Accuracy issues in Logistic Regression with L1 penalty

[JohnZed]

Observed only in multi-class case.

[JohnZed]

Hi, @dantegd - were you able to root cause this one?

[cjnolet]

@dantegd, did @tfeher's recent fix also take care of this one?

[tfeher]

The recent changes introduces by PRs #2543 and #2494 do fix issues with logistic regression using L1 penalty. I would expect that any converged model has correct accuracy (it is a different issue that from the Python side it is not necessarily obvious whether we are converged #2546). I believe we can close this issue.

[tfeher]

Artem's recent fix for the QN solver stopping condition (#3766, #3774, #3777) also affects convergence for the case with L1 regularization.

@dantegd do we have any concrete example where the accuracy still differs? If yes, then @achirkin could have a look. Otherwise we could close this issue.

[beckernick]

Closing given the above comments. Please feel free to re-open if a scenario is found with accuracy issues.