Split Selection Methods for Classification Trees
Prof. Wei-Yin Loh
Statistics Department, UW-Madison
2:30pm, Friday May 5 in 2310 CS
The talk will review the greedy search method of split selection used
in the AID and CART algorithms and compare it with a more modern
approach based on linear statistical techniques such as t-tests and
discriminant analysis. The decision trees produced by the two
approaches tend to have comparable classification accuracy and tree
size, but the new approach enjoys vastly superior computational
speed. Further, the split points chosen by the new approach tend to
have smaller variation than those obtained by greedy search.