You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Sean Owen (JIRA)" <ji...@apache.org> on 2016/12/13 20:43:58 UTC
[jira] [Commented] (SPARK-18844) Add more binary classification
metrics to BinaryClassificationMetrics
[ https://issues.apache.org/jira/browse/SPARK-18844?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15746212#comment-15746212 ]
Sean Owen commented on SPARK-18844:
-----------------------------------
Yeah I think we discussed something like this before, and the drawback was just filling up the API with variations that are mostly not used. False positive and specificity might see some use. This would have to go in MulticlassMetrics too, and the Python API of both, for completeness.
Still that doesn't mean it's not doable. I tend to agree that it makes sense if anything to expose the 'computer' API, but then it's not clear how to translate that to multiclass and Python.
> Add more binary classification metrics to BinaryClassificationMetrics
> ---------------------------------------------------------------------
>
> Key: SPARK-18844
> URL: https://issues.apache.org/jira/browse/SPARK-18844
> Project: Spark
> Issue Type: Improvement
> Components: MLlib
> Affects Versions: 2.0.2
> Reporter: Zak Patterson
> Priority: Minor
> Labels: evaluation
> Fix For: 2.0.2
>
> Original Estimate: 5h
> Remaining Estimate: 5h
>
> BinaryClassificationMetrics only implements Precision (positive predictive value) and recall (true positive rate). It should implement more comprehensive metrics.
> Moreover, the instance variables storing computed counts are marked private, and there are no accessors for them. So if one desired to add this functionality, one would have to duplicate this calculation, which is not trivial:
> https://github.com/apache/spark/blob/v2.0.2/mllib/src/main/scala/org/apache/spark/mllib/evaluation/BinaryClassificationMetrics.scala#L144
> Currently Implemented Metrics
> ---
> * Precision (PPV): `precisionByThreshold`
> * Recall (Sensitivity, true positive rate): `recallByThreshold`
> Desired additional metrics
> ---
> * False omission rate: `forByThreshold`
> * False discovery rate: `fdrByThreshold`
> * Negative predictive value: `npvByThreshold`
> * False negative rate: `fnrByThreshold`
> * True negative rate (Specificity): `specificityByThreshold`
> * False positive rate: `fprByThreshold`
> Alternatives
> ---
> The `createCurve` method is marked private. If it were marked public, and the trait BinaryClassificationMetricComputer were also marked public, then it would be easy to define new computers to get whatever the user wanted.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org