search string, the data sources for the search, and the inclusion and exclusion criteria. Then, we carried out the study selection process in four steps, with snowballing being the last stage. Altogether 12 studies were selected to compose SLR. All studies automated the task of ML model search of the pipeline, one study automated the task of feature engineering of the pipeline, all were related to Multi-label Classification, and only one addressed multi-target regression. The search space consisted of algorithms/neural operations and hyperparameters, the studies employed optimization algorithms (such as Genetic Algorithms and Hierarchical Task Networks) to produce increasingly better candidate solutions and one metric to assess the quality of candidate solutions. Only two studies employed Transfer Learning to contribute to AutoML. This article reviewed AutoML, multi-label classification, and multi-target regression and, by answering the SLR research questions, showed how current studies address these issues and gave insights into future directions for AutoML and multi-target tasks.