• We used a mixture of fraud and non-fraud cases in our credit card fraud example. In this case, we are lucky enough to have the correct label for each case. So perhaps in this case, it makes more sense to do the reconstruction of the normal transactions only. Would the detection rate improve if we only used non-fraud cases? Run the same experiment but use only the non-fraud cases for the training set.
• Using the Credit Card dataset, use the reconstructions from the autoencoder as inputs for a classification model, pretty much in the same way you would use PCA. Does this improve the accuracy of the classification? Note that you can do this in this particular dataset, because you have information about the class of the transaction (fraud/non-fraud) which might not be available in other data.