search

interpretml / interpret

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

Lookup Table for single feature and feature interaction terms #530

Open sauvikd opened 3 weeks ago

sauvikd commented 3 weeks ago

Screenshot 2024-04-19 035202

How can I extract a lookup table for the red-line (score values for the each of the features) and similarly the data for the heatmap.

I had tried downloading the ebm_global.data(1)['scores'], but those are values for each of the bins, from what I understand. Even if I'm able to extract the bin boundary values, I will be able to connect them with the feature values.

paulbkoch commented 3 weeks ago

Hi @sauvikd --

One of the very nice things about EBMs, and GAMs in general, is that the scores are used when calculating the predictions AND what are shown on the graphs. EBMs have no hidden information.

For graphing though, you also need the "bins" attribute, which for continuous features are the locations on the X-axis where the bin transitions occur.

There's some code which creates graphs here: https://github.com/interpretml/interpret/issues/325

And you might also find our documentation on EBM internals useful: https://interpret.ml/docs/ebm-internals-regression.html