07-05-2016 07:27 AM
I have plotted this months*tenure against the target event rate. Can neural network node in EM approximate such surface with sufficient hidden nodes?
07-05-2016 11:21 AM
Are you ultimately trying to predict a binary target (event/nonevent) using two interval inputs?
Judging from the plot, there is clearly an interaction between the two inputs (for high month values, event rate is high for low tenure values but drops for middling tenure values). I'd try the Neural node but also logistic regression with interaction terms and polynomials (i.e., x*x, y*y, x*y).
Did you know you can easily create surface plots in EM? After you've built your model, use the Graph Explore node in EM to create response surface plots using predicted probabilities in the Z axis. To do so, run the GE node then in the Results window choose "3D Plots". I've attached an example below.