模式识别技术11_Ensamble.pdf

Ensembles of Classifiers Ensembles of Classifiers 多分类器方法 多分类器方法 1 1 Ensembles of Classifiers Ensembles of Classifiers Why? Why? Goal: improve accuracy in supervised learning Goal: improve accuracy in supervised learning Discovery: ensembles are often much more Discovery: ensembles are often much more accurate than the individual classifiers that accurate than the individual classifiers that make them up. make them up. We have to …… We have to …… How? How? 2 2 1 Methods for Constructing Methods for Constructing Ensembles Ensembles Subsampling the Training Examples Subsampling the Training Examples Idea: train every classifier with a different Idea: train every classifier with a different subset of the training examples. subset of the training examples. 3 3 Bagging Predictors Bagging Predictors 4 4 2 Bagging Predictors Bagging Predictors Background Background There might be some good but unstable procedures. There might be some good but unstable procedures. How to push them towards optimality. How to push them towards optimality. Unstable: e.g. Performance of Predictors changes Unstable: e.g. Performance of Predict