hi to all,

is there a way to calculate a measure that represents the quality of a 
certain svm-class prediction. of course, the first look is at the training 
accuracy but how could one judge the quality/confidence of each individual 
classification in addition to that.

the straight forward idea would be, to look at the absolute value of the 
decision function with high values implying better predictions - because 
'better' point should be farer away from the hyperplane... right/wrong ?

however, we werde discussing this matter intensively and found no way to 
normalize this distance to scale to, lets say values from 0 to 100, because 
the decision value itself could potentially be somewhere close to 
infinity... right/wrong ?

can you think of any normalization ?

can you think of other approaches for the problem ?

thanks a lot

steve

[ comp.ai is moderated.  To submit, just post and be patient, or if ]
[ that fails mail your article to <[EMAIL PROTECTED]>, and ]
[ ask your news administrator to fix the problems with your system. ]
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to