DALEX
DALEX copied to clipboard
Adding partial derivative and partial log-derivative
I think it could useful a function to calculate numerically not just the partial plot of the predictor, like in the function variable_effect_partial_dependency , but also the partial derivative of the predictor for the additive effect:
and the logarithmic partial derivative for the multiplicative effect:
It could be possible to calculate it like tha partial response in the variable_effect_partial_dependency function: fixing all other covariates and changing just X_j.
Hi,
thank you so for your valuable ideas! We will try to implement them.
@pbiecek what do you think about adding derivative_partial type to model_profile
and thus to aggregate_profiles
in ingredients
. IMO it fully complies with the definition of the Model Profile task. We could estimate the partial derivative of the predictor by setting a very dense grid of points to evaluate for standard profiles and then just dividing proper values to get the derivative.
Hi,
thank you so for your valuable ideas! We will try to implement them.
@pbiecek what do you think about adding derivative_partial type to
model_profile
and thus toaggregate_profiles
iningredients
. IMO it fully complies with the definition of the Model Profile task. We could estimate the partial derivative of the predictor by setting a very dense grid of points to evaluate for standard profiles and then just dividing proper values to get the derivative.
If you make some progress, please reply and tell me how you called the function (or if it was a failure and it doesn't work), I am curious to test it and understand if it gives interesting results.
@pbiecek are you continuing on this?