[SPARK-8849] expose multiclass scores in LogisticRegression

• Type: New Feature
• Status: Closed
• Priority: Minor
• Resolution: Won't Fix
• Component/s: MLlib
• Labels: LogisticRegression mllib multiclass
• Assignee: Unassigned
• Reporter: Asher Krim
• Created: 06/Jul/15 21:10
• Updated: 08/Jul/15 19:31
• Resolved: 08/Jul/15 19:31

Description

In SPARK-2309, dbtsai implemented multinomial logistic regression for mllib. It would be great if users could get the score for each class directly, since they are already computed. It doesn't look like this would be a terribly complicated thing to do and I've sketched out what it could look like: https://gist.github.com/Krimit/f61b5e3be2b1e380d1ca

We could even move the score methods further up into GeneralizedLinearModel, since in the binary case we could still just return a 2-element vector of the scores (or margin in SVM's) for each of the 2 class, so algorithms lacking multiclass support at the moment can still support this.