PARTIALDEPEND Procedure

The PARTIALDEPEND procedure computes partial dependence (PD) functions for machine learning models that are built using algorithms such as gradient boosting, forest, and neural networks and for many other types of models that are created using SAS software. You can compute partial dependence functions for any model that produces an analytic store or DATA step code.

For a given model, PD values represent the model’s average prediction for each value of a particular input variable. Thus, PD plots show how changes to the model’s input values affect the model’s predictions. For example, a PD plot can show whether predicted automobile price increases linearly with horsepower or whether the relationship is of another type, such as a step function, curvilinear, and so on. You can also use PD plots to show interactions between two model inputs and to evaluate competing machine learning models. PROC PARTIALDEPEND produces plots of partial dependence only if you specify just one analysis variable. You can plot the output from a two-way PD by using SAS graphics procedures like PROC SGPLOT.

PROC PARTIALDEPEND also produces individual conditional expectation (ICE) tables. Whereas a partial dependence function represents the average model prediction, an ICE table represents the variation of individual model predictions around the overall average. ICE curves therefore help you identify interactions between model variables as well as identify subgroups that diverge from the overall average.