What? Can you explain the mechanism than a NN can “extrapolate” an invoice where a tree model couldn’t? This is all just how the modeler builds the features.

Also all models are a “mean of the subgroup of the data.” The prediction is by definition the conditional mean as a function of the input values.