Large language model meets chaos: A new deep learning model for fake review detection

Detecting fake online reviews is crucial for the e-commerce ecosystem. However, existing studies often fail to mine the intrinsic attributes of reviews, which limits detection performance. In this paper, we introduce a novel fake review detection model, LLMChaos, which investigates reviews from the perspectives of large language models (LLMs) and chaos theory. Specifically, we first propose a method that blends energy selection with LLMs to generate review time series. Second, we construct a space mapping mechanism with multiple chaotic attributes, embracing the intrinsic attributes of reviews. Finally, we design a hierarchical learning network that trains in deep contrastive learning across LLM layers, chaotic attribute layers, and Transformer layers. Extensive experiments demonstrate that LLMChaos is robust and state-of-the-art. For instance, on the Hotel dataset, LLMChaos achieves 94.78% F1, outperforming recent models by 1.42%-19.78%; on the Amazon dataset, LLMChaos achieves 93.15% F1, surpassing recent models by 1.45%-18.39%. Moreover, we contribute novel discoveries, for example, chaotic behaviors of reviews generally exhibit bounded ranges: Lyapunov exponent (0–0.0125), Correlation dimension (0.25–0.5), Kolmogorov entropy (0.75–0.85), Fractal spectrum (0–1.5), and Recurrence rate (0.005–0.015); real and fake reviews display distinct chaotic distributions.