Master's Thesis : Machine Learning Techniques for Money Laundering Detection
Promotor(s) : Geurts, Pierre
Date of defense : 7-Sep-2020/9-Sep-2020 • Permalink :
|Master's Thesis : Machine Learning Techniques for Money Laundering Detection
|Translated title :
|[fr] Techniques d'apprentissage automatique pour la détection du blanchiment d'argent
|Date of defense :
|Committee's member(s) :
|Number of pages :
|[en] money laundering
[en] machine learning
[en] imbalance problem
[en] data generation
[en] F1 score
[en] outlier detection
|Engineering, computing & technology > Computer science
|Université de Liège, Liège, Belgique
|Master : ingénieur civil en informatique, à finalité spécialisée en "management"
|Master thesis of the Faculté des Sciences appliquées
[en] Some economic phenomena have a significant negative impact on financial institutions. Money laundering is one of them. Money laundering is the processing of illicit funds into the financial system to make them appeared from legitimate sources. It is usually detected through rule based monitoring. Unfortunately, it does not exploit the potential of data. Machine leaning approaches learn automatically from data patterns that cannot be captured by rule-based ones. Money laundering in transaction logs are defined as an imbalance problem with regard to machine learning. The challenge is to detect few fraudulent transactions over a huge population which compose a financial database.
This thesis presents the design of a simulator for transaction data generation to answer to the lack of available data needed for machine learning algorithms. Then a theoretical research has been done in order to identify machine learning algorithm designed for classification problem with skew class distribution. The machine learning models which have been selected from this research are the Hellinger Distance Decision Tree (HDDT), the Isolation Forest (iForest), the One-Class SVM (ocSVM) and an approach based on successive hypothesis testing using the Fisher, George and Pearson methods.
These approaches have been tested over synthetic data from the simulator. The model with the highest F1 score is a Random Forest composing of fully expended decision trees using the Gini split criterion. The second best performing model was the Isolation Forest.
Cite this master thesis
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.