Practical Bias Variance Decomposition

Authors:
Remco R. Bouckaert
Affiliations:
Computer Science Department, University of Waikato, New Zealand
Venue:
AI '08 Proceedings of the 21st Australasian Joint Conference on Artificial Intelligence: Advances in Artificial Intelligence
Year:
2008

Citing 8
Cited 2

Data mining: practical machine learning tools and techniques with Java implementations

Data mining: practical machine learning tools and techniques with Java implementations
MultiBoosting: A Technique for Combining Boosting and Wagging

Machine Learning
An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants

Machine Learning
Variance and Bias for General Loss Functions

Machine Learning
Database Mining: A Performance Perspective

IEEE Transactions on Knowledge and Data Engineering
A Unifeid Bias-Variance Decomposition and its Applications

ICML '00 Proceedings of the Seventeenth International Conference on Machine Learning
Bias-Variance Analysis and Ensembles of SVM

MCS '02 Proceedings of the Third International Workshop on Multiple Classifier Systems
Decision Tree Instability and Active Learning

ECML '07 Proceedings of the 18th European conference on Machine Learning

On Feature Selection, Bias-Variance, and Bagging

ECML PKDD '09 Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases: Part II
A MapReduce-based distributed SVM ensemble for scalable image classification and annotation

Computers & Mathematics with Applications

Quantified Score

Hi-index	0.00

Visualization

Abstract

Bias variance decomposition for classifiers is a useful tool in understanding classifier behavior. Unfortunately, the literature does not provide consistent guidelines on how to apply a bias variance decomposition. This paper examines the various parameters and variants of empirical bias variance decompositions through an extensive simulation study. Based on this study, we recommend to use ten fold cross validation as sampling method and take 100 samples within each fold with a test set size of at least 2000. Only if the learning algorithm is stable, fewer samples, a smaller test set size or lower number of folds may be justified.