So, I was wondering how LIME's linear model approach compares with other explanation metrics, in special, with Mutual Information? For those unfamiliar with how LIME works:

  1. Choose the instance you want to explain
  2. Probe your model with a few perturbations on that instance
  3. Fit a linear model to those points, points closer to the original instance are worth more (weighted heavier)
  4. You can now look to the (simpler) linear model, instead of your RBF-kernel SVM (much harder to interpret)

What would be advantages and disadvantages of using a different "model" in step 3, in this, calculating the mutual information of the model-assigned label and each of the features?

  • $\begingroup$ Can you please also provided references about how mutual information is used to explain individual predictions? I am only aware of the very recent work of Chen et al. but that's about it... $\endgroup$
    – usεr11852
    Commented Sep 23, 2018 at 20:24
  • $\begingroup$ In this case in particular, instead of fitting a linear model to the probed instances, one would instead compute how much information each feature provides for determining the model-assigned label, again, probed instances would be weighted heavier if they are closer to the instance that is being explained, for reference, here is the LIME repo: github.com/marcotcr/lime, the general approach here is replacing the Linear model with the Mutual Information as a mean to assign feature importance values. My question is: is there any reason to prefer this over Linear models? $\endgroup$ Commented Sep 27, 2018 at 13:31
  • $\begingroup$ Apologies, Maybe I was unclear, I am not asking about LIME; I am asking about the MI part of your question. Where have you seen a methodology using it to explain individual predictions. $\endgroup$
    – usεr11852
    Commented Sep 27, 2018 at 19:05


Your Answer

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