Mlr3extralearners: Expanding the Mlr3 Ecosystem With Community-Driven Learner Integration

MCML Authors

Marc Becker

→ Group Bernd Bischl
 Statistical Learning and Data Science

Martin Binder

→ Group Bernd Bischl
 Statistical Learning and Data Science

Lennart Schneider

→ Group Bernd Bischl
 Statistical Learning and Data Science

Lukas Burk

→ Group Bernd Bischl
 Statistical Learning and Data Science

Henri Funk

→ Group Helmut Küchenhoff
 Statistical Consulting Unit (StaBLab)

Andreas Bender

Dr.

→ Group Bernd Bischl
 Statistical Learning and Data Science

Bernd Bischl

Prof. Dr.

Director

Statistical Learning and Data Science

Abstract

The mlr3 ecosystem is a versatile toolbox for machine learning (ML) in R (R Core Team, 2019) that is targeted towards both practitioners and researchers (Bischl et al., 2024). The core mlr3 package (Lang et al., 2019) defines the standardized interface for ML, but its goal is not to implement algorithms. This is, e.g., done by the mlr3learners extension (Lang, Au, et al., 2024) that connects 21 stable learning algorithms from various R packages to the mlr3 ecosystem that serve as a good starting point for many ML tasks. In addition, mlr3extralearners is a community-driven package that integrates many more methods. The package currently wraps 147 different ML algorithms from many different R packages, for tasks such as classification, regression, and survival analysis. This enables users to seamlessly access and utilize these learners directly within their workflows. One of the strengths of mlr3 is the design and implementation of large-scale benchmark experiments. For example, datasets for such experiments can be easily obtained from the OpenML1 repository (Vanschoren et al., 2014) via the mlr3oml package (Lang & Fischer, 2024). Furthermore, strong support for parallelization, including execution on high-performance computing clusters via batchtools (Lang et al., 2017) and its mlr3 integration mlr3batchmark (Becker & Lang, 2024), is available and well documented (Fischer et al., 2024). In combination, these tools allow for large-scale empirical investigations, which has, for example, been used to collect and analyze data about hyperparameter landscapes of ML algorithms (Binder et al., 2020). An overview of all mlr3 learners, including those introduced through mlr3extralearners, is available on the mlr3 website. Beyond accessibility, mlr3extralearners also allows mlr3 users to easily connect their own algorithms to the interface. This enriches each learner with extensive metadata about its hyper-parameter space, prediction types, and other key attributes. Furthermore, mlr3extralearners includes robust mechanisms for quality assurance, such as regular automated sanity checks and verification tests that ensure learner parameters are consistent and up-to-date with the latest versions of their underlying R packages. In order to allow the integration of learners that are not available on CRAN, the package is hosted on the mlr R-universe3. By providing a standardized interface and comprehensive metadata for each learner, mlr3extralearners enhances the FAIRness (findability, accessibility, interoperability, and reusability) of ML algorithms within the R ecosystem (Wilkinson et al., 2016).

article FZS+25

The Journal of Open Source Software

Authors

Links

Research Areas

 A1 | Statistical Foundations & Explainability

 C4 | Computational Social Sciences