scikit-learn

What it is

scikit-learn is a Python library dedicated to statistical learning ^[2]. In programming, a library is a body of ready-made code that other programs call instead of reimplementing every step from scratch; statistical learning is the practice of fitting a model to data so that the model can classify new examples or estimate values for them. scikit-learn is open source and released under the BSD license ^[2] — a permissive license that allows the code to be used, modified and redistributed. Inria, the French research institute whose project team became involved in 2009, describes the library as one of the top three most popular machine learning software programs on GitHub ^[2].

The library works on complex data, including databases, texts and images ^[2]. Inria lists fraud and spam detection, medical image analysis and industrial process optimization among the applications it serves ^[2]. Fraud detection has to pick out unusual transactions from a stream of ordinary ones, medical image analysis has to find the parts of an image that deserve attention, and industrial process optimization uses production records to relate settings to outcomes. The data in those cases look very different from one another, but the workflow is the same: prepare the data, choose a model, fit it, and check how well it does.

Origins

By the project's own account, scikit-learn began in 2007 as a Google Summer of Code project by David Cournapeau ^[1]. Matthieu Brucher later started work on the project as part of his thesis, also in 2007 ^[1]. Inria's version of the story emphasizes the community: the project was initially launched in 2007 by members of the Python scientific community, and had a new start in 2009 with the investment of Inria's Parietal project team ^[2]. The two accounts agree on the starting year and on the setting — the work grew out of the Python scientific community rather than out of a company's product plan.

First public release

The turning point came in 2010. Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort and Vincent Michel, of INRIA, took leadership of the project and made the first public release on 1 February 2010 ^[1]. After that the project settled into a rhythm: several releases have appeared following a roughly three-month cycle ^[1]. A predictable cadence is a practical virtue in software — users can plan their upgrades, and fixes reach them within months rather than after an indefinite wait.

How it is cited

Research software has an awkward trait: it is used constantly and cited rarely. The official page deals with this by asking users to cite a paper — Scikit-learn: Machine Learning in Python by Pedregosa and colleagues, volume 12, pages 2825 to 2830 of JMLR, the Journal of Machine Learning Research, 2011 ^[1]. Describing a library in a journal paper gives academic users a standard way to credit it in their own publications, and gives readers of those papers a fixed reference to point at.

Where it stands

From a 2007 summer-of-code project, to a 2009 restart backed by a research team, to a public release on 1 February 2010 and a roughly three-month release cycle since then ^[1][2]: the sequence describes a common path for scientific software, in which a community builds a tool and a research institution supplies people, funding and continuity. Inria now counts scikit-learn among the three most popular machine learning programs on GitHub ^[2], and the project continues to publish releases on its cycle ^[1].