Skip to content

Contributing

Thank you! 🎉

Thank you for your interest in mljar-supervised and for willing to contribute! 👍 😃

How can I contribute? 💪

Reporting Bugs

When creating a new bug report, please make sure to:

  • search for existing issues first. If you find a previous report of your issue, please update the existing issue with additional information instead of creating a new one.

  • please review the template and guidelines and include as many details as possible.

Suggest Enhancements

Please try first to see if an issue exists which describes your enhancement. If there is none, please create the new one with a detailed description. Thank you!

Pull Requests

  • Before any contribution, please reach out and discuss what do you want to implement and how (ideally in the created issue).
  • For code formatting we use Black.

To run all tests locally:

pytest tests

To run only tests for the selected file (for example, NN algorithm): 

pytest tests/tests_algorithms/test_nn.py -vsx

I add vsx flags to see what's going on.

Adding an algorithm

Algorithm implementations inherit from BaseAlgorithm in supervised/algorithms/algorithm.py. It uses Python's ABC and requires these methods, either implemented directly or inherited from an adapter:

Method Responsibility
fit Train using the training data and supported validation arguments.
predict Return task-appropriate predictions (positive-class probabilities for binary classification, class probabilities for multiclass classification, or numeric predictions for regression).
is_fitted Report whether the model is loaded and ready for prediction.
file_extension Return the saved model's extension without a leading dot.
copy Return an independent copy of the learner and its current model state.
save Persist the model to the supplied path.
load Restore the model from the supplied path.

Incomplete subclasses cannot be instantiated. Shared adapters such as SklearnAlgorithm can remain abstract until a concrete algorithm implements the remaining methods.

The base class provides metadata serialization, lazy loading, interpretation, and probability formatting. update is an optional no-op hook; get_metric_name returns None unless the algorithm supplies a metric name. Override these defaults only when the algorithm needs different behavior.

The abstract contract checks that methods are implemented; it does not validate their signatures or behavior. Add tests for training and prediction, independent copies, saving and loading, and prediction after lazy loading. Registered algorithms must also pass tests/tests_algorithms/test_algorithm_contract.py.

Documentation 📜

The project documentation source is in this repository under docs/. Deployed docs are available at https://supervised.mljar.com.

Contributor License Agreement 👔

Before merging any pull request with code updates, we'll need each person contributing to the code base to sign a Contributor License Agreement here.