Network-based methods for anti-money laundering

Abstract

This study concentrates on the contributions of network-based methods in antimoney laundering. Being responsible for reporting suspicious activities, time-consuming analyses conducted by the banks through rule-based systems result in high false positive rates. Today, machine learning methods draw attention for their ability to reduce computational cost and false positive rates in anti-money laundering. In this research, the goal is to both reduce dimension of large and complex financial networks while minimizing the loss of valuable information and identify suspiciousness by feeding machine learning algorithms by network-based features using bank transaction networks. Dimensional reduction of networks in the first place enhanced the efficiency of following subgraph construction algorithms. Based on the hypothesis that money launderers disguise organizedly in licit activities, hence network relations bear the trace of hidden suspects, candidate subgraphs are constructed. Network-based features of each subgraph such as inside transaction volume, amount of outgoing transactions, size and so forth are calculated and these subgraphs are labeled as innocent or suspicious according to presumed risk aversion behaviors of banks. Created candidate subgraphs data set is used for training supervised classifiers and the evaluations on separate test data sets indicate that a satisfactory discriminative performance is obtained. These outcomes support the initial hypothesis that network-based analysis in anti-money laundering would contribute more compared to accustomed rule-based systems or individual analysis methods. These findings are promising for declining false positive rate and easing the burden on labor force in the event of a prospective implementation of the suggested system on real data to be acquired from banks.