Automated machine learning with dynamic ensemble selection

Automated machine learning (AutoML) has been developed for automatically building effective machine learning pipelines. However, existing AutoML frameworks use a single individual pipeline or a weighted ensemble of several pipelines to create the final predictive model, which ignores the difference between the unseen instances and leads to undesirable performance. To construct customized models for different unseen instances, we propose a novel AutoML method based on dynamic ensemble selection of machine learning pipelines, where the most competent combination of base pipelines is selected and aggregated to predict a specific unseen instance. First, an effective base pipeline pool is generated by filtering out the underperforming pipelines. Second, when an unseen instance appears, we deploy a new dynamic balanced accuracy criterion to select the most competent ensemble of base pipelines according to its local region. Finally, the outputs of the selected pipelines are integrated to give the final prediction. Comprehensive experiments on 39 publicly available datasets demonstrate the superiority of the proposed method compared to some state-of-the-art AutoML frameworks.