[SPARK-15699] Add chi-squared test statistic as a split quality metric for decision trees

• Type: New Feature
• Status: Resolved
• Priority: Minor
• Resolution: Incomplete
• Affects Version/s: 2.0.0
• Component/s: ML, MLlib
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Erik Erlandson
• Created: 01/Jun/16 14:37
• Updated: 20/Jan/20 00:05
• Resolved: 21/May/19 04:37

Description

Using test statistics as a measure of decision tree split quality is a useful split halting measure that can yield improved model quality. I am proposing to add the chi-squared test statistic as a new impurity option (in addition to "gini" and "entropy") for classification decision trees and ensembles.

I wrote a blog post that explains some useful properties of test-statistics for measuring split quality, with some example results:
http://erikerlandson.github.io/blog/2016/05/26/measuring-decision-tree-split-quality-with-test-statistic-p-values/

(Other test statistics are also possible, for example using the Welch's t-test variant for regression trees, but they could be addressed separately)

Issue Links

• Is contained by: SPARK-14045 DecisionTree improvement umbrella
• is related to: SPARK-13868 Random forest accuracy exploration
• links to: [Github] Pull Request #13440 (erikerlandson)
• links to: GitHub Pull Request #13440