Fraud Investigation using LLMs and Graph Databases

Fraud Investigation using LLMs and Graph Databases

The Challenge

A company aims to develop a question answering system for fraud investigators to detect factors leading to fraudulent transactions from fraud data. The challenge lies in transactions possibly sharing relationships among them. To address this, Tatras developed a solution using network graphs and GenAI, significantly enhancing efficiency, accuracy, and scalability in fraud investigation.

Hypothesis

  • Graph-based data structures can enable easy backtracking and relation finding for transactions.
  • GenAI advanced agents can generate dynamic graph queries to retrieve data.
  • Retrieved data can be used to generate inferences and answer investigator queries.

Execution

  • Utilized Neo4j Graph DB to store data.
  • GenAI Cypher QA agents analyze user queries and graph schema to generate and execute graph queries.
  • REACT agent analyzes fetched data to discover intriguing patterns and provides the final answer based on its findings.
  • Scikit-learn, Langchain, PyTorch, Llama Index Transformers, Neo4

Outcomes

  • Question Answering system to investigate fraud.
  • Improved ability to track transactions seamlessly from start to endpoint using the graph-based system.
  • GenAI agents capability to generate dynamic queries enhances system flexibility and generality.
  • Graph-based data structure to effectively investigate the fraud pattern.

Project Highlights

  • Significantly reduced manual efforts and increased accuracy of fraud detection.