Training

When you use either a neural net or a decision tree to create a predictive model for a symbolic field, the model automatically assigns a confidence value. This value reflects the level of confidence that the model has in its predictions. Unfortunately, the confidence value makes no distinction between the categories of the outcome field. For example, if you model a flag field with the values “stay” and “leave,” the confidence values for each category can vary from 0 to 1. However, a high degree of confidence does not help you determine whether a customer will actually stay or leave. Instead, the confidence value merely indicates the degree of confidence that you have in a specific prediction.

It is easy to create a scoring model that modifies the confidence values: a high value indicates customers who are likely to leave and a low value identifies those customers who are most likely to stay. To assign higher scores to those customers that the model predicts will leave, simply divide the model confidence value by 2 and add the result to 0.5. To calculate lower scores for the customers that the model predicts will stay, divide the model confidence by 2 and subtract it from 0.5.

Here is an example of the Derive node transforming confidence values from a neural net model into scores:

Figure 1.
Click to enlarge.

A histogram of the new field, NN_SCORE, now shows a clear distinction in the confidence score between those customers that the model predicts will leave and those that are predicted to stay:

Figure 2.
Click to enlarge.

With the new field, you can score a database so that customers can easily be selected for a marketing campaign based on their propensity to leave, as defined by their value on NN_SCORE.