Search

In the field of online lending, the key problem for fraud detection is how to judge whether the user is a fraudster or a normal user based on the collected historical transaction data of the user. At present, the representative research methods treat any user as an independent node and ignore the related information among users. Considering that the fraud is gradually becoming a group behavior, the relationships among fraud nodes and non-fraud nodes are sparse in social networks, and the relationships among fraud nodes are closely related, we propose a collective classification fraud detection method with label propagation. A call-records-based user association network is constructed based on the phone call records between users of online lending company, and we use the label propagation algorithm to spread the label information of fraud node to determine whether the unlabeled node is a fraudulent user. In addition, we improve the initialization method of transition probability matrixin label propagation algorithm by the operation of weights powering to avoid the performance degradation of label propagation algorithm caused by the unbalanced distribution of fraud data. Finally, the validation experiment is conducted in a real loan data set with a very low proportion of labeled samples and unbalanced training sample distribution. By using the proposed method in this article, the accuracy rate of fraud user detection reaches 17%, and the F1 value and accuracy rate are both better than those of the classic WvRn algorithm. [ABSTRACT FROM AUTHOR]