FEDOT

FEDOT is an open-source AutoML framework for constructing data-driven composite models. It suits developers and researchers who want pipeline automation they can steer in code, especially across mixed data types and forecasting tasks. The trade-off is a self-hosted, code-first workflow and narrowly defined GPU model support.

Overview

FEDOT addresses binary and multiclass classification, regression, clustering, and univariate or multivariate time-series forecasting. It covers preprocessing, feature engineering, model selection, tuning, cross-validation, explainability, and serialization, making it a broad foundation for an end-to-end machine-learning workflow rather than a single modeling utility.

The project is maintained by the NSS Lab within ITMO University's National Center for Cognitive Technologies. It is released under the 3-Clause BSD license, including for projects and research, and welcomes bug reports and enhancement proposals through GitHub issues.

Key features

InputData can come from CSV files, pandas DataFrames, NumPy arrays, or time-series CSV data. FEDOT works with tabular, text, and image data, including multimodal combinations. Its preprocessing handles infinite and missing values, binary and non-binary categorical features, and stray spaces in categorical data, reducing common preparation work but not replacing domain-specific data checks.

Automation is configurable. Omitting parameters enables full pipeline generation; supplying parameters lets users constrain the search or compose parts manually. That makes FEDOT useful for establishing a baseline quickly and then imposing expert decisions, provided the user is comfortable working through an API. The default preset is auto; alternatives include best_quality, fast_train, stable, gpu, ts, and automl.

Graph-based pipeline optimization and learning use the open-source GOLEM library and meta-heuristic methods. Models are drawn mostly from scikit-learn, statsmodels, and Keras; the project also describes support for scikit-learn, CatBoost, and XGBoost, with custom libraries integrable for teams that need to extend the model set. Five-fold cross-validation is the default, and additional metrics can be supplied to the optimizer to address potential bias.

Installation is available with pip install fedot, with optional image, text-processing, and deep-neural-network dependencies through fedot[extra]. The API can also be invoked from a console without writing Python, with predictions saved as CSV files. GPU evaluation uses RAPIDS but is limited to Ridge, Lasso, LogisticRegression, RandomForestClassifier, RandomForestRegressor, KMeans, and SVC; that restriction matters if acceleration is central to your workload.

Pricing

PlanPriceIncludesFEDOTfreeOpen-source AutoML framework; BSD 3-Clause licenseFEDOT is free, with no trial period because the sole plan is open source. There are no published seat, usage, or renewal charges. You install and operate it yourself, so the value is strongest for users able to provide their own development and compute environment rather than those seeking a managed service or graphical workspace.

Platforms

FEDOT is available through an API and runs on Linux, macOS, and Windows. Its hosting model is self-hosted, matching the package-based installation and code workflow. This suits teams that need local control; it is less suitable when a web application or vendor-managed runtime is a requirement.

Who it's for

Choose FEDOT if you build models in Python-oriented environments and want to decide how much of pipeline search, preprocessing, and model selection to automate. Its task coverage and multimodal inputs fit varied research and applied ML projects, while presets and manual controls accommodate both quick baselines and more directed experiments.

Look elsewhere if nontechnical users need a visual interface, if you require a managed cloud service, or if your GPU workload depends on algorithms outside FEDOT's supported RAPIDS list.

Pros and cons

Pros

Broad support for classification, regression, clustering, and univariate or multivariate forecasting.

Handles tabular, text, image, and multimodal data with practical missing-value and categorical-data preprocessing.

Automation can range from hands-off generation to parameter-guided, manually composed pipelines.

Free BSD-licensed software with extensibility across established ML libraries and custom integrations.

Includes cross-validation, tuning, explainability, and serialization within one workflow.

Cons

The workflow is code-based and self-hosted, excluding readers who need a polished visual or managed experience.

GPU evaluation covers only seven named model types, limiting acceleration choices.

Using presets, metrics, and partial automation effectively still requires machine-learning and programming judgment.

Alternatives

AutoML Software is the broader directory starting point when you want to compare more tools. Pick BigML instead for a freemium API, web, or self-hosted service with a free tier capped at 16 MB per task, two parallel tasks, and one user. JADBio is the alternative when a web-accessible freemium product and defined project, upload, storage, and export allowances fit your team.

Choose Amazon SageMaker Autopilot for a paid, web/API service billed pay-as-you-go rather than a local open-source package. MLJAR AutoML, AutoKeras, LightAutoML, and AutoGluon are other free, self-hostable library options. Siemens Building X is a subscription platform for readers seeking a broader commercial environment.

Verdict

FEDOT is a strong choice for developers and researchers who want free, extensible AutoML with control over pipeline composition and support for varied data and forecasting problems. Choose it chiefly for that adjustable, code-driven breadth; look elsewhere for a visual interface, managed hosting, or unrestricted GPU algorithm coverage.