Predictor Importance

Typically, you will want to focus your modeling efforts on the predictor fields that matter most and consider dropping or ignoring those that matter least. The predictor importance chart helps you do this by indicating the relative importance of each predictor in estimating the model. Since the values are relative, the sum of the values for all predictors on the display is 1.0. Predictor importance does not relate to model accuracy. It just relates to the importance of each predictor in making a prediction, not whether or not the prediction is accurate.

Predictor importance is available for models that produce an appropriate statistical measure of importance, including neural networks, decision trees (C&R Tree, C5.0, CHAID, and QUEST), Bayesian networks, discriminant, SVM, and SLRM models, linear and logistic regression, generalized linear, and nearest neighbor (KNN) models. For most of these models, predictor importance can be enabled on the Analyze tab in the modeling node. See the topic Modeling Node Analyze Options for more information. For KNN models, see Neighbors.

Note: Predictor importance is not supported for split models. Predictor importance settings are ignored when building split models. See the topic Building Split Models for more information.

Calculating predictor importance may take significantly longer than model building, particularly when using large datasets. It takes longer to calculate for SVM and logistic regression than for other models, and is disabled for these models by default. If using a dataset with a large number of predictors, initial screening using a Feature Selection node may give faster results (see below).

• Predictor importance is calculated from the test partition, if available. Otherwise the training data is used.
• For SLRM models, predictor importance is available but is computed by the SLRM algorithm. See the topic SLRM Model Nuggets for more information.
• You can use IBM® SPSS® Modeler's graph tools to interact, edit, and save the graph.
• Optionally, you can generate a Filter node based on the information in the predictor importance chart. See the topic Filtering Variables Based on Importance for more information.

Predictor Importance and Feature Selection

The predictor importance chart displayed in a model nugget may seem to give results similar to the Feature Selection node in some cases. While feature selection ranks each input field based on the strength of its relationship to the specified target, independent of other inputs, the predictor importance chart indicates the relative importance of each input for this particular model. Thus feature selection will be more conservative in screening inputs. For example, if job title and job category are both strongly related to salary, then feature selection would indicate that both are important. But in modeling, interactions and correlations are also taken into consideration. Thus you might find that only one of two inputs is used if both duplicate much of the same information. In practice, feature selection is most useful for preliminary screening, particularly when dealing with large datasets with large numbers of variables, and predictor importance is more useful in fine-tuning the model.

Predictor Importance Differences Between Single Models and Automated Modeling Nodes

Depending on whether you are creating a single model from an individual node, or using an automated modelling node to produce results, you may see slight differences in the predictor importance. Such differences in implementation are due to some engineering restrictions.

For example, with single classifers such as CHAID the calculation applies a stopping rule and uses probability values when computing importance values. In contrast, the Auto Classifer does not use a stopping rule and uses predicted labels directly in the calculation. These differences can mean that if you produce a single model using Auto Classifier, the importance value can be considered as a rough estimation, compared with that computed for a single classifier. To obtain the most accurate predictor importance values we suggest using a single node instead of the automated modelling nodes.