Improve the default hyperparameter values for XGBoost models in AutoML

[exalate-issue-sync[bot]]

Improvements suggested for XGBoost default models inside AutoML:

• xgBoostParameters._max_depth should not be 20 (XGB def2). The deepest we should go on this is probably 15 or 13. For reference def1 is depth 10 and def3 is depth 5.
• xgBoostParameters._min_rows looks small. xgBoostParameters._min_rows vaguely scales with the number of rows. For a tiny dataset (20 K rows) it might be 3-15. For a 150K row dataset it might be 20-60. For a 500k row dataset probably 100-300.
• xgBoostParameters._sample_rate should probably be 0.8-1.0 by default. Frequently I end up with 1.0, so I might start with that. It's seldom as small as the 0.6.
• xgBoostParameters._col_sample_rate and xgBoostParameters._col_sample_rate_per_tree are cumulative so if they are both 0.8, that's 0.64 fraction of the features at each split. I vaguely expect the cumulative value to be larger (0.8?) for a small number of columns (20?) and smaller (0.3?) for a larger number of columns(100?).

[h2o-ops]

JIRA Issue Details

Jira Issue: PUBDEV-8173 Assignee: Sebastien Poirier Reporter: Erin LeDell State: Open Fix Version: Backlog Attachments: N/A Development PRs: N/A