Predictive models evaluation & inspection in scikit-learn

Scikit-learn is one of the fundamental open-source libraries for developing machine learning pipelines in both academic and industrial research.

When building a machine learning pipeline for a specific research problem, two key aspects are closely connected: (i) designing the pipeline and (ii) assessing, analyzing, and inspecting it.

Researchers strive to identify the optimal pipeline, maximizing specific evaluation metrics, while also seeking to explain the validity and rationale behind the pipeline's predictions. This is the cornerstone to properly answering research questions. With this proposal, we aim to improve and extend the available scikit-learn tools.

In the domain of model inspection, we aim to address several areas: (i) model inspection during training, (ii) enhancing user experience through interactive inspection, and (iii) model explainability.

To achieve these goals, we aim at implementing a "callback" framework, interactive tools for better integration in IDEs, visual displays for model evaluation, and a unified approach for model explainability.

On top of all these items, we intend to continue working on the general maintenance of the project, addressing bug reports and performance regressions. As a community-driven project, we want to dedicate time to reviewing external contributions.