XGBoost colsample_bylevel (col_sample_rate) not working with tree_method="approx"

[exalate-issue-sync]

When setting col_sample_rate to a value lower than 1.0 for an H2O XGBoost model, no sampling appears to happen. This is confirmed by:

• When col_sample_rate is the only parameter that triggers a stochastic process and the seed is fixed, models with different values for col_sample_rate (e.g 0.3, 0.5, 0.7, etc) are exactly the same
• When col_sample_rate is the only parameter that triggers a stochastic process (e.g. for a value of 0.5), models with different seeds are exactly the same

The value for col_sample_rate is logged correctly in Flow (also in the set of native XGBoost parameters), and it shows correctly in the MOJO file (see attachments). I did not notice this bug in earlier versions of H2O, but I have no way to verify this at the moment. I have verified that this problem shows up with different datasets. I verified that the parameter col_sample_rate_per_tree does seem to work as intended.

After investigation, the bug on the XGBoost side was found: [https://github.com/dmlc/xgboost/issues/7244|https://github.com/dmlc/xgboost/issues/7244|smart-link]

[exalate-issue-sync]

Veronika Maurerová commented: I created this new Jira ticket based on the [https://h2oai.atlassian.net/browse/PUBDEV-8266|https://h2oai.atlassian.net/browse/PUBDEV-8266|smart-link] , where we solved a different problem than [~accountid:5d9dc9eb87dd6f0dcb4d4d98] reported.

In this ticket I will add an error when someone wants to set tree_method=”approx” and some column sampling parameters together.

[h2o-ops-ro]

JIRA Issue Details

Jira Issue: PUBDEV-8368
Assignee: Veronika Maurerová
Reporter: Mathijs de Jong
State: Resolved
Fix Version: 3.34.0.4
Attachments: N/A
Development PRs: Available

[h2o-ops-ro]

Linked PRs from JIRA

#5811