Autosklearn

Auto-sklearn is an open-source automated machine learning (AutoML) system for tabular data. Built on top of scikit-learn, it aims to automate the process of selecting models, tuning hyperparameters, and choosing suitable data preprocessing steps, delivering a ready-to-use estimator with minimal manual intervention.

The system searches a large configuration space that includes classifiers or regressors, preprocessing steps, feature engineering, imputation, and normalization. It uses Bayesian optimization (via the SMAC framework) to efficiently explore configurations, while caching evaluations to avoid redundant work. Auto-sklearn also employs meta-learning to warm-start the search by leveraging experience from previous datasets, helping to choose promising configurations early. During the search, candidate pipelines are evaluated with cross-validation to estimate generalization performance. After a budget is exhausted, the best-performing pipelines are combined into an ensemble, typically using ensemble selection, to improve robustness and accuracy.

Users interact via Python code by instantiating AutoSklearnClassifier or AutoSklearnRegressor, providing the dataset, time and memory constraints, and letting the system run. It integrates with scikit-learn estimators and follows the same fit/predict API.

Limitations include computational requirements, particularly on large datasets, and a focus on tabular data. It may require careful preprocessing to avoid leakage and to handle categorical features, and results can be sensitive to the chosen time budget. The project is actively maintained by a community of researchers and developers.