Large Language Models for Fraud Detection

Large Language Models (LLMs) are reshaping the landscape of Machine Learning (ML) application development. This study addresses the problem of whether a general language model pre-trained on a large amount of data samples could be effective in social security data fraud detection tasks. We selected the GPT-4 model and designed a template for prompt engineering to explore its potential capability in analyzing. Additionally, we compared it with other methods such as Ensemble Bayesian, BP neural network, Bert, and Transformer. The experimental results indicate that the large language model, after prompt engineering, achieved an F1 score of 0.89 on the same test set, surpassing other state-of-the-art works.