Multiclass Classification
Ryan Rifkin
Description
We consider the problem of multiclass classification. We have seen
that regularization-based approaches (RLSC, SVM) perform very well on
binary classification tasks, and we hope to extend this benefit to the
multiclass scenario. We advance the hypothesis that a simple
"one-vs-all" scheme is an extremely effective approach to multiclass
classification. We review a number of other approaches, and present
experimental comparisons.
Slides
Slides for this lecture: PDF
Suggested Reading
Rifkin and Klautau, In Defense of One-Vs-All
Classification, submitted to Journal of Machine Learning Research.