Temporal Cycle Enumeration for Detecting Financial Fraud

Fraud detection is an important task of financial supervision and financial risk management. Several fraud behaviors can be connected with cycle transactions in which the money initially sent from one bank account eventually returns back to the same account. In this paper, we propose an efficient method for detecting such kind of frauds from large-scale financial transaction data. The method first constructs a transaction graph after pre-processing the original data, then divides the graph into its strongly connected components, and finally uses multiple threads to enumerate temporal cycles on different components in a parallel manner. Existing temporal cycle enumeration algorithms usually constraint the length of the cycle or the size of the time-window, which are not suitable for the specific application of financial fraud detection. In light of this, we extend the classical Johnson algorithm to enumerate temporal cycles without length and time-window constraints. To further improve the efficiency of enumeration, we introduce a block-time mechanism that avoids unnecessary multiple explorations of the same parts of the graph components. Experiments show that our method, with multithreading, is on average 15–20 times, and even 100 times faster than the existing competitor. Additionally, we adopt strategies such as amount constraint during cycle enumeration, which assist in reducing the false-positive rate of detected frauds.