AI Governance Engineer
GitHub

The explanation technique map

Explanation techniques on two axes, global or local and model-agnostic or model-specific, each tested before its explanation record is kept.

By Updated First published

The explanation technique map Explanation techniques on two axes, global or local and model-agnostic or model-specific, each tested before its explanation record is kept. Two axes: scope and access Global the whole model Local one output Model-agnostic inputs and outputs only Global surrogate models Permutation feature importance Partial dependence LIME KernelSHAP Counterfactual explanations Nearest-example explanations Model-specific uses the internals Coefficients of an interpretable model Tree structure Probing of internal representations TreeSHAP Integrated gradients Attention or circuit analysis (research) an output: test it Explanation tests fidelity, stability, sanity, comprehension Explanation record method, version, baseline, reason codes Layer 03 Evals & Red Teaming as Evidence Layer 04 Runtime Controls & Observability
The explanation technique map Explanation techniques on two axes, scope (global or local) and access (model-agnostic or model-specific), and the tests an explanation passes before its record is kept as evidence. Pick the quadrant your decision needs, then test the method for fidelity and stability before you rely on it. Drawn from chapter 16.

Text alternative

Explanations vary along two axes: scope (a global explanation describes the model's overall behaviour; a local one explains a single output) and access (a model-agnostic method needs only inputs and outputs; a model-specific one uses the model's internals). Model-agnostic and global: global surrogate models, permutation feature importance and partial dependence. Model-agnostic and local: LIME, KernelSHAP, counterfactual explanations and nearest-example explanations. Model-specific and global: the coefficients of an interpretable model, tree structure and probing of internal representations. Model-specific and local: TreeSHAP, integrated gradients and other gradient attributions, and attention or circuit analysis (research). An explanation is an output, so it gets evals like any other output (Layer 03 Evals & Red Teaming as Evidence): fidelity, stability and sanity tests without people, and comprehension tests with them. The unit of evidence is the explanation record, one structured object per explained decision, written at decision time by the runtime (Layer 04 Runtime Controls & Observability) with the method, its version, the baseline and the reason codes pinned.

Download

Every file carries the attribution band "aigovernanceengineer.com · CC BY 4.0 · v0.5.0" inside the image, and the version is in the file name, so a copy always says where it came from and which edition it shows. The SVGs keep the text live: the first follows the viewer's light or dark setting, the other two fix one theme for slides and print. The PNGs are drawn with the site's own typefaces.

Explanation techniques on two axes, global or local and model-agnostic or model-specific, each tested before its explanation record is kept.
The PNG, light, 1600 px wide, as it downloads (shown here as a lighter copy).

Reuse and credit

The figure is published under CC BY 4.0: you may copy, share and adapt it, commercially too, provided you give appropriate credit, link to the licence and say if you changed it. Keep the attribution band in the image. A credit line that covers title, author, source and licence:

“The explanation technique map” by Jorge García Aibar, aigovernanceengineer.com (https://aigovernanceengineer.com/figures/explanation-techniques), v0.5.0. Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

Embed with HTML

<figure>
  <img src="https://aigovernanceengineer.com/downloads/figures/explanation-techniques-v0.5.0-light-1600.png" alt="Explanation techniques on two axes, global or local and model-agnostic or model-specific, each tested before its explanation record is kept." width="800" height="1272" loading="lazy">
  <figcaption>
    <a href="https://aigovernanceengineer.com/figures/explanation-techniques">The explanation technique map</a> by Jorge García Aibar,
    aigovernanceengineer.com, v0.5.0.
    Licensed under <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a>.
  </figcaption>
</figure>

Embed with Markdown

![Explanation techniques on two axes, global or local and model-agnostic or model-specific, each tested before its explanation record is kept.](https://aigovernanceengineer.com/downloads/figures/explanation-techniques-v0.5.0-light-1600.png)

*[The explanation technique map](https://aigovernanceengineer.com/figures/explanation-techniques) by Jorge García Aibar, aigovernanceengineer.com, v0.5.0. Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).*