[
https://issues.apache.org/jira/browse/SPARK-8069?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14611244#comment-14611244
]
Joseph K. Bradley commented on SPARK-8069:
------------------------------------------
I like the idea of including it in an abstraction like ClassificationModel and
ProbabilisticClassificationModel, unless it is too difficult. If a developer
does not want to support thresholds/cutoffs (or wants to modify the API), the
developer does not have to use the abstraction.
The main difficulty I see is in trying to specify thresholds in a uniform way:
* Thresholding rawPrediction vs. probability: It would be easy to mimic the R
randomForest package for thresholding probabilities, for which we know which
values are in the range [0,1]. That won't work well for rawPrediction values,
which could be negative.
** We could initially only support thresholding for
ProbabilisticClassificationModel. I expect to modify trees & tree ensembles to
subclass ProbabilisticClassificationModel in release 1.5 (WIP).
** Do you have ideas for thresholding for rawPrediction?
* Binary vs. multiclass: It would be nice to think of a way to naturally
support binary, though it might mean modifying or deprecating HasThreshold.
Once we decide on a good way to specify thresholds, then perhaps the binary
case can be handled by providing a setter as in HasThreshold
({{setThreshold(value: Double)}}) but returning the generalized threshold in
the getter ({{Vector getThreshold}}).
> Add support for cutoff to RandomForestClassifier
> ------------------------------------------------
>
> Key: SPARK-8069
> URL: https://issues.apache.org/jira/browse/SPARK-8069
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Reporter: holdenk
> Priority: Minor
> Original Estimate: 240h
> Remaining Estimate: 240h
>
> Consider adding support for cutoffs similar to
> http://cran.r-project.org/web/packages/randomForest/randomForest.pdf
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
