Research On Several Anti-Money Laundering Algorithms Based On Network Representation Learning

In recent years,with the rapid development of the China’s economy,economic crimes occur frequently,and the types of economic crimes also increase frequently.The number of suspicious transactions involving money laundering crimes is also increasing day by day.The People’s Bank of China also receives more and more suspicious transaction reports on anti-money laundering each year.It is urgent to crack down on the crime of moneylaundering work.With the advent of the Internet era,criminal behaviors are more and more closely related to the Internet,facing the characteristics of diversified criminal means and high concealment of crimes,the investigation of money laundering cases is also more difficult.In the financial field,anti-money laundering has been deeply studied and a series of theories have been formed.However,the research methods in this field are all from the perspective of Commercial Banks,most of them mainly focus on the identification of suspicious transactions and abnormal accounts.If the analysis theories and methods in the financial field are directly applied to the investigation of public security cases,the actual effect is not ideal.Therefore,from the perspective of public security investigation,this paper proposes several anti-money laundering algorithms,combined with the characteristics of large amount of money laundering transaction data and high transaction amount,this paper proposes several anti-money laundering algorithms are to assist the case analysis of suspicious money laundering transactions.The main work and innovation of this paper can be summarized as the following four points:First,the layered analysis of the role of money laundering funds in macro roles.Because of the large amount of money laundering data,complex transactions,it is difficult to directly from the huge amounts of data to find out the various accounts of general character,this paper designs a money laundering money hierarchical model,according to the account of capital flow direction account division level,and to simplify the large-scale data,the accounts could be divided into money into account layer,middle layer and funds liquidated,from on macroscopic display account trading role in money laundering network.Second,hierarchical data visualization and automatic layout.Aiming at the capital data after layering,we design a constrained force directed automatic layout algorithm,which can display the hierarchical relationship between the accounts after layering,through automatic layout in a visual way,thus this algorithm can solve the problem that it is impossible to manually lay out the large amount of data.In the process of the realization of the algorithm,because the nodes’ initial layout influence on the final layout greatly,we use genetic algorithm to determine the initial layout of each account,optimize the intersection of the edges connected by each layer,and then compress each node in each layer so that each node can only move in the horizontal direction,then simulate the principle of the spring,according to the effect of gravity and repulsion of other nodes to achieve automatic layout,make nodes move until finally reaching a stable state.The visual layout of the hierarchical data can clearly show the macro roles of the accounts in the millions or even tens of millions of transaction flow data,including the account of the capital inflow layer responsible for the source of funds,the account of the intermediate flow layer responsible for the working capital of each other,and the account of the terminal realization layer responsible for the realization and precipitation of funds.Third,money laundering group found.According to the actual investigation results of the public security department,money-laundering crimes often show group characteristics,but in case investigations,we encounter massive capital transaction data,and it is impossible to divide groups manually.In the money-laundering cases,the group members have a high degree of connectivity in the transaction of funds inevitably.Therefore,in this paper,we design a group division algorithm based on network representation learning,which can automatically divide groups from the capital transaction data.Firstly,use the representation learning algorithm to learn the vector features of accounts,and then cluster the vector features to form groups.Compared with the traditional community discovery algorithm,which ignores money laundering funds’ direction,it only divides groups based on the transaction relationship between accounts,without considering attributes such as amount,transaction number,and other important transaction characteristics.The group division algorithm combines the characteristics of the transaction number,the transaction amount,and the centrality of the intermediary.We use the modularity to evaluate the divided groups,the value of the modularity is about 0.66,which has achieved good results.Fourth,suspicious transaction patterns of money laundering match.In order to analyze the suspicious transaction patterns of group member accounts in detail,it is necessary to match some common money laundering suspicious transaction patterns for accounts.Traditional pattern matching algorithms have limitations such as large calculation volume and low fault tolerance.This paper proposes a transaction pattern matching algorithm based on semi-supervised network representation learning.Firstly,we construct a fixed dataset of suspicious transaction patterns,which is considered as supervised dataset,and then process money laundering dataset of unknown patterns,put them together in a graph embedding algorithm based on structural similarity for training,and finally by calculating the structural similarity between the fixed suspicious transaction pattern node and the money laundering data node to match the suspicious transaction pattern of the account.The algorithm proposed in this article was applied to the actual cases of the Ministry of Public Security in 2018,achieved certain results all.With only capital transaction data,these algorithms assisted the public security organs to find each account’s main role in the money laundering data,divided the suspicious money-laundering groups,and found group member’s suspicious transaction behavior patterns in various accounts,provided important clues for case investigation.