-
Notifications
You must be signed in to change notification settings - Fork 39
Description
Describe the workflow you want to enable
I want to be able to plot a collection of local explanations as a dependence plot similar to SHAP. This should be possible for main effects / first-order values and also interactions.
Describe your proposed solution
Re-implement the scatter functionality in SHAP for also higher-Order scores to work with the InteractionValues object. For main effects or first-order scores the scatter plot would work exactly like shap. For higher-order scores the coloring of the instances would probably work similarly to the 1d case. However it’s unclear how to exchange the x-axis from a feature value with a value from feature interactions. @annprzy, solved this in the beeswarm plot with plotting multiple facets in the plot for each feature in an interaction. I don’t know if this would work here as well or if we should just limit the x-axis always to be only one feature. I think the best solution is to limit the x-axis to be one feature.
Describe alternatives you've considered, if relevant
Additional context
This is particularly interesting if one wants to do explanations in a regional setting.
Impact
Low (Nice to have)
Metadata
Metadata
Assignees
Labels
Projects
Status