You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First of all, thanks for the idea and implementartion!
We're adopting SHAP values in our model evaluations and find them very useful overall but struggle with the fact that there seems to be no established, programmatic way of determining directionality (think of + vs - coefficients in linear models and how this helps interpretation strongly).
From beeswarm plots, one can read the directionality of the top features by comparing where left/right of x==0 the red/blue dots concentrate. But this visual aid does not allow us to programmatically extract this info.
Alternative Solutions
One solution would be to compute spearman correlations between feature values and shap values. In our hands, this is very reliable and in essence equivalent to what our eye is doing when reading the beeswarm plot.
Do you think this a sensible approach? If approach sounds sensible perhaps one could add the directionality of each feature as a little symbol (perhaps +/- sign somewhere close to the y-axis in the beeswarm plot)?
If this sounds good I'd be happy to issue a PR.
Additional Context
No response
Feature request checklist
I have checked the issue tracker for duplicate issues.
I'd be interested in making a PR to implement this feature
The text was updated successfully, but these errors were encountered:
Problem Description
First of all, thanks for the idea and implementartion!
We're adopting SHAP values in our model evaluations and find them very useful overall but struggle with the fact that there seems to be no established, programmatic way of determining directionality (think of + vs - coefficients in linear models and how this helps interpretation strongly).
From beeswarm plots, one can read the directionality of the top features by comparing where left/right of x==0 the red/blue dots concentrate. But this visual aid does not allow us to programmatically extract this info.
Alternative Solutions
One solution would be to compute spearman correlations between feature values and shap values. In our hands, this is very reliable and in essence equivalent to what our eye is doing when reading the beeswarm plot.
Do you think this a sensible approach? If approach sounds sensible perhaps one could add the directionality of each feature as a little symbol (perhaps +/- sign somewhere close to the y-axis in the beeswarm plot)?
If this sounds good I'd be happy to issue a PR.
Additional Context
No response
Feature request checklist
The text was updated successfully, but these errors were encountered: