Machine Learning
A Unifeid Bias-Variance Decomposition and its Applications
ICML '00 Proceedings of the Seventeenth International Conference on Machine Learning
| Hi-index | 0.00 |
According to Domingos’ bias–variance decomposition framework, we study the bias–variance characteristics of the standard Multiple Criteria Linear Programming (MCLP) classification method. The experimental results show that, under Domingos’ bias–variance decomposition framework, bias is much bigger than variance, and boosting ensemble doesn’t behave better than bagging ensemble, and increasing training example can effectively reduce variance rather than bias. We conclude that MCLP intrinsically is a stable classification method, and that an appropriate ensemble method for MCLP rests with the characteristics of a specific data set. When data can be easily linearly separated, MCLP will have low bias and bagging can be employed to lessen variance. When data present complicated no-linear structure, MCLP will have high bias and boosting ensemble can be considered to reduce bias. But, when boosting is used, noises and over fitting should be considered.