"A new technique for separating the wheat
from the chaff in data."
Byron Roe, University of Michigan
Abstract:
In examining data, we are often faced with the problem of
classifying it into categories, e.g. "signal" and "background".
A new technique for classifying data is presented: boosted
decision trees. This technique is compared with the older
artificial neural net (ANN) technique using the mini-BooNE
neutrino oscillation experiment as a test bed. The new
technique does a better job of classification of data.
Furthermore, it requires less tuning of parameters, generally
seems more robust and can handle many more classification
variables than an ANN. We expect this technique to have
wide applications within physics, often supplanting ANN
techniques currently in use.
Reference:
"Boosted Decision Trees, an Alternative to Artificial Neural Networks"
(ps,
pdf)
eprint: physics/0408124