Calculating AIC for tree-based models? (Akaike's Information Criterion)

[Mike90]

Are there settings or existing SAS code to get the tree-based models to compute Akaike's Information Criterion (AIC)?

The target is the number of days, and is an exponential distribution.  Average Squared Error is not appropriate for comparing models that have a non-normal interval target.

I would like to compare the tree-based models to neural networks and GLMs in the Model Comparison node.

Tree based models: decision tree, random forest, gradient boosting.

[MikeStockstill · Accepted solution]

The Akaike's Information Criterion (AIC) (Akaike, 1973,1977) uses the log likelihood function for a model with k parameters to select models, choosing the model that maximizes 2(LL — k) or the model that minimizes –2(LL + k).  The tree-based models are non-parametric (there is no k), so there are no settings in the Enterprise Miner nodes that make this computation.

If you know of a formula that gives the computation that you want for the tree-based models, and you are seeking coding suggestions for that formula, please post the formula or a link to it.