Direct Optimization of Margins Improves Generalization in Combined Classifiers

Part of Advances in Neural Information Processing Systems 11 (NIPS 1998)

Authors

Llew Mason, Peter Bartlett, Jonathan Baxter

Abstract

Cumulative training margin dis(cid:173) tributions for AdaBoost versus our "Direct Optimization Of Margins" (DOOM) algorithm. The dark curve is AdaBoost, the light curve is DOOM. DOOM sacrifices significant training er(cid:173) ror for improved test error (hori(cid:173) zontal marks on margin= 0 line)_

-1

-0.8 -0.6 -0.4 -0.2 0

0.2 0.4 0.6 0.8