A Study on Fake Review Detection Based on RoBERTa and Behavioral Features

Abstract

With the rapid development of e-commerce platforms, how to better identify and filter fake reviews has become an urgent issue for the healthy development of the e-commerce industry. However, traditional fake review identification methods cannot effectively solve this problem. They are easily affected by text changes, language differences, and context, and fake reviewers may take measures to blur their behavioral characteristics, making them difficult to detect by behavior-based algorithms. Large language models can capture the contextual relationships of the entire text through self-attention mechanisms, thereby understanding the overall meaning and emotional tendency of the review, which enables them to more effectively judge the authenticity of the review. In addition, large models have strong generalization ability and a certain degree of interpretability, which also makes them suitable for research on fake review text identification. Since fake reviews have an asymmetric distribution feature, this paper uses the RoBERTa model to extract information and combine it with the behavioral features in traditional research for model training. Compared with traditional methods, the accuracy rate is improved by nearly 3%.