next up previous contents
Next: Wrapper Inducers Up: Inducers Previous: CN2


Some inducers are still being developed or have esoteric uses. We briefly mention them.

Accuracy estimator
is a wrapper inducer that runs a given inducer in ACC_INDUCER, estimates its accuracy using an accuracy estimation method, and returns that as the resulting accuracy. The AccEst utility provides a friendlier interface, but there are occasions where one wants to do two levels of accuracy estimation ( e.g. , cross-validation accuracy on holdout sets), where this inducer is very useful [].

always predicts unknown and thus gets 0% accuracy. It is mostly used internally, but can be used with the Inducer utility and DISP_CONFUSION_MAT set to true in order to view the distribution of the labels. The ``info'' utility is probably a better way of getting basic statistics about a data file.

is an inducer for building oblivious decision graphs top-down []. Cannot handle unknown values.

is a lazy decision tree algorithm, described in friedman-kohavi-yun-lazydt.

searches for an attribute ordering. Very researchy.

is a wrapper discretizer that searches for the best number of intervals for each attribute. Very slow.

is a wrapper discretizer that searches for the best weight for each attribute (from a uniform set of weights). Slow. Researchy. Not much improvement over feature subset selection.

Builds decision trees with categorizers you choose at the leaves. Researchy. Requires that inducers support copies, which very few do ( e.g. , naive-bayes).

Ronny Kohavi
Sun Oct 6 23:17:50 PDT 1996