A large language model-based approach for fake review detection: the implicit characteristics perspective

Detecting fake online reviews is crucial for maintaining the integrity and sustainable development of digital platforms. Prior work has focused on overt linguistic signals, but the implicit traits of deceptive content remain largely untapped, thus constraining detection accuracy. This study proposes a novel fake review detector called LLMIC, which leverages large language models (LLMs) and introduces complex system paradigms to quantify the implicit characteristics. Specifically, we propose a new multifractal method XSMF to explore multi-faceted fractal patterns of reviews, and a new recurrence plot approach MSFRP to analyze multi-layered recurrence behaviors of reviews. Extensive experiments demonstrate that LLMIC outperforms dozens of recent models, and also confirm the effectiveness of XSMF and MSFRP. In addition, we provide new insights: fake reviews tend to exhibit larger Hurst exponents, broader Hölder exponents, and more prominent recurrence diagonals, reflecting their greater complexity and variability.