Marc Kaminski created SPARK-21806:
-------------------------------------
Summary: BinaryClassificationMetrics pr(): first point (0.0, 1.0)
is misleading
Key: SPARK-21806
URL: https://issues.apache.org/jira/browse/SPARK-21806
Project: Spark
Issue Type: Improvement
Components: MLlib
Affects Versions: 2.2.0
Reporter: Marc Kaminski
Priority: Minor
I would like to reference to a [discussion in scikit-learn|
https://github.com/scikit-learn/scikit-learn/issues/4223], 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|https://github.com/scikit-learn/scikit-learn/issues/4223#issuecomment-215273613]
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 |
|1.0 | 1.0 |
|1.0 | 1.0 |
|1.0 | 1.0 |
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
