GBM Interaction Gain and Variable Importance Inconsistency about h2o-3 HOT 3 OPEN

Michal Kurka commented: Thank you for the report and your investigation - it does look suspicious.

I would be just cautious about trying to manually reconstruct the tree predictions, what I would do is to use leaf node assignment to see where the observations actually end-up: [https://docs.h2o.ai/h2o/latest-stable/h2o-r/docs/reference/predict_leaf_node_assignment.H2OModel.html|https://docs.h2o.ai/h2o/latest-stable/h2o-r/docs/reference/predict_leaf_node_assignment.H2OModel.html|smart-link]

Even simple comparison using {{>=}} can in some cases due to different precision considered (float32 vs float64)

{noformat}h2o.ifelse(train_data_h2o$x2 >= tree@thresholds[1], gbm_model@model$init_f + tree@predictions[[3]], gbm_model@model$init_f + tree@predictions[[2]]){noformat}

We will take a look.

from h2o-3.

Dominick Sullivan commented: [~accountid:557058:04659f86-fbfe-4d01-90c9-146c34df6ee6] - I’m the OP of this ticket (wasn’t signed in at the time of the post). Just wanted to follow up on the progress of this. Thanks again!

from h2o-3.

JIRA Issue Details

Jira Issue: PUBDEV-8884
Assignee: Adam Valenta
Reporter: N/A
State: Open
Fix Version: N/A
Attachments: N/A
Development PRs: N/A

from h2o-3.