[SPARK-21806] BinaryClassificationMetrics pr(): first point (0.0, 1.0) is misleading

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 2.2.0
• Fix Version/s: 2.3.0
• Component/s: MLlib
• Labels: releasenotes
• Assignee: Sean R. Owen
• Reporter: Marc Kaminski
• Created: 22/Aug/17 08:54
• Updated: 06/Jun/18 15:37
• Resolved: 30/Aug/17 10:36

Description

I would like to reference to a discussion in scikit-learn, as this behavior is probably based on the scikit implementation.

Summary:
Currently, the y-axis intercept of the precision recall curve is set to (0.0, 1.0). This behavior is not ideal in certain edge cases (see example below) and can also have an impact on cross validation, when optimization metric is set to "areaUnderPR".

Please consider blucena's post for possible alternatives.

Edge case example:
Consider a bad classifier, that assigns a high probability to all samples. A possible output might look like this:

Real label	Score
1.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	1.0
0.0	0.95
0.0	0.95
1.0	1.0

This results in the following pr points (first line set by default):

Threshold	Recall	Precision
1.0	0.0	1.0
0.95	1.0	0.2
0.0	1.0	0,16

The auPRC would be around 0.6. Classifiers with a more differentiated probability assignment will be falsely assumed to perform worse in regard to this auPRC.

Issue Links

• is related to: SPARK-24431 wrong areaUnderPR calculation in BinaryClassificationEvaluator
• links to: GitHub issue