With decision trees its easy to overfit the data, especially in the case where there are lots of variables. However, if you want to grab a subset of variables that are more meaningful and explain most of the data, it would be nice if some metrics were available.
Any suggestions, papers, software, comments,... will be much appreciated.
Dear Mahabir,
I dont have much idea about the decision tree, but recently I came across a paper in which the author used 1 Standard Error (known as 1-SE rule) above the mean cross-validate relative error to determine the size of the tree. May be you can get some clues in the paper.
De’ath, G., 2002. Multivariate regression trees: A new technique for modeling species-environment relationships. Ecology 83, 1105–1117. doi:10.1890/0012-9658(2002)083[1105:MRTANT]2.0.CO;2
Cheers,
My use of decision trees has been not to develop some huge network of complex interactions, but more confined to a specific problem needing a better understanding of the major factors, options or alternatives, and their display their relative effects before making a decision. However, I see it may also be used to develop complex, unwieldy networks of information and interactions, almost like brainstorming, but sometimes well beyond in complexity, like sending someone to the moon. I found this and it includes some advice on pruning. Sometimes it makes sense to be a lumper, and sometimes the problem may require splitting and great study of the fine detail. Maybe this could help.
http://www.ise.bgu.ac.il/faculty/liorr/hbchap9.pdf
Hi Ron. I have used an excellent, free, pdf, online from Stanford called "An Introduction to Statistical Learning with an Introduction in R " that might help. Chp. 8, p. 309 " the use of a tuning parameter and cross-validation or validation could help determine the possible threshold....then using the tuning parameter from cross-validation could generate a viable sub-tree". The free R code starts on p. 323. v/r Dave.
- Badges
- Science topic
- Similar topics
- Geoscience
- Geoinformatics