Generalized unified decomposition of ensemble loss

Authors:
Remco R. Bouckaert;Michael Goebel;Pat Riddle
Affiliations:
Xtal Mountain Information Technology, Auckland;Department of Computer Science, University of Auckland, New Zealand;Department of Computer Science, University of Auckland, New Zealand
Venue:
AI'06 Proceedings of the 19th Australian joint conference on Artificial Intelligence: advances in Artificial Intelligence
Year:
2006

Citing 4
Cited 1

Bagging predictors

Machine Learning
Improved Boosting Algorithms Using Confidence-rated Predictions

Machine Learning - The Eleventh Annual Conference on computational Learning Theory
An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants

Machine Learning
A Unified Decomposition of Ensemble Loss for Predicting Ensemble Performance

ICML '02 Proceedings of the Nineteenth International Conference on Machine Learning

A Hierarchical Face Recognition Algorithm

ACML '09 Proceedings of the 1st Asian Conference on Machine Learning: Advances in Machine Learning

Quantified Score

Hi-index	0.00

Visualization

Abstract

Goebel et al. [4] presented a unified decomposition of ensemble loss for explaining ensemble performance. They considered democratic voting schemes with uniform weights, where the various base classifiers each can vote for a single class once only. In this article, we generalize their decomposition to cover weighted, probabilistic voting schemes and non-uniform (progressive) voting schemes. Empirical results suggest that democratic voting schemes can be outperformed by probabilistic and progressive voting schemes. This makes the generalization worth exploring and we show how to use the generalization to analyze ensemble loss.