Expose Decision Tree as a stand-alone algo in H2O

[exalate-issue-sync[bot]]

We should decide whether we want to make a new REST endpoint for this (e.g. /3/ModelBuilders/decisiontree) or if we want to simply wrap the Random Forest or GBM function at the R/Py client level. However, if it's done in H2O core, then it will show up in Flow as well, which is nice.

You can get a single decision tree by setting the following args in RF:

• ntrees = 1
• mtries = # of features (would be determined dynamically at runtime)
• sample_rate = 1 Note: min_rows defaults to 1 and max_depth to 20

Or, you can get a single decision tree by setting the following args in GBM (maybe this is better because it skips the data-dependent mtries arg):

• ntrees = 1
• min_rows = 1
• sample_rate = 1
• col_sample_rate = 1 Note: min_rows defaults to 10 and max_depth to 5

I'm wondering if we also want to set max_depth to something large so that we don't apply regularization by default...

Related: https://0xdata.atlassian.net/browse/PUBDEV-4007

[exalate-issue-sync[bot]]

Patrick Hall commented: Would be really nice to ...

• Have a plot of the tree
• Have english language rules that define the tree splits

[exalate-issue-sync[bot]]

Arno Candel commented: Yes, agree, the only point of this would be to take it a step further.

[exalate-issue-sync[bot]]

Patrick Hall commented: Moving ongoing stories for machine learning interpretability into H2O Driverless AI project.

[exalate-issue-sync[bot]]

Patrick Hall commented: Moving ongoing stories for machine learning interpretability into H2O Driverless AI project.

[exalate-issue-sync[bot]]

Erin LeDell commented: Why was this closed? I am going to re-open it. This is not just relevant to MLI, this is a general H2O improvement.

[h2o-ops]

JIRA Issue Migration Info

Jira Issue: PUBDEV-4324 Assignee: Navdeep Gill Reporter: Erin LeDell State: Reopened Fix Version: N/A Attachments: N/A Development PRs: N/A