The problem with drawing an ROC curve with (default) IBk is that it

uses 1 nearest neighbour so the probability estimates you are getting

will be very discrete (mostly 1/0) and you won't get many points for

your curve. You could try using more neighbours and distance weighting.

(A totally different option is to use locally weighted learning in

conjunction with naive Bayes.)

Using the CostSensitiveClassifier in conjunction with IBk is pointless

if you are drawing an ROC curve. The structure of the classifier

doesn't change when the instance weights are changed and it's

sufficient to just change the threshold on the class probabilities to

get an ROC curve (i.e. what the Explorer does).

Cheers,

Eibe

On May 13, 2005, at 1:38 AM, Ewy Mathe wrote:

> Hello all,

> I have a one dimensional data to classify and am using the K-nearest

> neighbor (IBK) to classify my instances. Because the data is very

> unbalanced (one class may contain 20% of the data while the other has

> the remainder), I am using a cost sensitive matrix on top of that.

> However, it is very difficult for me to obtain "incremental" TP and TN

> values to draw an ROC curve, because the TP decreases by .4 or

> more...is it inappropriate to use a cost sensitive matrix with a KNN?

> Thanks for any feedback,

> Ewy

>

> _______________________________________________

> Wekalist mailing list

>

[hidden email]
>

https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist_______________________________________________

Wekalist mailing list

[hidden email]
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist