LoginSign Up
← Back to all sparks
M

midr

INFRA · APIS
Velocity0.0

Learning from Black-Box Models by Maximum Interpretation Decomposition

github.com/ryo-asashi/midr ↗

A black-box interpreter reaches CRAN, then learns multi-class and survival responses

explainable-aisurrogate-modelsshapleysurvival-analysismulti-classr-package
◆Current state
midr explains black-box models by fitting an interpretable surrogate through Maximum Interpretation Decomposition — main effects plus second-order interactions, with exact Shapley values for the surrogate. Two months after its first CRAN release it can take a matrix response, which covers multi-class classification and survival models, and hold collections of fitted interpretations in midlist and midrib objects for comparison.
◆Where it's heading
The releases move outward along two axes at once: what can be interpreted, and how much of it fits in memory. Version 0.5.3 rebuilt the fitting path to avoid materialising large design matrices and added a save.memory option; 0.6.0 widened the response from a vector to a matrix and added parametric link functions. Class and argument names were shortened in the same release, so the package is still willing to break itself this early.
◆Prediction
With multiple models now held in one object and visualisation methods for them, comparison across models is the surface most likely to fill out next — the collection classes exist but the notes describe manipulation and plotting rather than any comparison metric.

◆Recent moves

  1. 5mo ago

    Matrix responses bring multi-class and survival models in scope

    ⚡ SPARK

    Two months after arriving on CRAN, midr widens what it can interpret: interpret() takes a matrix for the response, which covers multi-class classifiers and survival models, and midlist and midrib objects hold several interpretations together. Class names are shortened and two arguments renamed in the same release.

    View source ↗
  2. 7mo ago

    First CRAN release: MID surrogate models for black-box explanation

    ⚡ SPARK

    The package's first stable release, and the base everything since builds on: decomposition of predictions into main and second-order interaction effects, a full set of base and ggplot2 visualisations, and exact Shapley values for the surrogate with shapviz integration.

    View source ↗
  3. 7mo ago

    Memory-efficient fitting for large design matrices

    Scaling work rather than new scope: the fitting path's space complexity is improved for datasets with many observations, a save.memory option is added, and the prediction engine is rewritten to stop storing term effect matrices. Note the stamps — this fourth release was published 28 seconds before the initial one.

    View source ↗