OneVsOneClassifier decision function shape non-standard #8049
Comments
|
well, you have the option to use a FutureWarning to help users anticipate
the change
…On 14 December 2016 at 02:56, Andreas Mueller ***@***.***> wrote:
For binary tasks, OvO has a shape of (n_samples, 2) which pretty much
violates our standards.
I'm not sure what the best solution is apart from just breaking it as a
bug-fix.
We could add a parameter and deprecate it and then remove the parameter if
we really want.
The OVR classifier also has a potential issue where the decision_function
for binary is (n_samples, 1) instead of (n_samples,). That also seems
non-standard. I'm not sure we have other multi-output classifiers with a
decision function, so I'm not entirely certain what the standard should be.
All the multi-label ones do (n_samples,) according to the tests.
Found via #8022 <#8022>.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#8049>, or mute the
thread
<https://github.com/notifications/unsubscribe-auth/AAEz63bs1PaG8mEY2d7mqxXZU_q7NSPrks5rHsA6gaJpZM4LL5WR>
.
|
|
yeah, I guess so. I don't like those because there is no easy way to silence them except filtering them. |
|
Note that, actually, this OvO decision_function is enshrined in the codebase in SVC which has a parameter |
|
@jnothman If I read this correctly, then I disagree. In SVC The decision function of |
|
Okay. Sorry for making claims without giving them full attention! |
|
For OvR classifiers, because of the following in return np.array([est.decision_function(X).ravel()
for est in self.estimators_]).TI think the only option is to add a separate binary check for consistency in the shape. |
|
I'm not sure what you mean. |
|
As @amueller mentioned that we are getting the shape as |
|
Well yeah fixing it is obvious, and I did that in my PR (for now, will
probably revert unless we can figure something out). The problem is fixing
it in a backwards compatible way.
Sent from phone. Please excuse spelling and brevity.
…On Dec 19, 2016 5:19 AM, "Aman Dalmia" ***@***.***> wrote:
As @amueller <https://github.com/amueller> mentioned that we are getting
the shape as (n_samples,1) instead of the standard (n_samples,) that we
have as standard for multi-label classifiers, I wanted to add that because
of the above line, even if len(self.estimators_)==1 the shape will always
be (n_samples,1) for the binary case. So, the only workaround that seems
plausible here is to add a separate check.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#8049 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAbcFmroFLbXiFNhrH24xi6VMFTb_0Wbks5rJlongaJpZM4LL5WR>
.
|
|
@amueller Yes, I get it now. I would like to work on the issue, but am not really sure as to where should I start. Could you please give me a head start? |
|
Hello @amueller, Is it feasible to add a parameter (say standard_binary_output) to decision_function() which can be either True or False and will be False by default.So, now if task is binary we can use a warning that output of shape (n_classes,2 or 1) is not standard and will be of shape (n_classes,) in future version.When passed parameter is True than we can return desired standard output through a binary check in decision_function. |
|
We need to handle the case where decision_function is being called by
library code, not user code, and hence requires a standard API for
decision_function.
…On 6 January 2017 at 22:19, akshay0724 ***@***.***> wrote:
Hello ***@***.*** <https://github.com/amueller>*, Is it feasible to add a
parameter (say standard_binary_output) to decision_function() which can be
either True or False and will be False by default.So, now if task is binary
we can use a warning that output of shape (n_classes,2 or 1) is not
standard and will be of shape (n_classes,) in future version.When passed
parameter is True than we can return desired standard output through a
binary check in decision_function.
Please correct me if I'm wrong.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#8049 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAEz6wefttUeMwhgZI7Zw81C9N6YKW31ks5rPiNXgaJpZM4LL5WR>
.
|
|
@lucyleeow I don't feel like I can close... this comment makes me think that something is missing... @amueller do you mind clarifying? This will probably be useful also for contributors willing to address this issue. Thanks! |
|
Ah good point! I missed that. Though if we haven't been backwards compatible since 2016, changing it now will make it not backwards compatible 😅 |
For binary tasks, OvO has a shape of
(n_samples, 2)which pretty much violates our standards.I'm not sure what the best solution is apart from just breaking it as a bug-fix.
We could add a parameter and deprecate it and then remove the parameter if we really want.
The OVR classifier also has a potential issue where the decision_function for binary is
(n_samples, 1)instead of(n_samples,). That also seems non-standard. I'm not sure we have other multi-output classifiers with a decision function, so I'm not entirely certain what the standard should be. All the multi-label ones do(n_samples,)according to the tests.Found via #8022.
The text was updated successfully, but these errors were encountered: