|
what is the most important/interesting/best-performing recent research in boosting and ensemble methods? AdaBoost is fantastic, but very old - there must be something by now that takes things up a notch! |
|
One method that comes to mind is BrownBoost, which ignores data that is misclassified in several successive iterations. This is intended to limit the affect of noisy data. There's also gradient boosting, which works with decision trees. My impression is that ensemble learning is good when you have next to no apriori knowledge of the data, linear classifiers only perform slightly better than random, and more sophisticated or nonlinear classifiers are not applicable. However, it can be useful for increasing robustness when dealing with random algorithms. Either way, Polika's "Ensemble-based systems in decision making" can get you started there. 1
Thanks. Just found that there is an entire course on "Ensemble-based decision making": http://www.cs.gmu.edu/~carlotta/teaching//CS-795-s09/info.html
(Jul 01 '11 at 02:11)
Amit Kumar
|
|
I recommend you the "Ensemble Methods in Data Mining" , by Giovanni Seni and John F. Elder. it has just 126 pages but is a really good survey on ensembling. Both classic and modern methods are discussed and compared to each other. Structure of the book and the classification of different methods of ensembling is very interesting. You may also download the pdf version from publisher official website. http://www.morganclaypool.com/doi/abs/10.2200/S00240ED1V01Y200912DMK002 |
