Auto-sklearn
If you already know your way around scikit-learn, the fastest way to see what AutoML actually does is to let Auto-sklearn take over the part of the workflow where you were guessing. You hand it a table and a time budget; it hands back a fitted pipeline.
How it works
Auto-sklearn, from the AutoML group at the University of Freiburg, searches the combined space of scikit-learn preprocessors, estimators and their hyperparameters — the CASH problem — using SMAC, a Bayesian optimiser built on random forests. Two ideas lift it above random search:
- Meta-learning warm start. It computes meta-features of your dataset, finds the most similar datasets in an archive of past experiments, and begins from the configurations that won there.
- Ensemble selection. Every model trained during the search is kept, and the final answer is a greedy weighted ensemble assembled from that library — so the compute spent on runners-up is not thrown away.
Using it
from autosklearn.classification import AutoSklearnClassifier
clf = AutoSklearnClassifier(time_left_for_this_task=3600)
clf.fit(X_train, y_train)The budget is the main dial you turn. Auto-sklearn 2.0 adds precomputed portfolios, successive halving to kill weak configurations early, and an automatic choice of validation strategy, which makes short runs far more competitive. Expect tabular data: this is the reference research system for structured problems, not a tool for images or raw text.
Auto-sklearn is an open-source AutoML library that automates model and hyperparameter choice for scikit-learn pipelines. It searches the combined space of preprocessors, estimators and their settings with the Bayesian optimiser SMAC, warm-starts from meta-features of similar past datasets, and finishes by building a greedy weighted ensemble from every model it trained. It matters as the reference research system for tabular AutoML and as a drop-in replacement for manual tuning.
Suppose you have a spreadsheet and a question — which customers are about to cancel, say — and you've heard that a computer can learn the answer from the data. The hard part isn't the learning. It's the hundreds of small decisions: which method to use, which settings, how to clean up the columns.
Auto-sklearn is a free tool that makes those decisions for you. You give it your table and tell it how long it is allowed to think — ten minutes, an hour — and it tries out many combinations, leans on what worked for similar data in the past, and hands back the best model it managed to find. You get a working predictor without having to be an expert in the options.