Semi-supervised Co-ensembling for AutoML

Automated machine learning (AutoML) is an increasingly popular approach for selecting learning algorithms and for configuring their hyperparameters in an effective, principled and fully automated way. So far, AutoML techniques have been focussed primarily on supervised learning scenarios. In this work, we extend AutoML to semi-supervised learning. Specifically, we propose co-ensembling, a generic procedure that uses unlabelled data to enhance the performance of high-quality ensembles of learners obtained from a state-of-the-art supervised AutoML system. Co-ensembling can be applied to any set of two or more base learners, and bypasses direct exploration of the large design space of semi-supervised learning algorithms. We demonstrate substantial performance improvements on multiclass classification problems by applying a single iteration of co-ensembling. On a large, diverse suite of multiclass data sets, single-step co-ensembling yields improved performance on 75% of the multiclass problems we considered, and achieves an average relative reduction in error rate of 7%.