Grouping options for predictors
Grouping allows you to combine values or ranges based on the similarity of predictor behavior.
There are no absolute rules for grouping. The profile should be smooth, the differences in behavior should be meaningful, and the number of cases should be reliable. Prediction Studio uses sensible defaults, but some experimentation is usually worthwhile for producing more predictive power and reliability. The predicted behavior should not to be monotonic, too flat, or jagged without a good reason.
If the profile is too flat, try increasing the level of detail (granularity), setting a maximum probability of error, and decreasing the minimum size (the level of evidence required for the behavior of a bin to be judged as representative). The number of bins should increase and the profile should become more varied.
If the profile is too jagged, try decreasing the level of detail and increasing the minimum size. The number of score bands should decrease and the profile should become smoother.
- Auto grouping option for predictors
If you want to set the best grouping options automatically, you can select the Auto grouping option. When this option is enabled, Prediction Studio attempts to automatically determine the best granularity and the minimum size of cases falling into each bin. When you select auto grouping, all other grouping parameters are disabled.
- Managing grouping options for numeric predictors manually
Combine values or ranges based on the similarity of numeric predictor behavior.
- Managing grouping options for symbolic predictors manually
Combine values or ranges based on the similarity of symbolic predictor behavior.
Previous topic Binning symbolic predictors Next topic Auto grouping option for predictors