Discussion and Conclusions” We have

Discussion and Conclusions” We have shown how to
use boosting to reduce error rates in classification and
prediction error in regression using classification trees
and regression trees, respectively. Of critical importance
is the pruning of the trees by using a separate set of
pruning data after the tree has been initially grown using
the training set. Pruning both increases the speed and
improves the generalization to samples as yet unseen.
By picking the appropriate number of trees in the
ensemble, one can trade off speed versus performance.
If the speed performance is not acceptable (but the error
rate or prediction error is), then one choice is a parallel
architecture. Generally, there are between forty and
seventy trees in an ensemble and therefore the speed
would increase by those factors (but only approximately
since not all trees are the same size). If error rate or
prediction error is the issue, then another possibility is to
build a single neural network. Generally a single neural
network is slower than these ensembles because trees
just implement simple IF statements, while neural
networks need to implement both multiplication and
some sigmoid function as the transfer functions.
However, these issues must be decided on a case-by-case
basis.