An AutoML framework that evolves ML pipelines as graphs

What it does

FEDOT is an open-source AutoML framework that automatically assembles machine learning pipelines for classification, regression, clustering, and time series forecasting. It wraps common libraries like Scikit-learn, CatBoost, and XGBoost into a unified system where preprocessing and model blocks are composed automatically. The framework handles tabular data, text, images, and multi-modal inputs under a single evolutionary optimizer.

The interesting bit

Rather than flattening AutoML into a hyperparameter grid search, FEDOT represents pipelines as graphs—nodes are operations, edges are data flows—and evolves the topology itself using genetic programming. This structural-learning approach is flexible enough that the documentation claims applicability to ODE and PDE problems, not just standard ML tasks.

Key highlights

• Graph-based pipelines: data preprocessing and model blocks connect as a directed graph, not a rigid linear stack.
• Evolutionary core: genetic programming optimizes both the pipeline structure and its hyperparameters.
• Multi-modal support: text, images, and tabular data can coexist in the same automated pipeline.
• Pluggable backends: integrates with Scikit-learn, CatBoost, and XGBoost, with hooks for custom models.
• Reproducible exports: final pipelines serialize as JSON or bundled ZIP archives for experiment replay.

Caveats

• The README offers no head-to-head benchmarks against mainstream AutoML frameworks, so relative performance is unclear.
• Tutorials and video content are largely in Russian, which may slow onboarding for English-only users.
• The roadmap section was truncated in the provided source, leaving near-term priorities unspecified.

Verdict

Data scientists who want inspectable, graph-shaped pipelines for messy, multi-modal problems will find FEDOT worth a spin; those seeking a battle-tested, benchmark-heavy AutoML commodity should probably shop around.