In this paper, we introduce a novel machine learning approach for regression based on the idea of combining bagging and boosting that we call BagBoo. Our BagBoo model borrows its high accuracy potential from. Friedman's gradient boosting [2], and high efficiency and scalability through parallelism from Breiman's bagging [1]. We run empirical evaluations on large scale Web ranking data, and demonstrate that BagBoo is not only showing superior relevance than standalone bagging or boosting, but also outperforms most previously published results on these data sets. We also emphasize that BagBoo is intrinsically scalable and parallelizable, allowing us to train order of half a million trees on 200 nodes in 2 hours CPU time and beat all of the competitors in the Internet Mathematics relevance competition sponsored by Yandex and be one of the top algorithms in both tracks of Yahoo ICML-2010 challenge. We conclude the paper by stating that while impressive experimental evaluation results are presented here in the context of regression trees, the hybrid BagBoo model is applicable to other domains, such as classification, and base training models.
[1]
Tie-Yan Liu,et al.
Learning to rank for information retrieval
,
2009,
SIGIR.
[2]
Maksims Volkovs,et al.
BoltzRank: learning to maximize expected ranking gain
,
2009,
ICML '09.
[3]
Leo Breiman,et al.
Random Forests
,
2001,
Machine Learning.
[4]
J. Friedman.
Greedy function approximation: A gradient boosting machine.
,
2001
.
[5]
Tao Qin,et al.
LETOR: Benchmark Dataset for Research on Learning to Rank for Information Retrieval
,
2007
.
[6]
Jianjun Xie,et al.
A Combination of Boosting and Bagging for KDD Cup 2009 - Fast Scoring on a Large Database
,
2009,
KDD Cup.
[7]
Rich Caruana,et al.
Additive Groves of Regression Trees
,
2007,
ECML.