how it is possible that neural network which is used for recontruction of input values can be used as a classifier for fraud transcations
so answer to this curiousity is very simple as we know autoencoder is useful for reconstruciton of values but if I train it on non-fraudulent transaction then it will be able to contruct non-fraudlent only so if I pass fraudulent transaction with non-fraudulent one then mse will be high for fraud transaction one why because it's weight are made on the basis of non-fraudulent transaction then at last I will decide a perfect threshold for classify fraudulent vs non-fraudulent transaction