I hadn't thought of that. I'll run some tests...
-N
On 8/4/09 11:49 AM, Tobias Sing wrote:
>> Is the "probability of the true label" the best prediction to feed to
>> the ROCR package, or is it better to use the "decision.value"
>>
> Since AFAIK they are related by a monotonous transforma
> Is the "probability of the true label" the best prediction to feed to
> the ROCR package, or is it better to use the "decision.value"
Since AFAIK they are related by a monotonous transformation, both
approaches should lead to the same ROC curve, shouldn't they? (not
tested)
On Tue, Aug 4, 2009
Good point. I'm not sure how I missed that.
This does lead to an additional question:
Is the "probability of the true label" the best prediction to feed to
the ROCR package, or is it better to use the "decision.value"
Anybody have any experience on this one?
Thanks!
-N
On 8/4/09 3:28 AM, Ch
Hi,
you need the score value , have a look at ?svm.predict and in the ROCR
example.
traindata <- as.data.frame(matrix(runif(1000),ncol=10))
trainlabels <-
as.factor(sample(c("win","lose"),nrow(data),replace=T,prob=c(0.5,0.5)))
model <- svm(traindata,trainlabels, type="C-classification",
4 matches
Mail list logo