Enhancing Illicit Activity Detection using XAI: A Multimodal Graph-LLM Framework

Explainability is critical for financial crime investigators: The paper emphasizes that while deep learning models are effective at detecting illicit activity, investigators need to understand why the model flagged something as suspicious. Current systems have failed to provide this transparency, making it difficult for human analysts to act on AI recommendations.

Multiple data types require different analytical approaches: The framework combines three distinct deep learning models to analyze transaction sequences, network connections between entities, and narrative descriptions. This multimodal approach recognizes that financial crime detection cannot rely on a single type of data or analysis method.

AI-generated narratives can aid human decision-making: The system uses large language models (LLMs) to generate contextual explanations of transactions and their metadata. This bridges the gap between complex model outputs and human understanding, helping analysts make informed investigative decisions rather than blindly trusting algorithmic flagging.

Responsible AI requires centering the human investigator: The paper explicitly positions "the investigative analyst at the heart of any deep learning platform," suggesting that effective financial crime prevention systems must be designed with human oversight and decision-making in mind, rather than treating AI as a replacement for human judgment.