Fine-Tuning Pre-Trained Model for Consumer Fraud Detection from Consumer Reviews

Consumer fraud is a significant problem that requires accurate and prompt detection. However, existing approaches such as periodic government inspections and consumer reports are inefficient and cumbersome. This paper proposes a novel approach named CFD-BERT, to detect consumer fraud automatically based on the group intelligence from consumer reviews. By applying the correlation between consumer reviews and official regulations to accurately mine consumer fraud patterns, and fine-tuning a pretrained model BERT to better model their semantics, which can detect fraudulent behaviors. Experimental evaluations using real-world datasets confirms the effectiveness of CFD-BERT in fraud detection. To explore its potential application and usefulness in real world scenarios, an empirical study was conducted with CFD-BERT on 143,587 reviews from the last three months. The results confirmed that CFD-BERT can serve as an auxiliary tool to provide early warnings to relevant regulators and consumers.