A subscription business with a spreadsheet of past customers, and a column saying who cancelled, could give that table to an AutoML tool. The tool tries several model families, tunes each one’s settings, scores every candidate on data it did not train on and returns a ranked list, sometimes with a combined model at the top. Work that could take an analyst days of trial and error becomes one search with a time budget.
Under the hood, AutoML is mostly hyperparameter optimization widened to include the choice of model and preprocessing steps. The auto-sklearn project, for example, combines Bayesian optimization, lessons from earlier datasets (meta-learning) and ensemble methods. It is a different job from autotuning, which picks fast implementations or runtime settings for a workload that is already defined.
AutoML automates the search, not the judgment around it. People still decide what to predict, which data is allowed, whether information from the future has leaked into the training table and how the model will be monitored once deployed. A large search can also overfit the validation data it is scored on, so keep a separate test set that the search never sees.
Sources
- auto-sklearn documentation — Describes auto-sklearn as freeing a user from algorithm selection and hyperparameter tuning, using Bayesian optimization, meta-learning and ensemble construction; checked October 24, 2026.
- Hutter, Kotthoff and Vanschoren (eds.): Automated Machine Learning: Methods, Systems, Challenges — Open-access book whose chapters cover hyperparameter optimization, meta-learning, neural architecture search and AutoML systems such as Auto-sklearn and TPOT; checked October 24, 2026.