gamma="auto" in SVC

[amueller]

It's a bit unfortunate that we just changed the parametrization of SVC's gamma (from 0 to "auto"), but I just realized that using 1 / (n_features * X.std()) would be a much much better default than the current 1 / n_features.

Not sure there's a good way to change it though.

I wonder whether we should at some point do a breaking release like matplotlib did, but that might lead to more issues than it solves... (I kind of regret that we didn't change the cv=3 to cv=5 default when doing the model_selection refactoring).

[GaelVaroquaux]

Agreed on every sentence of this issue (including the last one :( ).

[jnothman]

'auto-scale'

…

On 16 February 2017 at 02:18, Gael Varoquaux ***@***.***> wrote: Agreed on every sentence of this issue (including the last one :( ). — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#8361 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6-tRiq9IaVfwCAxHZvKaiDNhvT7-ks5rcxc8gaJpZM4MB1or> .

[GaelVaroquaux]

'auto-scale'

Or maybe simple "scale". But what worries me most is yet another deprecation cycle. The danger is the instability that it brings to scikit-learn. But if, on average, it make things work really better...

[amueller]

I think this is worth it, so let's do scale.

[amueller]

For whoever fixes this, it needs a deprecation cycle

[neokt]

I'm currently working on this.

[gxyd]

We had a release 0.19 in which this change wasn't included (we still have gamma='auto' as default). Is this issue in-line to be ready to be present in the next release?