The output score of an individual-level feature

[JWKKWJ123]

Hi all,
I would like to ask if I want to use the ebm.eval_terms(.) to get local explanations or the ebm.term_importances(.) to get global explanations, what is the exactly of output score of an individual-level feature k in a classification task? Is it exactly the output of shape functions f_k(.), or the a normalized number?
If the feature importance score is a normalized number, is it the output of sigmoid function like this:

[paulbkoch]

Hi @JWKKWJ123 -- eval_terms finds the vertical value on the shape plots for the feature values in X. For EBMs, that value is equal to the local explanations.

term_importances are not normalized, and no sigmoid is applied. term_importances can be calculated with eval_terms. If you take the mean of the absolute values returned from eval_terms, it will be equal to the term importances.

[JWKKWJ123]

Hi @JWKKWJ123 -- eval_terms finds the vertical value on the shape plots for the feature values in X. For EBMs, that value is equal to the local explanations.

term_importances are not normalized, and no sigmoid is applied. term_importances can be calculated with eval_terms. If you take the mean of the absolute values returned from eval_terms, it will be equal to the term importances.

Hi Paul, Thank you very your reply!
Indeed I used to took the mean absolute value of the local explanations and it have the same result as the output of the term_importances .
I have another question that if I want to compare the feature importance among EBMs who trained on same kinds of features, do I need to do the normalization? If it is needed, what kind of normalization do you recommend (L1/L2/sigmoid/minmax/...)?

[paulbkoch]

For classification, the scores are in logits and should be comparable across datasets. For regression, the scores are in the units being predicted, so you'll probably want to normalize across datasets. There isn't a generally agreed upon normalization for regression. I use the interquartile range in our benchmarks.

[JWKKWJ123]

For classification, the scores are in logits and should be comparable across datasets. For regression, the scores are in the units being predicted, so you'll probably want to normalize across datasets. There isn't a generally agreed upon normalization for regression. I use the interquartile range in our benchmarks.

Hi Paul,
Thank you for your reply!