ODM uses the Predictor Variance algorithm to build Attribute Importance models. Predictor Variance estimates the variances of the predictor target combinations and the variance with respect to the other predictors. Predictor variance assesses the relative usefulness of each attribute for making predictions for rows in general, instead of making a prediction for any particular case.
The Predictor Variance Algorithm measures univariate correlation to the target, that is, an attribute considered as a one-predictor model of the target. If an AI model indicates that an attribute has negative importance, the attribute is not correlated with the target. It is still possible that such an attribute might be useful in an interaction model, such as a decision tree. If you wish to reduce the number of features, an attribute with negative importance is a prime candidate for exclusion. Models such as decision tree and ABN perform internal feature reduction, so other than for performance reasons, there is no need to exclude such an attribute with those algorithms.
Copyright © 2005, Oracle. All rights reserved.