I am following this example to learn a bit about the use of auto-encoders in fraud detection. Now that I reached the end of the article, two questions rose in mind:
- Can we train the network in an unsupervised way, freeze the layers weights, add a sigmoid layer as an output then finally use this model as a binary classifier?
- Are auto-encoders considered as the state-of-the-art method in fraud detection? What other alternatives can be used in this case?