Hi There,
I got some results from using nnet on a two-class problem, and I'd like
to hear your comments to understand well about the algorithm. In the training
set, the ratio of class 1 to class 2 is about 23:77. I did a 5-fold cross
validation. The networks were trained twice, one with 'weights=1', one
with 'weights=ifelse(species=="class1", 77/33, 1)'(pointed out
by Prof. Brian Ripley).All other settings are same. The average Matthew
Correlation Coeffience for the one with weights=1 is 0.80, significantly larger
than that of the other, 0.74. So, it seems weighting the unbalanced samples does
not help performance on evaluations, which is against my initial thoughts. My
question would be, does that mean the training data is not unbalanced enough?
then how unbalanced is enough? Or it was totally just a signal event? Or it was
just some suboptimal results? Any references regarding this issue in particular?
Thanks!
Best regards,
Baoqiang Cao
