Enhancing Graph Database Interaction through Generative AI-Driven Natural Language Interface for Financial Fraud Detection

The rapid evolution of financial services has led to an increase in fraudulent activities, necessitating advanced detection methods. Traditional techniques often struggle with the volume and complexity of financial transactions, leading to inaccuracies and delays. This research addresses the enhancement of graph database interaction through a generative AI-driven natural language interface for more accurate fraud detection. We employ the T5 and LLaMA-2 models, trained on a balanced dataset of financial transactions, to recognize fraud patterns. The $\mathbf{T 5}$ model handles natural language queries, extracting key details such as transaction amounts and account numbers. The LLaMA-2 model generates predictions on the likelihood of transactions being fraudulent. The system then converts these processed queries into Cypher queries for the Neo4j graph database to retrieve relevant transaction data. This approach enables efficient pattern analysis and interactive data exploration. Experimental results demonstrate that our system accurately identifies fraudulent transactions, significantly reducing the time and expertise needed for detection. This approach makes use of Neo4j’s powerful features to increase the efficacy and precision of financial detection systems for fraud.