LIME and multi-correlated variables

[elopezfune]

Hi Community!

I’ve noticed a potential issue with using the LIME package for ML model explainability, and I’d appreciate some clarification. Given that LIME is a local method, it generates data samples around the given data instance without considering any multicollinearity present in the data. This oversight can introduce significant variance in the estimation of the linear regression coefficients, thereby compromising the reliability of the explanations.

Has anyone else encountered this issue? If so, how have you addressed it?

Thanks!