You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Hayri Volkan Agun (JIRA)" <ji...@apache.org> on 2016/07/07 08:40:10 UTC
[jira] [Commented] (SPARK-16098) Multiclass SVM Learning
[ https://issues.apache.org/jira/browse/SPARK-16098?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15365819#comment-15365819 ]
Hayri Volkan Agun commented on SPARK-16098:
-------------------------------------------
Hi Sean,
I will work on creating correct JIRA issues. The problem of SVM in spark is
class-imbalanced problem. This happens in large sample sizes. It can be
solved by either introducing a new optimization routine for multiclass case
or OneVsOne classifier (with voting). However OneVsOne creates n*(n-1)/2
number of binary classifier for n labels. So I think the best way to solve
imbalanced class problem in SVM for large datasets (~1000 sample for each
class and 50 classes) is writing a multiclass optimization routine.
--
Hayri Volkan Agun
PhD. Student - Anadolu University
> Multiclass SVM Learning
> -----------------------
>
> Key: SPARK-16098
> URL: https://issues.apache.org/jira/browse/SPARK-16098
> Project: Spark
> Issue Type: Improvement
> Components: ML, MLlib
> Affects Versions: 2.0.0
> Environment: Spark MLLib and ML 1.6.1
> Reporter: Hayri Volkan Agun
> Priority: Minor
> Original Estimate: 1,512h
> Remaining Estimate: 1,512h
>
> There exists a OneVsRest classifier for using all binary classification classifiers in multi-class classification. However for Linear SVM using OneVsRest may create an imbalanced dataset scenarios where SVM of Spark certainly fails. I verified this by creating LinearSVM classifier and implemented predictRaw method of ClassificationModel class. In all experiments the results came very poor in terms of F-Measure. The only explanation is SVM is very sensitive to imbalanced dataset, and naturally OneVsRest classifier creates an imbalanced dataset.
> For multi-class classification, linear SVM can be optimized by considering imbalanced datasets.
--
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