Similar to what we can do in terms of variable importance in regression type routines where we determine the partial R2 of a given variable and then divide by the total R2 of the entire equeation.
The approach to assessing variable importance in regression only makes sense if the the variables are uncorrelated with one another since he regressor has a coefficient that reflects what it adds over and above the other regressors. In some cases, adding another variable to a model can make a variable already in the model seem 'more' important or 'less' important. As a result, any importance is only relative importance to a defined set of other predictors in a model. Scorecards are interpretable because they first bin all interval variables so that there can be no notion of 'correlation' among the inputs and the model is truly additive. Of course, importance also assumes the model reflects the true relationships so any importance measures are less meaningful the less they reflect the actual relationship.
In neural nets, I understand we use weight of a given variable that are indicated across all the nodes and then sum it up and then divide by the weights of all variables across all the nodes.
I am not aware of any broadly accepted approach to evaluating variable importance in a neural network which is essentially a complex nonlinear model. The relative importance of a variable at a given point might differ wildly depending on which set of coordinates you are using to define the 'neighborhood'. Overall importance is even more tricky since a variable important near one data point might be non-informative near another data point.
does anyone have SAS code to do this.
I don't know who investigated the approach you are describing but you might consider contacting the author who proposed it.
Hope this helps!
Doug
