Posted by SUN on November 19, 19103 at 06:44:34:
I am working on a project of finding objects of interest from an unconstraint enviroment. Therefore I have aboundant data for one class and very limited number of data for another class. Moreover, the risks for two classes are highly unbalanced. I have implemented an algorithm based on Adaboosting. The overall test errors do decreased with the training error. However, the test error for the class with less data (object) is increased (very bad!). So my question is: how to implement boosting on unbalanced data with unbalanced risk? How to include the risk term into the boosting algorithm? Are any papers available to deal with these problem? Many thanks.