Bagging [Breiman, 96] and its variants is one of the most popular methods in aggregating classifiers and regressors. Its original analysis assumes that the bootstraps are built from an unlimited, inde-pendent source of samples. In the real world this analysis fails because there is a limited number of training samples.
We analyze the effect of intersections between bootstraps to train different base predictors, which shows that the real-world bagging behaves very differently than its ideal analog [Breiman, 96]. Most importantly, we provide an alternative subsampling method called design-bagging based on a new construction of combinatorial designs. We prove that this is universally better than bagging. Our ana-lytical results are backed up by experiments on general classification and regression settings, and sig-nificantly improved all machine translation systems we used in the NIST-15 C-E competition.
Associate Professor, Institute of Computing Technology in Chinese Academy of Sciences
Jia Xu is an associate professor at ICT/CAS, after being an assistant professor in Tsinghua University and a senior researcher at DFKI lecturing at Saarland University in Germany. She worked at IBM Watson and MSR Redmond during her Ph.D. advised by Hermann Ney at RWTH-Aachen University. Her current research interests are in Machine Learning with a focus towards highly competitive machine transla-tion systems, where she led and participated in teams winning first place in WMT-11, TC-Star -05-07 and NIST-08. In NIST-15 she led one more team that won 4th place, which is the 1st among academic institutions.