Redes neurais de grafos aplicadas à detecção de lavagem de dinheiro

Abstract

Financial crimes exist in all world countries, and one of the most recurrent ones is Money Laundering. It is capable of causing enormous damage, both financial and reputational, to the companies and government agencies involved in the process. Currently, such organizations use algorithms involving Artificial Intelligence techniques to detect suspicious Money Laundering financial transactions. However, such methods generate many suspicious transactions, often requiring a posterior human evaluation to confirm the suspicion, increasing financial costs and time spent. The literature has presented more robust alternative methods to solve these problems, often involving Machine Learning techniques. In this scenario, since it is possible to represent financial transactions through graphs, methods involving Graph Neural Networks (GNN) have proven to be a promising solution for detecting suspicious Money Laundering transactions. It is possible to represent transactions both as vertices and edges through graphs, impacting the choice of the GNN model for the detection process. This study evaluates the well-known Convolutional Graph Network (GCN) and Skip-GCN, as well as the recent Node and Edge Neural Network (NENN), for the Money Laundering automated detection problem solution, testing them in financial transactions generated by the AMLSim simulator. Four databases were generated to test the influence of class imbalance on detection quality: AMLSim 1/3, AMLSim 1/5, AMLSim 1/10, and AMLSim 1/20, with imbalance rates of 3, 5, 10, and 20, respectively. Initially, the GNN models were tested on all datasets, with the classification done by Softmax and XGBoost. Then, a hyperparameter optimization was performed on the models on the AMLSim 1/20 database, aiming to improve the results for the highest imbalance rate. The precision increase through classification performed by Softmax + XGBoost combination arranged in cascade was also evaluated so that the next classifier confirms the detection of suspicion by the previous one. In the initial results, although the GCN and Skip-GCN models performed better overall, the combination NENN + XGBoost achieved better results for the AMLSim 1/20 set, with a macro-F1 of 86.69%, indicating the positive influence of the representation of transactions as edges of the graph. After hyperparameter optimization, all models improved their results, and the combination with the highest F1 (88.77%) became Skip-GCN + Softmax. Using the combination of Softmax + XGBoost classifiers, the Skip-GCN model obtained the best F1 (88.90%).