Talk:Multi-label classification

Label Combinations
Just someone with a causal interest in multi-label classification (i.e. definitely no expert), but assuming they are referring to the same approach, is the name `Label Powerset' not more common than Label Combinations? Certainly I have come across the former far more often in the literature (and indeed had never heard of Label Combinations before reading this article - relatively few hits on Google Scholar too). 129.215.59.52 (talk) 21:23, 25 February 2013 (UTC)
 * While Label Powerset is the more common name in the literature, Label Combination is used in the WEKA/MEKA constellation of software. As this is the reference implementation in many cases, it is often used.  However I have changed the article to your recommendation as I believe your analysis is correct.  Doctorambient (talk) 00:57, 7 April 2014 (UTC)

Clarification of intro paragraph needed
The intro paragraph perhaps makes sense to those already familiar with the distinctions it tries to make, but for those who need the explanation I don't think it's possible to discern what it's trying to say.

"multi-label classification and the strongly related problem of multi-output classification are variants of the classification problem where multiple target labels must be assigned to each instance. Multi-label classification should not be confused with multiclass classification, which is the problem of categorizing instances into more than two classes. Formally, multi-label learning can be phrased as the problem of finding a model that maps inputs x to binary vectors y, rather than scalar outputs as in the ordinary classification problem."


 * What is a "target label" as opposed to just a label?
 * Are "labels" related or unrelated to levels of a categorical (nominal) classification variable?"
 * What is the significance of the switch from "Multi-label classification" to "multi-label learning" in sentences 2 and 3? Are they synonyms, or is this a non-sequitur?
 * What is a "binary vector"?

Here's my best guess as to what this is trying to say:

"Multi-class classification" apparently refers to the case where, in addition to the x input variable(s), there is a single class variable, that variable is categorical (nominal), and has three or more levels (alternatives). Not sure why this needs a special name, it seems to me that the case of having only two possible levels is the oddity. But whatever.

"Multi-label classification" appears to refer to the case where there are multiple class variables. I suppose that if these several variables are represented as a "vector" (ie: array with only one index), AND each array entry can only have one of two values ("binary"), then this is the "binary vector" referred to? But why would this be called "multi-label" rather than just "multiple class variables"? And what happens if the class variables have >2 levels?

So, I'm not too confident that this series of suppositions is correct. In any case, the intro paragraph could perhaps be revised to make all this clear.

Gwideman (talk) 03:55, 30 December 2014 (UTC)


 * A belated thanks for your well-explained clarification request. I've rephrased the lead taking your points into account. Loraof (talk) 16:41, 17 August 2017 (UTC)

Ambiguity/Discrepancy between this article's mention of the Multiclass Classification article and that article itself
> This method of dividing the task into multiple binary tasks has something in common with the one-vs.-all (OvA) or one-vs.-rest (OvR) method for multiclass classification. Note though that it is not the same method, following are two main differences. In binary relevance we allow multiple labels to be predicted for one instance, not just the most likely one. In binary relevance we train one classifier for each label, not one classifier for each possible pair of values for the label in the case of OvA, however OvR is the same in this respect.

This implies that "one-vs.-all" and "one-vs.-rest" are different. While the multiclass classification article says they're the same.

Enamex (talk) 20:33, 23 May 2018 (UTC)