Abstract: Over the past few years,telecommunication fraud has caused enormous economic losses for bank users.Exis-ting detection methods firstly extract statistical features,such as RFM (Recency,Frequency,Monetary Value) of user transactions,and then use supervised learning algorithms to detect fraud transactions or fraud accounts through training classifiers.However,the RFM features don’t make use of the network structure of the transaction network.This paper designed a pairwise markov random field to capture the characteristics of the network structure in telecommunication fraud.Then,it exploited a linear loopy belief propagation algorithm to estimate the posterior probability distribution and predict the label of an account.Finally,it compared the proposed method with CIA and SybilRank on both synthetic dataset and real-world dataset.The results show that the proposed method outperforms other methods and can improve the F1-score of the RFM features based method.

Key words: Data mining, Fraud detection, Markov random field, Semi-supervised learning, Telecommunication fraud