Machine mob

Abstract

For the purpose of evaluating long term performance of fraud detection algorithms, we are designing an evaluation method to yield the expected long term worst case performance. Machine Mob reuses the non-illicit transactions from the training set of a fraud detection model to iteratively generate the predicted progression of the licit transaction network. The synthetic networks is then used for adversarial learning between fraudulent agents against a copy of the fraud detection model. That is, Fraudsters commit fraud by adding edges to iterations of the synthetic network while the copy of the fraud detection model learns to distinguish edges from the fraudulent agents. After some iterations an average score over the last rounds are returned as the total score of the model. We believe this method could prove useful to compare multiple fraud detection algorithms.