On 03/28/2012 06:18 PM, Lars Buitinck wrote:
> Hi all,
>
> I just suggested [1] adding a decision_function to @robertlayton's new
> NearestCentroid classifier and tried to implement one, but I got stuck
> on the binary case. In the multiclass case, the decision function is
> simple: it would be
>
>      D = -pairwise_distances(X, self.centroids_, metric=self.metric)
>
> which has shape (n_samples, n_classes), so that classification can be
> performed by D.argmax(axis=1).
>
> But in the binary case, linear models return an (n_samples, 1)-shaped
> array and I was wondering if NearestCentroids should do the same? We
> could do that by subtracting D[:, 1] - D[:, 0], but doing so we seem
> to lose interesting information.
>
> What to do?
>    
How about something like

(D[:, 1] - D[:, 0]) / (D[:, 1] + D[:, 0])



Would that solve the problem?

How would you formulate the information as uncertainty?

------------------------------------------------------------------------------
This SF email is sponsosred by:
Try Windows Azure free for 90 days Click Here 
http://p.sf.net/sfu/sfd2d-msazure
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to