To clarify the purpose of this PR, can you please add a new test for the multi-output multiclass classification problem?

https://scikit-learn.org/stable/modules/multiclass.html#multiclass-multioutput-classification

This can break backward compatibility, since the output now of shape (n_samples, n_classes, n_outputs).

I think the shape should not change for existing problem (e.g. fitting a ClassifierChain on a multi-label classification problem, which is equivalent to multioutput binary classification, should still behave as in scikit-learn 1.0, without having to adjust the existing tests).

@ogrisel Sure I will add a new test case for the multi-output multiclass classification problem?

This can break backward compatibility, since the output now of shape (n_samples, n_classes, n_outputs).

I think the shape should not change for existing problem (e.g. fitting a ClassifierChain on a multi-label classification problem, which is equivalent to multioutput binary classification, should still behave as in scikit-learn 1.0, without having to adjust the existing tests).

Yes, that can done. But in that case "ClassifierChain.decision_function" will produce two test cases one for multi-output binary-class classification of shape (n_samples, n_outputs) and the other for multi-output multiclass classification of shape (n_samples, n_classes, n_outputs)

Yes, that can done. But in that case "ClassifierChain.decision_function" will produce two test cases one for multi-output binary-class classification of shape (n_samples, n_outputs) and the other for multi-output multiclass classification of shape (n_samples, n_classes, n_outputs)

I replied concurrently here: #21942 (review)

@rana-akkumar would you still like to work on this? Thanks