Enhancing Decision-Making in Anti-Money Laundering with Explainable AI (XAI) Visualization: Aligning Human and Machine Learning Decision Making Processes

Abstract

This thesis explores how the incorporation of Explainable AI (XAI) visualization can improve decision making processes in Anti Money Laundering (AML) investigations. AML investigators face challenges in combining human expertise with complex machine learning systems and handling labor-intensive tasks like labeling data. However, existing Visual Analytics (VA) tools tend to focus on the needs of data analysts rather than decision makers. They often lack the necessary flexibility to adequately support the dynamic decision making process in AML. Therefore, this study expands the concepts of weak supervision and interactive data programming by developing a prototype tool that incorporates AI assistance, particularly through the use of a large language model, to facilitate labeling functions (LFs) creation. We conducted a user study with 10 participants, divided into an AI-assisted group and a control group. We then observed how these participants engaged with our VA prototype using a synthetic AML dataset. Quantitative data on task completion time, System Usability Scale (SUS) scores, and performance metrics were collected, alongside qualitative data from interviews and observations. The quantitative analysis indicated that participants in the AI-assisted group finished the LFs creation tasks more quickly and reported a higher average SUS score. This suggests that the use of AI contributed to enhanced efficiency and better usability overall. The qualitative findings also indicated that although participants valued the assistance provided by the AI in formulating LFs, the integration of AI also brought about certain challenges. Participants who faced uncertainty in the LF creation process expressed preferences for better interpretability about the reasoning behind the AI's decisions. Additionally, some participants found the LFs generated by AI to be somewhat simplistic or less useful once they developed their own understanding. Taken together, the incorporation of AI support within our VA tool demonstrates promise for improving both efficiency and user experience in AML labeling tasks. However, there remain opportunities for enhancement in explainability, and future studies should focus on this aspect to optimize the human and AI collaboration in this domain.