An Explainable Data Depth-Based Method in Anti-Money Laundering (AML) Domain

This paper proposes a novel approach for anomaly detection in Anti-Money Laundering (AML) using data depth-based methods, specifically $\beta$-Skeleton Depth and Projection Depth. These unsupervised techniques leverage geometric principles and statistical analysis, with hyperparameters fine-tuned through Pareto optimization to enhance detection performance. Depth functions deliver good results while offering advantages such as interpretability and independence from normalization. Experimental results validate their effectiveness, with Logistic Regression and Autoencoders serving as baselines. Autoencoders achieved the best performance among unsupervised models but face challenges in explainability. Logistic Regression, while yielding strong results, requires labeled data, limiting its applicability in unsupervised scenarios. Correlation analysis further highlights the strong relationship between depth values and the target variable, showcasing their potential as features for AML detection. This study underscores the value of depth-based methods as robust, interpretable tools for anomaly detection in financial data.