I'm dealing with a low event rate problem (e.g. credit card fraud). I've balanced my data with SMOTE, and ran a neural net model (cross validated with recall as the measure).
However my precision (as can be expected) is very low.
I want to use the positive predictions (TP + FP) from my model as an input to another model - a subset of the original data. Creating a sort of "pipeline". Practically speaking, If a credit card company were to use my model, all the positive predictions will be handled by a human who would than decide whether or not it's a fraud, so why not automate?
Just like model stacking, but the stacked model gets a subset of the data and not the previous models predictions.
I'm not sure if that's OK. Any thoughts?