0

How to compute conditional permutation importance from h2o.gbm?

I have a data set with many highly correlated variables(>0.9). And fed this data set to h2o.gbm. As it turned out, RMSE increases (on CV) when I drop down correlated variables.

Now I'm trying to get variable importance and found just this function: h2o.varimp(). Which is (I guess) differs from classic party::varimp(model, conditional = T).

Bear
  • 662
  • 1
  • 5
  • 20
Amirullo Faiazov
  • 1
  • 1

1 Answers1

0

This is a duplicate question, you can find previously posted answers here and here.

The short answer is this is not currently an option in H2O-3, for alternative options please see the linked questions above.

In addition a jira ticket has been created for this issue that you can track and comment on here: https://0xdata.atlassian.net/browse/PUBDEV-4027

Lauren
  • 5,640
  • 1
  • 13
  • 19
  • Thank you for you answer. It is simple to calculate 'original' permutation importance, but way harder to make it conditional. Previous questions where about 'original'. Looking forward for updates. – Amirullo Faiazov Oct 11 '18 at 07:35
  • okay great thanks for emphasizing that detail, since the jira ticket encompasses adding more importance metrics in general I've updated my comment on that ticket to specifically call out conditional permutation. – Lauren Oct 11 '18 at 14:35