An experimental bias-variance analysis of SVM ensembles based on resampling techniques

摘要: Recently, bias-variance decomposition of error has been used as a tool to study the behavior learning algorithms and develop new ensemble methods well suited characteristics base learners. We propose procedures, based on Domingo's unified theory, evaluate quantitatively measure in ensembles machines. apply these compare single support vector machines (SVMs) SVMs resampling techniques, their relationships with cardinality training samples. In particular, we present an experimental analysis bagged random aggregated order verify theoretical variance reduction properties. The characterizes between bagging aggregating, explains reasons why built small subsamples data work large databases. Our also suggests directions for research improve classical bagging.