topic Re: How to access Variable importance in neural network in EM? in SAS Data Science

How to access Variable importance in neural network in EM?

munitech4u — Wed, 02 Mar 2016 12:22:17 GMT

I created a neural network, but I can't find any option within enterprise miner, where I can access the variable importance, similar to decision tree. I wanted to get to a final list of reduced variables. Like in regression, we keep only the most significant variables for final model building.

Re: How to access Variable importance in neural network in EM?

WendyCzika — Wed, 02 Mar 2016 14:21:55 GMT

Neural networks don't directly give you variable importance, and all of the inputs are included in the model (no selection is done, so there is no reduced set). But see this post about a technique of using a decision tree as a surrogate model for calculating variable importance based on the neural network model that was fit: https://communities.sas.com/t5/SAS-Data-Mining/Interpreting-Neural-Network/m-p/250372/highlight/true#M3705

Re: How to access Variable importance in neural network in EM?

munitech4u — Thu, 03 Mar 2016 08:04:07 GMT

It does not allow to change the role of the variables in decision tree, from the output of neural network. That option is freezed

Re: How to access Variable importance in neural network in EM?

WendyCzika — Thu, 03 Mar 2016 14:56:57 GMT

You need to use the Metadata node following the Neural Network node to change the roles of the observed target to REJECTED and the predicted target (posterior probability for the event level if a nominal target) to TARGET. Then use a Decision Tree node after that.