Adding support for sparse label matrix in coverage_error.

[elanmart]

Currently the implementation of sklearn.metrics.coverage_error expects dense indicator matrix for ground-truth labels (code)

Do you think it could be beneficial to extend it to support sparse matrices for y_true as well?

I could open a PR, as I needed this functionality in my personal project.

[jnothman]

I don't get why this would provide substantial benefit if y_score is dense. would you make that sparse too?

[elanmart]

I'd argue that it's a natural choice to keep y_true as a sparse matrix. And if a user chooses to do so, she can still use coverage_error without expensive calls to .toarray()

[jnothman]

Yes, it's a natural choice, but it provides no substantial memory benefit if y_score is still dense. PR welcome to support y_true being sparse.

…

On 24 June 2018 at 20:24, Marcin Elantkowski ***@***.***> wrote: I'd argue that it's a natural choice to keep y_true as a sparse matrix. And if a user chooses to do so, she can still use coverage_error without expensive calls to .toarray() — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#11348 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz62HFZ5yggm_e6jw4KZQIFT1vGkzaks5t_2j3gaJpZM4U08WN> .