DOC Specify primal/dual formulation in LogisticRegression#26294
DOC Specify primal/dual formulation in LogisticRegression#26294lorentzenchr merged 9 commits intoscikit-learn:mainfrom
Conversation
|
Pinging some recent contributors to |
|
It would help to point to a formula (eg in the user guide) for the regularized one such that it is crystal clear. |
|
Thanks, @lorentzenchr. I found a regularized formulation at https://scikit-learn.org/stable/modules/linear_model.html#binary-case. How would I link this and refer to an equation? |
|
@glemaitre @lucyleeow Could you help out how to link to a formula in the user guide? |
|
Maybe Otherwise, potentially .. math::
:name: <eq name>
|
|
Hmm it does look a bit odd now that this specific equation is the only one on the page that has a number next to it. Maybe .. math::
:name: <name>
I need to refer to this :ref:`eq title <name>` |
|
With all three approaches and the equation receives a number |
|
Thanks for testing it all! I would just go with We could think about working on adding labels to all equations in our docs WDYT @lorentzenchr ? |
|
@lucyleeow First of all, thanks for helping out. I would not add labels to all equations we have, but make more use of labels on a case by case where needed base. |
|
No worries, probably unnecessary. Just thought the numbering looked odd, especially as it isn't the first equation in the page. No matter. |
|
I've added the same text to |
|
@lucyleeow @lorentzenchr Is there anything still needed to do before we can merge this? |
lorentzenchr
left a comment
There was a problem hiding this comment.
LGTM.
@mlondschien Thanks for this doc improvement.
What does this implement/fix? Explain your changes.
It is ambiguous which of the two formulations for (Logistic) Ridge is dual / primal. As there is no duality gap, each formulation is the dual of the other. Someone coming from optimization would typically consider the constrained variant "primal", whereas someone from machine learning would typically consider the regularized variant "primal". This PR removes this ambiguity by explicitly referring to the variants as "constrained" and "regularized".