16 Bad model vs Bad model
The model until now it’s working very well but it is because it clearly overfitting.
I then tried to create models with lower or same units as the input but they couldn’t recreate the original data well.
It was necessary to decide if I should use the overfitting model (Model A) or the not accurate one (Model B)
In general it is better to use neither of them, obviously, but in our case which is the worse model?
The issue with the model A is that it is probably only capable of reconstructing data very similar to the training data.
The models B on the other hand don’t encode and decode the more uncommon event at all. They are treated as not existent.
The objective of the MedGAN (and AE-WGAN-GP) is to generate synthetic data similar to the original one. With model A the data will be very similar almost identical to the original while model B won’t create uncommon variables as they will always be 0.
This crucial in our case since in epidemiological studies we are often more interested in rare events since there is less information regarding them in the litereature.
Taking into account the above mentioned reasons I decided to use model A, so the overfitting one.