search

Tilakkumar / cleartk

Automatically exported from code.google.com/p/cleartk
0 stars 0 forks source link

OVASVMlight doesn't do anything to ensure that the models' output is comparable #73

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago 
SVMlight models, per se, don't produce comparable outputs. While for one model 
a higher output 
value can be considered to indicate higher "confidence" in the prediction, the 
same cannot be said 
when comparing values produced by different models.

Platt's algorithm for fitting a sigmoid to the model's output to estimate 
probabilities can be used to 
make multiple models' output more comparable. Other methods may exist. ClearTK 
should, by 
default, apply such a method to OVASVMlight models. This would improve overall 
performance and 
is closer to The Right Thing, theoretically.

Original issue reported on code.google.com by phwetz...@gmail.com on 17 Mar 2009 at 6:29

GoogleCodeExporter commented 9 years ago 
Added an implementation of the modified Platt algorithm by Lin, Lin and Weng in 
r1341. Also modified the 
OVASVMlight components to use it.

This still needs to be tested.

Original comment by phwetz...@gmail.com on 25 Sep 2009 at 12:32

GoogleCodeExporter commented 9 years ago 
Apparently an issue being "fixed" here means "changes have been made, to be 
verified". While I don't know that anyone's actually going to verify it, this 
change was made a while and I've seen no complaints, so I'm going to mark this 
issue "fixed".

Original comment by phwetz...@gmail.com on 26 Jun 2010 at 2:09