search

interpretml / interpret

Fit interpretable models. Explain blackbox machine learning.
https://interpret.ml/docs
MIT License
6.04k stars 715 forks source link

Question: what is the name of the 3-dimensional plot / analysis visibile in the output? #490

Open alatif-alatif opened 6 months ago

alatif-alatif commented 6 months ago

Hi,

thanks for creating and maintaining this project!

I have some slightly off-topic questions:

For my data, I am interested in this part of interpretml:

logit of predicted target dependent on two variables I understand that output as the logit of the target dependent on two specific features

My questions are:

  1. As it seems, I would be able to build a certain degree of predictiveness if I just take the conditions visible in that plot and have a simple algorithm that makes predictions of the target by taking the two specific features and some rules like mean texture > 27 and worst radius < 15. Is my understanding correct here?

  2. How is an analysis called that leads to such an insight (it should be doable without machine learning, if I am not wrong...)? Is there perhaps a specific expression for this?

  3. Are there perhaps any (python?) tools / frameworks that perform such an analysis that I could use in addition to interpretml, if I would like to get more possibilities to control the process of getting such a result?

Thanks in advance!

paulbkoch commented 6 months ago

Hi @alatif-alatif --

Yes, you are correct. You can convert the pairwise interaction graph above into a big if statement that you can then use in code directly. There is another project called ML_2_SQL that does this for EBMs if you're interested in seeing an implementation: https://github.com/kaspersgit/ml_2_sql . Someday we hope to have converters that turn EBMs into things like javascript, python, C++, etc.. for zero dependency scenarios.