feature_importances_ of BaseGradientBoostedSetTree (gbest) errors

[liorshafir]

Hi,

When trying to get the feature importances of a trained gbest, the following errors occur:

1. line 622 in gbest.py: 'SetSplitNode' object has no attribute 'node_count'
   I tried to fix it by replace the 'node_count' with 'tree.n_nodes'
   but even then, another error appears:

2. When trying to execute:
   relevant_trees[0].tree_.compute_feature_importances(normalize=False)
   we get
   AttributeError: 'SetSplitNode' object has no attribute 'compute_feature_importances'

It looks like SetSplitNode (or rather settree.set_tree.SetTree) does not implement compute_feature_importances method.

Can you suggest any alternative to calculate the feature importances?

Thanks in advance

[RoyHirsch]

Hi Lior,
Thank you for your interest in our project. You are right, I've added support for impurity-based feature importance for Set-Tree and GBeST (same mechanism as in Sklearn).
19f75f7
Our implementation is based on Sklean but we do not fully support all of Sklearn's functionality (for example training with weighted samples is not fully supported).

I would like to highlight another result of our work - the item-based importance score as opposed to the feature-based importance. We believe that item-based insights will also be useful for many set prediction tasks. Don't hesitate to reach out with other questions and comments.

Thank you,
Roy

[liorshafir]

Hi Roy,

Thank you for your response. Appreciate it.
I will definitely look into the the item-based importance.

Thanks,
Lior