We support the PassiveAggressive classifier and regressor, which simply are BaseSGD with hinge loss and a specific learning rate schedule (+ some regularization).

I'm wondering whether we actually have a compelling reason to keep them around? Are these estimators useful to anyone?

They are quite poorly documented, and the only examples they're used in are just general benchmark comparison examples.

I feel like if we can't give users an incentive to use an estimator (i.e. in what cases the estimator should perform well), then we have no reason to support it. The UG doesn't do that right now.

Also note that they are apparently sub-optimal #1327

I have no strong opinion, I'm mainly genuinely curious.

We could also just move them to e.g. scikit-learn-extras. But that would be a superficial removal since we would still need to support their specific learning rate schedule as deep as in _fast_sgd().

If we decide to keep them around I'll make a PR to document them because this is really bad right now.

ping maybe @ogrisel and @GaelVaroquaux since you have reviewed the original PR #1259