The HPSPLIT Procedure

PRUNE Statement

PRUNE C45 </ confidence>;

PRUNE NONE;

PRUNE COSTCOMPLEXITY </ $\gamma$ >;

PRUNE by-metric </ until-metric operator value>;

PRUNE by-metric / <until-metric> MIN;

The PRUNE statement controls pruning. It has five different syntaxes: one for C4.5-style pruning, one for no pruning, one for cost-complexity pruning, one for pruning by using a specified metric and choosing the subtree based on the change in a specified metric, and one for pruning by using a specified metric and choosing the subtree based on the minimum of a specified metric.

The default decision tree pruning method is entropy, using the change-in-entropy subtree selection method. The following PRUNE statement example is equivalent to having no PRUNE statement for a nominal target:

prune entropy;

The default regression tree pruning method is ASE, using the change-in-ASE subtree selection method. The following PRUNE statement example is equivalent to having no PRUNE statement for an interval target:

prune ASE;

The preceding statements are also equivalent to the following statements, respectively:

prune entropy / entropy >= 1.0;
prune ASE / ASE >= 1.0;

You can specify the following pruning options:

C45 </ confidence>

requests C4.5-based pruning (Quinlan, 1993) based on the upper error rate from the binomial distribution (Wilson, 1927; Blyth and Still, 1983; Agresti and Coull, 1998) at the confidence limit. The default confidence is 0.25.

This option is available only for decision trees (nominal targets).

COSTCOMPLEXITY </ $\gamma$ >

requests cost-complexity pruning (Breiman et al., 1984). The optional argument, $\gamma$ , is the penalty for the size of the tree. The final per-leaf penalty value is $\gamma$ multiplied by the sum of squares error (SSE) of the largest tree and divided by the number of leaves in that tree.

NONE

turns off pruning.

by-metric < / until-metric operator value>

chooses a node to prune back to a leaf by the specified by-metric. Optionally, you can specify an until-metric, operator, and value to control pruning. If you do not specify these arguments, until-metric is set to the same metric as by-metric, operator is set to ">=," and value is set to 1.

You can specify the following value for by-metric for decision trees (nominal target) or for regression trees (interval target):

ASE: chooses the leaf that has the smallest change in the average square error (ASE).

You can specify the following values for by-metric only for decision trees (nominal target):

ENTROPY: chooses the leaf that has the smallest change in the entropy.
GINI: chooses the leaf that has the smallest change in the Gini statistic.
MISC: chooses the leaf that has the smallest change in the misclassification rate.

You can specify the following values for until-metric for decision trees (nominal target) or for regression trees (interval target):

ASE: stops pruning when the per-leaf change in average square error rate (ASE) is operator value times the per-leaf change in the ASE of pruning the whole initial tree to a leaf.
N: stops pruning when the number of leaves is operator value.

You can specify the following values for until-metric only for decision trees (nominal target):

ENTROPY: stops pruning when the per-leaf change in entropy is operator value times the per-leaf change in the entropy of pruning the whole initial tree to a leaf.
GINI: stops pruning when the per-leaf change in the Gini statistic is operator value times the per-leaf change in the Gini statistic of pruning the whole initial tree to a leaf.
MISC: stops pruning when the per-leaf change in misclassification rate is operator value times the per-leaf change in the misclassification rate of pruning the whole initial tree to a leaf.

You can specify any of the following values for operator for decision trees (nominal target) or for regression trees (interval target):

<=: less than or equal to
LE: less than or equal to
>=: greater than or equal to
GE: greater than or equal to
<: less than
LT: less than
>: greater than
GT: greater than
=: equal to
EQ: equal to

by-metric / <until-metric> MIN

chooses a node to prune back to a leaf by the specified by-metric and selects the subtree at the minimum of the specified until-metric as the optimal subtree. Optionally, you can specify an until-metric. If you do not specify an until-metric, until-metric is set to the same metric as by-metric.

You can specify the following value for by-metric for decision trees (nominal target) or for regression trees (interval target):

ASE: chooses the leaf that has the smallest change in the average square error (ASE).

You can specify any of the following values for by-metric for nominal targets only:

ENTROPY: chooses the leaf that has the smallest change in the entropy.
GINI: chooses the leaf that has the smallest change in the Gini statistic.
MISC: chooses the leaf that has the smallest change in the misclassification rate.

You can specify the following value for until-metric for decision trees (nominal target) or for regression trees (interval target):

ASE: selects the subtree with the minimum absolute square error (ASE) as the optimal subtree.

You can specify any of the following values for until-metric for nominal targets only:

ENTROPY: selects the subtree with the minimum entropy as the optimal subtree.
GINI: selects the subtree with the minimum Gini statistic as the optimal subtree.
MISC: selects the subtree with the minimum misclassification rate as the optimal subtree.