Research On Machine Learning Based Money Laundering Risk Identification Method

The present model in the process of the continuous development of the world economy,money laundering as a financial offense for also emerged with the development of the world economy.After paying attention to the phenomenon that a large number of criminals around the world transfer illegal funds through the financial system every year and lead to difficulties in tracing and disposing of illegal funds^[1,2],how to effectively detect abnormal financial activities has become a great challenge for governments and financial institutions.Through extensive research and combing existing studies,it is found that many mature techniques can provide solution ideas for money laundering risk identification.However,for the problems of high false positive detection rate and high labor cost caused by the most primitive rule-based expert systems,integrated tree-based learning models show more potential in solving this problem.However,the classification performance of such models is often greatly affected by the feature selection of training data and the parameters of the model,and it is worth thinking about how to effectively perform feature selection,unbalanced sample and parameter setting.And the problem of concept drift often arises in money laundering transaction detection,and most studies neglect to think about this problem.In addition to being able to detect suspicious transactions from individual transaction features,existing graph learning techniques provide new ideas for money laundering transaction detection from the perspective of transaction networks,while most of the existing solutions provided from the perspective of graph learning ignore the practical problem of dynamic changes in transaction networks over time,which only focus on the static structure of the network or require learning the dynamic changes of the graph over the entire time span and are not applicable to transactional networks with frequent change characteristics.Therefore,the following studies are carried out in this paper to solve the above problems:（1）XGBoost based on incremental learning for money laundering transaction detection.In order to solve the impact of"dimensional disaster"on the performance of the model,the post-pruning random forest is used to calculate the feature importance and select the features larger than a certain threshold value as the training features of the model.To address the impact of sample imbalance on the model detection performance,an improved SMOTE algorithm is proposed to oversample the imbalanced category data.By comparing the model with the conventional parameter search algorithm,the particle swarm optimization（PSO）algorithm can search for the combination of XGBoost parameters faster and more effectively,which further improves the classification performance of the model.Finally to solve the problem of conceptual drift of the detection model due to the change of transaction patterns in real transaction scenarios,this model improves this situation by performing incremental learning on XGBoost.To verify the effectiveness of the model,the final binary classification task is performed on the Elliptic dataset,and the experimental results validate the effectiveness of this model for money laundering transaction detection.（2）A spatio-temporal attention-based graph convolutional network for anti-money laundering.At the end of（1）,it is experimentally demonstrated that the features of to that undergo graph embedding by GCN,i.e.,the features that fuse the transaction features of neighboring nodes,can further improve the classification performance of the model,and then this model proposes the Temporal GAT model from considering transaction networks.It performs node-preferred embedding of the spatial structure of the transaction graph by using a multi-headed attention mechanism;and adapts the multi-headed graph attention network to the dynamics of the graph sequence by introducing a dynamic update mechanism.Finally,the node（transaction entity）classification task is evaluated on the publicly available dataset Elliptic dataset,and the results of the ablation and hyperparameter experiments show that Temporal GAT has better classification performance on all classification evaluation metrics compared to existing methods.（3）Design and implementation of financial risk control systems.Considering that existing financial institutions may still be stuck in rule-based expert systems when dealing with fraud and money laundering risk problems,they lack effective integration with advanced artificial intelligence technologies.Therefore,this paper addresses this challenge by proposing a financial risk control system consisting of a fraud detection subsystem for network transactions,a money laundering risk identification subsystem,and a user rights and information management subsystem.In Chapter 4,the requirements analysis,outline design,and detailed implementation of the money laundering risk identification subsystem are presented in detail,and the incremental learning-based XGBoost model and the spatio-temporal attention-based graphical convolution model proposed in the previous two chapters are applied in this subsystem to verify the effectiveness of the proposed model.