Global Importance Feature does not work in crossvalidation

[mdibiase2]

Hi all,

i am trying to compute global importance among model predictors.
I ran a gradient boosted tree model and i followed (as it looks like to me) all the steps needed by reading knime XAI guide.

Despite of this, the node does not work properly.
May it be related to cross-validation loop wrong setting ?
I set a workflow start-end couple of node in order to link it to the global importance node, but i think i am wrong in something i don’t know…

Here my workflow
dibmir-xai.knwf (3.5 MB)

[k10shetty1]

Hello @mdibiase2,

I cannot test the shared workflow since your data is on your local machine. Here you can find how to share the data with your workflow using relative paths.

I can see that you’re using cross-validation while calculating the global feature importance metric. This metric requires two inputs: first, the input model, representing the model to be explained, needs to be captured within a Workflow Object via Integrated Deployment. While you’ve captured both the learner and predictor nodes in the workflow, it’s only the final model or predictor node requiring explanation should be captured

Since cross-validation is mostly used for model selection and hyperparameter tuning, it’s best to first train multiple models, select the best-performing one, store it, and then capture it before passing it to the component. This makes sure that we are using the most optimal model for feature importance analysis.

Hope this helps.

Best,
Keerthan