midr - Learning from Black-Box Models by Maximum Interpretation Decomposition
The goal of 'midr' is to provide a model-agnostic method for interpreting and explaining black-box predictive models by creating a globally interpretable surrogate model. The package implements 'Maximum Interpretation Decomposition' (MID), a functional decomposition technique that finds an optimal additive approximation of the original model. This approximation is achieved by minimizing the squared error between the predictions of the black-box model and the surrogate model. The theoretical foundations of MID are described in Iwasawa & Matsumori (2025) [Forthcoming], and the package itself is detailed in Asashiba et al. (2025) <doi:10.48550/arXiv.2506.08338>.
Last updated
actuarialimlinterpretable-machine-learningxaicpp
5.64 score 6 stars 1 dependents 18 scripts 519 downloads
midnight - A 'tidymodels' Engine and Other Extensions for the 'midr' Package
Provides a 'parsnip' engine for the 'midr' package, enabling users to fit, tune, and evaluate Maximum Interpretation Decomposition (MID) models within the 'tidymodels' framework. Developed through research by the Moonlight Seminar 2025, a study group of actuaries from the Institute of Actuaries of Japan, to enhance practical applications of interpretable modeling. Detailed methodology is available in Asashiba et al. (2025) <doi:10.48550/arXiv.2506.08338>.
Last updated
imlinterpretable-machine-learningparsnipxaicpp
4.15 score 2 stars 3 scripts 458 downloads