self : object for consistency with other places in the package

yes, but I don't think it's correct. The base class for all estimators is BaseEstimator not object, so I think there is no point in indicating the object dtype.

ok

self : object

The initial version was correct for cases where there are no types, and it's a bit pointless to indicate object dtype for estimators which is also not accurate https://numpydoc.readthedocs.io/en/latest/format.html#parameters

cc @glemaitre

it's a bit pointless to indicate object dtype for estimators which is also not accurate

I agree. However, we have this way of documenting everywhere. I would prefer to keep this inaccurate convention until we do a find/replace regex in another PR. Would it not be easier to make the change?

Right, but my point is there is no issue with this line of docstring. It should pass the numpydoc validation, I think? And if so I would rather we didn't change it instead of introducing unnecessary code churn with a solution we know is not helpful / correct.

self : object is used elsewhere, but 27% of estimators still use the current conversion,

$ rg "^\s+self$"  | grep self | wc -l
38
$ rg "^\s+self\s*:$"  | grep self | wc -l
2
$ rg "^\s+self\s*:\s*object$"  | grep self | wc -l
103

@rth Thank you so much for reviewing my PR. The numpydoc validation requires a description and that's the reason a added a tautology. I'll be more than happy to make any changes according to your guidance.

I think the description is ok. The point of @rth is about self : object vs self alone on the current line, not the description below.

I think,

but need to check the rendering (ci/circleci: doc artifact CI job), currently it doesn't look ideal: https://142780-843222-gh.circle-artifacts.com/0/doc/modules/generated/sklearn.decomposition.LatentDirichletAllocation.html

It might raise an error if the article is not linked (somewhere we expect [2]_) in the docstring. If this is not the case, we could either link where it is meaningful, otherwise, we could as well remove the reference.