I have a binary classification - "BAD" and "GOOD" samples. The features are binary as well, either 0 or 1 (each sample is a boolean vector of size 264. I got about 3000 "BAD" samples, and 3000 "GOOD" samples.

I want answer the question: "which feature is critical to the sample being BAD".

Currently, When i use the SHAP values, I order them by mean(abs(shap_vals)) i get which feature is improtant to the decision in general, not necessarily "which feature when 1 is likely to turn the sample to BAD".

I thought about sorting them by mean(shap_vals) without the abs. But i have no theoretical knowledge on the affect of the change.

Any recommendations?

  • $\begingroup$ a straight-forward approach is to train multiple models each one leaving out some features and compare the bad results to identify which features are more likely to guide the BAD identification, IMO $\endgroup$
    – Nikos M.
    Oct 21, 2021 at 17:13
  • 2
    $\begingroup$ I suggest to more precisely define "is critical" in the first place because it's a very vague phrase. $\endgroup$
    – Jonathan
    Oct 22, 2021 at 22:24


Your Answer

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

Browse other questions tagged or ask your own question.