Training

When you build predictive models in Clementine, you have to choose from hundreds of potential fields to use as predictors.

While there is no substitute for domain expertise when selecting fields, you can shorten the process considerably with the Feature Selection node. This node allows you to automatically screen out fields based on a variety of criteria, such as having too much missing data or too little variation. You can also rank order the variables based on level of importance. To do so, each predictor is tested against the target variable using the appropriate bivariate analysis to produce a probability value. If both the predictor and the target are continuous, a Pearson correlation is used. If they are both categorical, then the node performs a Chi-square test. The probability values are then turned into an importance measure by subtracting them from 1. An importance measure equal to 1 indicates that the two fields are highly related. 

To use the Feature Selection node:

• Add a Feature Selection node to the stream and connect it to the Type node. You can also attach it to your Source node if that is where you type your fields.
• Edit the Feature Selection node and click the Fields tab
• Click the “Use custom setting” button
• Select your Target field (Churn is selected in Figure 1) and all the fields you want to consider as potential predictors. In this example, we are selecting 127 fields.

Figure 1: You’ve selected “Churn” as your target field and chosen 127 input fields.
Click to enlarge.

The Model tab and Options tab can be used to refine the criteria used to screen variables and to define their importance. (In this example, however, we will use the default settings.) Before receiving the model’s output, you’ll need to:

• Execute the node by clicking on the Execute button
• Right-click on the Feature Selection-generated model and select “Browse”

Figure 2: The model selected 63 predictor fields and rejected seven.
Click to enlarge.

At the bottom of the output screen (see Figure 2), you’ll notice that seven of the predictors were rejected because of too much missing data or too little variation. The model selected 63 of the remaining fields as important, considerably reducing your task of field selection.

Now you can add the generated Feature Selection model to the stream and it will filter out the unimportant variables.