Single Decision Tree

[exalate-issue-sync[bot]]

Implement Standard Decision Tree.

Specific task for student Yuliia Syzon.

[exalate-issue-sync[bot]]

Adam Valenta commented: Currently, we are focusing on binomial classification with UniformAdaptive histogram.

In progress:

• Fix bug with different prediction on different machines
• Benchmark training time and scalability agains DRF
• Fix the binning to go through data only ones and not for each bin and each feature

Major stuff on roadmap:

• Support categorical splits
• Support NaNs splits
• Support other types of histograms: UniformAdaptive (/) Random QuantilesGlobal RoundRobin
• Support Tree printing
• Support multinomial classification
• Support regression
• Support scoring in training process
• Support grid search
• Support weights
• Support early stopping
• Make findBestSplit parallel
• Support cross validation
• Support MOJO
• Support Machine Learning Interpretability stuff: Shapley, PDP plot,…

Minor stuff to be done:

• Restructuralize code to StandardDecisonTree class and try to guess interface for shallow tree
• add parameters nbins, mtries, custom_metric, min_rows, (nbins_cats), min_split_improvement, categorical_encoding, check_constant_response
• Investigate parameters nbins_top_level
• API test, save-load test,
• Benchmark of training time compared to DRF
• Benchmark of prediction performance to DRF, Decision Tree (/)
• Sphinx documentation
• Enhance code documentation (Python examples, R examples…)

Other

• Support sample_rate_per_class, col_sample_rate_change_per_level, checkpoint,

[exalate-issue-sync[bot]]

Wendy Wong commented: For this JIRA, the focus is on:

• binary classification trees only
• numerical predictors only
• python client support
• R client support

For performance, please also add R single decision tree to comparison. Wendy Wong can assist in this effort.

[h2o-ops]

JIRA Issue Details

Jira Issue: PUBDEV-8691 Assignee: Yuliia Syzon Reporter: Adam Valenta State: In Progress Fix Version: 3.42.0.1 Attachments: N/A Development PRs: Available

[h2o-ops]

Linked PRs from JIRA

https://github.com/h2oai/h2o-3/pull/6182 https://github.com/h2oai/h2o-3/pull/6447

[wendycwong]

Next steps:

1. add NA support;
2. add multinomial classification;
3. add regression.