search

slds-lmu / iml_methods_limitations

Seminar on Limitations of Interpretable Machine Learning Methods
https://slds-lmu.github.io/iml_methods_limitations/
57 stars 15 forks source link

two questions about 04-09-lime-neighbourhood.Rmd #93

Open QWERTZ11 opened 4 years ago

QWERTZ11 commented 4 years ago

  1. locally fidele models: "fidele" is unknown word? And I can't find this in the paper mentioned.

  2. 328-th row: “For globally non-linear associations, we also find that we prefer a small kernel which however also produces stable coefficients.” However,  339-th row: The neighborhood problem can be described briefly by the following.
    A (too) small kernel width creates unstable coefficients whilst a too large kernel width fits a global surrogate model. (problematic?? Conflict ??)

 469-th row: At the same time, our study agrees with @alvarez2018robustness who find that local explanations are highly unstable. (problematic?? Conflict ??)

QWERTZ11 commented 4 years ago

@christophM Hi Christoph, Here are two small questions, could you help me?

SebGGruber commented 4 years ago
  1. Maybe try "fidelity" instead - does this make more sense for you?
  2. is a typo - smaller kernel width usually always worsens coefficient stability
QWERTZ11 commented 4 years ago

Thank you! @SebGruber1996