How does it compare with NOCATS such as started in this pull request: #12866

From my understanding of NOCATS, the differences are:

• NOCATS approaches handle categorical variables in the predictor, ordinal classification adresses the response variable.
• NOCATS try to get rid of the ordinality of features, or allows it to be unknown, while ordinal classification assumes the user knows the order. In more detail: There are two classical ways to treat categorical features. WIth dummy-encoding, the split is one group vs rest (which requires a lot of nodes). With integer-encoding, an order of the categories is assumed at the splitpoint (because all observations below the cut-off are sent left, the rest sent right). The NOCATS approach outlined in https://github.com/jblackburne uses random splits to generate the grouping at the split (and hopes to find useful ones).
• NOCATS and similar efforts have a bigger scope in the sense that they seem to be relatively big changes in terms of line count, require compiled code etc.
• NOCATS implementations like NOCATS: Categorical splits for tree-based learners (ctnd.) #12866 are specific for one (base) learner like trees, while the learner proposed in this issue can turn any classifier into a classifier that respects the order of the y values. In that sense, the scope of this isuse is bigger than NOCATS.

I completely missed that it was referring to encoding the target and not the data. So this is not related to NOCATS.

So this is something that could be adequate for the sklearn.multiclass module in addition to OvO and OvR classifier.

Hello
Thanks for mentioning me I'm the writer of the article mentioned above.
I think it's important to be aware that there is a major flaws in the implementation that I wrote in that article: since each classifier are independent from each other, the method won’t give the true probability (it won’t sum up to 1)

this article by Christopher Coffee might help deal with the problem mentioned above
https://medium.com/@deathswithbenefits/heres-the-actual-recipe-ec9aa0915d40

I think it's important to be aware that there is a major flaws in the implementation that I wrote in that article: since each classifier are independent from each other, the method won’t give the true probability (it won’t sum up to 1)

Ok. But I don’t think that’s a big problem. We can just normalize the probability, no?

Also. here is some other related work:
https://github.com/leeprevost/OrdinalClassifier

I think it's important to be aware that there is a major flaws in the implementation that I wrote in that article: since each classifier are independent from each other, the method won’t give the true probability (it won’t sum up to 1)

Ok. But I don’t think that’s a big problem. We can just normalize the probability, no?

Also. here is some other related work: https://github.com/leeprevost/OrdinalClassifier

I've checked the link that you share, I think it has a better implementation than the one that I provide in the medium post

Is there still interest in including this feature?

Figured I'd mention it, @FBruzzesi made an implementation of a OrdinalClassifier that's a meta estimator in scikit-lego. Somewhat experimental feature (most of the library is that) but could also be relevant to folks reading this.