Aggregating local mask for an aggregate feature importance mask

[ss8319]

Hi guys, thank you for implementing the codebase for this.

I was just wondering if there is any feature to aggregate the local interpretability mask at every decision into a global decision mask.

I cant seem to find that function in the codebase/documentation

The original paper details combining the local mask

[Optimox]

The explanability method give you both the aggregated feature importance mask and the individual steps mask no ?

tabnet/pytorch_tabnet/tab_network.py

Line 190 in 2c0c4eb

def forward_masks(self, x):

[Optimox]

The global feature importance from the training set is accessible from using clf.feature_importances_

tabnet/pytorch_tabnet/abstract_model.py

Line 750 in 2c0c4eb

def _compute_feature_importances(self, X):

[ss8319]

Sure, thanks! is there a way to compute for the test set? But it does look like simple edits to the code should be enough to make it work too for test set.

…

On Sat, 9 Aug 2025 at 7:08 AM, Optimox ***@***.***> wrote: *Optimox* left a comment (dreamquark-ai/tabnet#581) <#581 (comment)> The global feature importance from the training set is accessible from using clf.feature_importances_ https://github.com/dreamquark-ai/tabnet/blob/2c0c4ebd2bb1cb639ea94ab4b11823bc49265588/pytorch_tabnet/abstract_model.py#L750 — Reply to this email directly, view it on GitHub <#581 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARMWSSYHRBIL6KQARYIYSAT3MUGVDAVCNFSM6AAAAACDMYTUP2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTCNRZGMYDQNJTHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Optimox]

clf._compute_feature_importances(X_test) should work