adding sparse support to shap linear explainer #645
Conversation
| @@ -35,7 +36,8 @@ class LinearExplainer(Explainer): | |||
| input is correlated with another input, then both get some credit for the model's behavior. The | |||
| independent option stays "true to the model" meaning it will only give credit to features that are | |||
| actually used by the model, while the correlation option stays "true to the data" in the sense that | |||
| it only considers how the model would behave when respecting the correlations in the input data. | |||
| it only considers how the model would behave when respecting the correlations in the input data. | |||
| For sparse case only independent option is supported. | |||
| """ | |||
|
|
|||
| def __init__(self, model, data, nsamples=1000, feature_dependence=None): | |||
There was a problem hiding this comment.
one thing I am wondering about is the binary classification scenario - I believe we should be consistent with other explainers and output a list of shap values which the code is currently not doing (I believe we would need to take -coef for the negative class case if we are doing binary classification). One thing I am not sure about is how to determine if this is a binary classification model or a multiclass model with a single class, which would seem to have the same structure but should output different shap values:
# sklearn style model
elif hasattr(model, "coef_") and hasattr(model, "intercept_"):
# work around for multi-class with a single class
if len(model.coef_.shape) > 1 and model.coef_.shape[0] == 1:
self.coef = model.coef_[0]
self.intercept = model.intercept_[0]
else:
self.coef = model.coef_
self.intercept = model.intercept_
There was a problem hiding this comment.
I know this is a future direction we need to sort out. But it seems like for consistency we could eventually just always return a list for all multi-class outputs.
imatiach-msft
force-pushed
the
ilmat/linear-sparse
branch
from
9063fd9
to
dc309d1
Compare
|
Thanks! At first I was thinking that we would also want to have a sparse output (instead of dense), but since the mean offset is usually non-zero this is not actually that helpful, so dense seems best (as you have done). I am going to go ahead and merge this, with the idea of getting multi-output consistency as a later issue. |
similar to kernel explainer, adding scipy sparse support to linear explainer