How is calculated the Variable Importance in the Gradient Boosting algorithm? (Viya 3.5)

[AngeloRD]

Hello,

I'm trying to understand how Variable Importance is calculated in the PROC GRADBOOST(in SAS Studio) and the Gradient Boosting Node(in SAS Model Studio). It's clear to me that the VI metric is useful to determine which variables have more impact in my prediction, but my question is how this metric works.

I work in a field in which adequately justifying the metrics used is essential, and that includes understanding how the metric works and is calculated. But the documentation I find about it is vague. 

I hope someone here can help me. Thanks in advance.

[GuyTreepwood · Accepted solution]

According to the documentation, the variable importance statistic for PROC GRADBOOST is calculated using the Residual Sum of Squares Importance Method. You can find the details behind this method here: 

SAS Help Center: Residual Sum of Squares Importance Method

[AngeloRD]

Thank you very much!