2013/2/15 David Lambert <[email protected]>:
>
> On Feb 14, 2013, at 6:52 PM, Lars Buitinck <[email protected]> wrote:
>
>> 2013/2/15  <[email protected]>:
>>> how about softmax?
>>
>> The model is not intended for probability outputs. A predict_proba is
>> not required; just implement a decision_function instead.
>
> Fair enough, and already done.  For my own education though, would
> softmax do the trick, or would something else be suitable?  (I thought
> more or less any classifier could be persuaded to give probabilities
> somehow…)

Softmax with a scale / slope of 1. and and intercept of 0.5 would
yield uncalibrated (as in arbitrarily wrong) probabilities estimates.
Generating good class probabilities estimates would require
calibrating the sigmoid shape using the Platt scaling method or a non
parametric calibration based on Isotonic Regression for instance.

http://www.machinelearning.org/proceedings/icml2005/papers/079_GoodProbabilities_NiculescuMizilCaruana.pdf


--
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

------------------------------------------------------------------------------
Free Next-Gen Firewall Hardware Offer
Buy your Sophos next-gen firewall before the end March 2013 
and get the hardware for free! Learn more.
http://p.sf.net/sfu/sophos-d2d-feb
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to