Graph Neural Network for Credit Card Fraud Detection

With the rapid development of network payment, events of electronic transaction fraud take place often and result in lots of losses to customers and merchants. Therefore, transaction fraud detection is very important, and many detection methods have been studied and applied based on big data and machine learning. Some focused on the relationship of transactions, some on the relationship of original transaction features, and some on the dynamic changes of behaviours of users. However, there is few method to comprehensively characterise them. Hence, we design a weighted multiple graph called Transaction Graph (TG), and use Graph Neural Network (GNN) and TG to train a detection model so that the above characters can be represented comprehensively and thus the detection performance is enhanced obviously. We first construct a group of rules that are represented by a group of logical propositions and used to describe some transaction characters, and assign different weights for these rules to reflect their different importances. In our TG, a node corresponds to a transaction record. If two records satisfy a rule, then the two nodes corresponding to the two records are connected by an edge whose weight is equal to the weight of the rule. Based on our TG, we propose a new sampling policy which can sample neighbour nodes with more useful information. Additionally, we propose a new attention mechanism for GNN in order to decrease the interference of abnormal samples and increase influence of normal ones. Our experiments on a big dataset of real transactions illustrate the advantages of our method compared with some state-of-the-art ones.