Grouped/groupwise permutation importance: permute all features in a group jointly and report one score per group. This was the most-requested feature in the tracker's history (#62, #117 — closed as not-planned, but I said I'd take a PR, and it's small enough to just do).
Sketch: feature_groups = named list of character vectors; each group becomes one "feature" in the permutation loop (permute the group's columns with a common row shuffle); Variable column takes the group names. DALEX::model_parts(variable_groups = ...) is prior art.
Grouped/groupwise permutation importance: permute all features in a group jointly and report one score per group. This was the most-requested feature in the tracker's history (#62, #117 — closed as not-planned, but I said I'd take a PR, and it's small enough to just do).
Sketch:
feature_groups= named list of character vectors; each group becomes one "feature" in the permutation loop (permute the group's columns with a common row shuffle);Variablecolumn takes the group names.DALEX::model_parts(variable_groups = ...)is prior art.