AutoML that breeds scikit-learn pipelines by evolution

What it does

TPOT is a Python AutoML tool that treats scikit-learn pipelines as evolving organisms. You give it a dataset and a fitness metric; it explores candidate pipelines—handling preprocessing, feature selection, model choice, and hyperparameters—and keeps the winners. The goal is to remove the manual grunt work of pipeline tuning while staying inside the scikit-learn ecosystem.

The interesting bit

Rather than flattening pipeline search into a giant grid, TPOT models pipelines as graphs and mutates their structure directly. The recent ground-up rewrite (previously dubbed TPOT2) introduced multi-objective optimization and a modular evolutionary framework, meaning you can customize the genetic algorithm itself, not just the operators it searches over.

Key highlights

• Evolves complete scikit-learn pipelines end-to-end, including preprocessing and feature selection.
• Graph-based representation allows structural mutations—swapping nodes, not just tweaking knobs.
• Modular framework lets you customize the evolutionary algorithm’s behavior.
• Parallel execution via Dask, though the README insists you wrap script logic in if __name__ == "__main__" to avoid multiprocessing chaos.
• Optional preprocessing = True will impute missing values and encode categoricals automatically.

Caveats

• TPOT does not validate that your data matches your operators; pass a dataframe with categoricals and you must either enable preprocessing or supply compatible steps.
• The built-in preprocessing is currently fit on the entire training set before cross-validation splits, which can leak information; per-fold preprocessing is noted as future work.
• sklearnex acceleration extensions are explicitly flagged as untested or unstable on Arm-based CPUs such as Apple Silicon.

Verdict

Best for practitioners who want to automate scikit-learn pipeline discovery and can spare the CPU cycles for evolutionary search. Look elsewhere if you need deterministic tuning or tight control over every preprocessing step.