Wednesday, March 21, 2012

PMML: One node in a decision tree containing two states of an attribute as the rule for spli

Hi,
is there a way to import a decision tree-model from pmml where a node contains two or more states of an attribute as the split-rule?

Example:

...
<Node recordCount="600">
<CompoundPredicate booleanOperator="or">
<SimplePredicate field="color" operator="equal" value="red" />
<SimplePredicate field="color" operator="equal" value="green" />
</CompoundPredicate>
<ScoreDistribution value="true" recordCount="200"/>
<ScoreDistribution value="false" recordCount="400"/>
</Node>
...

This node shoud contain all cases, whose color is red or green (The Microsoft DecisionTree-Algorithm would build a model with two steps like red/ not red and then green / not green). According to the DMG, this is valid PMML 2.1, but when trying to import the server complains about an unexpected value in the SimplePredicate-tag.

How can i import such a node in SqlServer 2005?

Thank you in advance for any help

Chris

No, the Microsoft_Decision_Trees algorithm does not support splits of this type.|||You could however, implement the prediction logic of a tree as a plug-in algorithm and have your implementation parse PMML bodies with OR conditions.|||Ok, maybe we will try that. Thanks for your answer!

No comments:

Post a Comment