Hi,
I have a set of weak predictors that I have been trying to boost, to
improve on simple voting.
On the first pass I used simple voting and got decent improvement from
51~52% accuracy for individual predictors to 55~57% accuracy for the
unanimous voters.
I was expecting to do better with jboost but actually got results that
are worse. Not by much, but still, I would like to understand what I
can do better.
I have tried AdaBoost, LogLossBoost and BrownBoost and messed around
with number of interations and various other options.
Initially, and fairly quickly, both insample and outofsample error
rates go to about 46% and more or less stay there. Eventually, I can
observe overfitting, whereby the in-sample error is reduced while out
of sample error increases with additional interations.
I was wondering if anybody might have any thoughts on what might be going on.
I can send sample files if somebody might be willing to play with the data.
Thanks,v
|
