positive class in logistic operator? Disagrees with AUC

bobdobbs

First off, I want to thank everyone here, ESPECIALLY SIMON, for all the great help with my last question.  I couldn't have gotten this far without you guys!!

My next question should be much simpler...

I am following a suggestion read in a paper to use a "conditional logit" function for some training.  This leads to two questions:

1) Is there a "conditional logit" function in RM, or is this the same as the W-Logistic operator??

2) I learned in the past how to indicate which class is positive and which is negative in RM.  There seems to be a discrepency with how RM handles it.  In my current process, I indicate the "not sick" class as negative by making sure it is the first label in the arff file.  The problem is that the performance classifier and the w-logistic operator seem to disagree on this:
      a) The AUC for the binomial performance operator correctly states "sick" as the positive class and draws a nice curve for it.
      b) The resulting model for the W-logistic operator indicates the "not-sick" class as the trained for class.

This seems wrong to me.  Shouldn't the two operators agree on what is the positive class??

Thanks!

land

Hi,
probably WEKA does interpret the values different. I don't think we can do much about it, but you could replace the WEKA learner with the RapidMiner Logistic Regression, which should be the same for binominal values.

Greetings,
  Sebastian