Skip to content

Probabilities calibration error & ROCR

2 messages · Roberto Perdisci, Tobias Sing

#
Hello,
  I'd need to compute the calibration error of posterior class
probabilities p(y|x) estimated by using rpart as classification tree.
Namely, I train rpart on a dataset D and then use predict(...
type="prob") to estimate p(y|x).

  I've found the possibility to do that in the ROCR package, but I
cannot find a link to a paper/book which explains the details of the
implemented algorithm. Do you know of any reference where I can find
the details of the algorithm that computes the calibration error
implemented in ROCR (apart from ROCR's source code)?  Is there any
other function/package I can use to compute the calibration error?

thank you,
regards,
Roberto
#
Roberto,
On 1/24/07, Roberto Perdisci <roberto.perdisci at gmail.com> wrote:
we use it as defined in Caruana & Niculescu-Mizil: "Data mining in
metric space: An empirical evaluation of supervised learning
performance criteria". Knowledge Discovery and Data Mining (KDD) 2006.
http://www.cs.cornell.edu/~caruana/perfs.rocai04.revised.rev1.ps

Also, have a look at ?performance.

HTH,
  Tobias