1
$\begingroup$

I'm trying to find reliable ways of comparing explainability of models. Parsimony is one, but are there other ways to tell whether a model is more explainable than another?

For instance, a black-box model that involves post-hoc explainability methods may be more complicated to explain than an intrinsically interpretable model. How can that explanation be compared to say that the intrinsically interpretable model is better (from an explainability perspective)?

$\endgroup$

1 Answer 1

0
$\begingroup$

There are many dimensions of model explainability:

  • Intrinsic or post hoc - Is the model itself explainable or does the explainability happen in addition to the model?
  • Model-specific or model-agnostic - Are the explainability tools limited to certain classes of models or can they be applied to any model?
  • Degree of locality - Does the explainability apply to a single datapoint, several datapoints, parts of the model, or entire model behavior?
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.