Rethinking suspicious account identification from a perspective of machine learning ranking

Abstract Telecom fraud has raised much concern due to its negative social and commercial influence, in which a scammer cheats a victim to transfer money to his/her account. To prevent the fraud, it is essential to detect suspicious accounts from transaction data provided by banks. Conventional works usually formulated the problem as a classification problem, and employ rule-based methods or extract features to train a classifier for fraud prediction. This approach raises three primary concerns. Firstly, labeling suspicious accounts often involves significant manual effort. Secondly, identifying suspicious accounts requires expert labeling, which means there is no objective standard. Thirdly, the classifier’s accuracy may decline over time because fraudulent behavior evolves, rendering the labeled data outdated. To tackle this challenge, rather than formulating it as a classification problem, this paper reframes the task of identifying suspicious accounts as a ranking learning problem. Based on this idea, this paper proposes a framework DMLRank ( D eep M achine L earning Rank ing) for suspicious account ranking. DMLRank is composed of a network for representation learning and a model for ranking. The network for representation learning learns the representation of a transaction via multi-tasking learning which learns the distance function between two transactions and an autoencoder for each transaction. The model for ranking can then use the learned distance function to rank accounts based on their suspicious levels. We evaluate the performance of our model and compare it with some state-of-the-art methods on a real-world transaction dataset from a commercial bank. The results show that our method outperforms than others.