Posted by Gunnar Raetsch on May 21, 19102 at 22:25:29:
In Reply to: boosting for unbalanced data posted by Alexandre Pitti on April 12, 19102 at 23:23:01:
: i'm searching algorithm, publication for applying (improving...) boosting's method in the special case where data are unbalanced
There is little work on this topic. There was a paper at NIPS last year on face recognition that had very unbalanced data sets and had some fix for that.
But think about the problem: Boosting finds a separation of the training data. The (maximum margin) separation is indepentent of the class size. So there is only the bias one can adapt. However, if you consider soft margins then the situation is different. Then you might want to apply the ideas that seem to work well for SVMs (in particular Grace Whaba's paper on applying SVMs in non-standard situations). The same techniques that work thare also work for softmargin boosting.
I hope this helps,
Gunnar