Jump to content

Talk:Multiclass classification

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

Multi-label is not multi-class

[edit]

The current version of the page mixes two different concepts: multi-label classification and multi-class classification. The two are not exclusive but at the same time are NOT equivalent. In multi-class classification, each input is assigned a single output class out of multiple possible output classes. This is in contrast to the simpler problem of binary classification where just two output classes are allowed. On the other hand, multi-label classification does not assign just a single output label but rather a subset of output labels to each input. The article was originally on multi-class and there are links pointing to it as such, so imho, the multi-label part should be (re)moved. Tomash (talk) 17:24, 13 September 2011 (UTC)[reply]

I agree. Multi-label classification should be put somewhere else, possible even a new topic. Some further info is here Dcorney (talk) 12:49, 20 September 2011 (UTC)[reply]
Rewrote the article, see below. Qwertyus (talk) 11:34, 27 October 2011 (UTC)[reply]

Plagiarism

[edit]

I entirely rewrote this article, removing most of the information contained here. Not only was the article previously wrong in conflating multiclass and multi-label learning, it also seemed to have been copy-pasted of the PDF linked above. (Which, btw., is one CiteSeer, but looks to me like an unreviewed manuscript, so I haven't cited it or covered its use of "multiclass" to refer to what is commonly known as "multi-label".) Qwertyus (talk) 11:34, 27 October 2011 (UTC)[reply]

"Extension from binary" section questionable

[edit]

The "Extension from binary" section is very debatable. Not only are the qualitative statements ("Decision tree learning is a powerful classification technique", "Naive Bayes is a successful classified") too general and not required, it is unclear why k-nearest neighbors, Naive Bayes, Decision trees and the likes are even mentioned here, as they don't really need to be "extended" to handle non-binary.

Also for SVM, which indeed would need to be extended, a generic "extensions have been proposed to handle the multiclass classification case as well. In these extensions, additional parameters and constraints are added to the optimization problem to handle the separation of the different classes." without any further details, references or links to other pages doesn't provide any help.

MagRudolfMayer (talk) 17:23, 1 May 2020 (UTC)[reply]

Yes: this article is actually about taking binary classifiers and using them for the multi-class case, and should therefore (1) make that clear in the title, (2) avoid any mention of methods that don't even need extension (as noted above). A page with the title "Multiclass classification" should primarily concern itself with the basics of that subject, not with extensions from binary classifiers.

Mfrean (talk) 00:06, 5 May 2020 (UTC)[reply]