ballet
A lightweight framework for collaborative, open-source data science projects through feature engineering.
- Free software: MIT license
- Documentation: https://hdi-project.github.io/ballet
- Homepage: https://github.com/HDI-Project/ballet
Overview
Do you develop machine learning models? Do you work by yourself or on a team? Do you share notebooks or are you committing code to a shared repository? In contrast to successful, massively collaborative, open-source projects like the Linux kernel, the Rails framework, Firefox, GNU, or Tensorflow, most data science projects are developed by just a handful of people. But think if the open-source community could leverage its ingenuity and determination to collaboratively develop data science projects to predict the incidence of disease in a population, to predict whether vulnerable children will be evicted from their homes, or to predict whether learners will drop out of online courses.
Our vision is to make collaborative data science possible by making it more like open-source software development. Our approach is based on decomposing the data science process into modular patches - standalone units of contribution - that can then be intelligently combined, representing objects like "feature", "labeling function", or "prediction task definition". Collaborators work in parallel to write patches and submit them to a repo. Our software framework provides the underlying functionality to merge high-quality contributions, collect modules from the file system, and compose the accepted contributions into a single product. It also provides a familiar notebook-based development experience that is friendly to data scientists and other inexperienced open-source contributors. We don't require any computing infrastructure beyond that which is commonly used in open-source software development.
Currently, Ballet focuses on supporting collaboratively developing feature engineering pipelines, an important part of many data science projects. Individual features are represented as separate Python modules, declaring the subset of a dataframe that they operate on and a scikit-learn-style learned transformer that extracts feature values from the raw data. Ballet collects individual features and composes them into a feature engineering pipeline. At any point, a project built on Ballet can be installed for end-to-end feature engineering on new data instances for the same problem. How do we ensure the feature engineering pipeline is always useful? Ballet thoroughly validates proposed features for correctness and machine learning performance, using an extensive test suite and a novel streaming logical feature selection algorithm. Accepted features can be automatically merged by the ballet GitHub app into projects.
Next steps
Are you a data owner or project maintainer that wants to organize a collaboration?
Are you a data scientist or enthusiast that wants to join a collaboration?
Want to learn about how Ballet enables Better Feature Engineering
Want to see a demo collaboration in progress and maybe even participant yourself?
