Empirical Comparison of Ensembles… Could There be Modeler’s Effect Also?

Popular Ensemble Methods – An Empirical Study – Journal of Artificial Intelligence Research 11 (1999) 169-198

Graphical version of Table 2:

23 Datasets Ensembling Error Rates - ANN first four - Next is DT

 

 

From Table 2:  “Test set error rates for the data sets using (1) a single neural network classifier; (2)an ensemble where each individual network is trained using the original training set and thus only differs from the other networks in the ensemble by its random
initial weights; (3) an ensemble where the networks are trained using randomly
re-sampled training sets (Bagging); an ensemble where the networks are trained
using weighted re-sampled training sets (Boosting) where the re-sampling is based
on the (4) Arcing method and (5) Ada method; (6) a single decision tree classifier;
(7) a Bagging ensemble of decision trees; and (8) Arcing and (9) Ada Boosting
ensembles of decision trees.” – as noted by the author.

Let us just look at key measures only in two graphs, first on for ensembling using ANN and the next one is ensembling using decision tree method.  In both the cases it looks like you can just trust boosting (and hence combined with bagging).  However the author interprets it as 95% CI basis there is no difference between boosting and non-boosting.

23 Datasets Ensembling Error Rates - Adaboost wins in ANN23 Datasets Ensembling Error Rates - Adaboost wins in Decision Tree

 

Leave a Reply

Your email address will not be published. Required fields are marked *