From Data to Insights: Entity Resolution, Knowledge Graphs, and Visualization for Money Laundering Detection

Abstract

The increasing complexity of financial systems, combined with an increase in the number of transactions, has made money laundering detection and financial transparency increasingly challenging. This thesis addresses these challenges by proposing a framework that in- tegrates Entity Resolution, knowledge graph analytics and a Pruden- tial Multiple Consensus model. Complex relational patterns which are often missed by rule-based, traditional approaches are uncovered by leveraging graph-based algorithms. A data science pipeline is introduced that combines optional, prob- abilistic Entity Resolution with topological feature extraction to en- hance the accuracy of detecting fraud. Furthermore, to bridge the gap between technical analysis and application, we developed an interactive dashboard that provides the analyst with explainable visualizations and allows for decision traceability. A case study on the dataset of the paper "Enhancing Anti-Money Laundering: Development of a Syn- thetic Transaction Monitoring Dataset" [Ozt+23] demonstrates that while individual classifiers are sensitive to decision thresholds in real- istic scenarios, they maintain high discriminative power (AUC > 0.9). Building on this, the Prudential Multiple Consensus model we intro- duced makes use of topological features in an effective way that in- creased the F1-score in realistic, low-fraud scenarios. However, the results indicate that probabilistic Entity Resolution did not yield the anticipated improvements; in fact, in specific topological scenarios, it introduced noise that degraded the detection accuracy of the consensus model.