Yu Shuo, Ren Jing, Li Shihao, Naseriparsa Mehdi, Xia Feng
School of Software, Dalian University of Technology, Dalian, China.
Institute of Innovation, Science and Sustainability, Federation University Australia, Ballarat, VIC, Australia.
Front Artif Intell. 2022 Jun 20;5:922589. doi: 10.3389/frai.2022.922589. eCollection 2022.
Fake reviews have become prevalent on various social networks such as e-commerce and social media platforms. As fake reviews cause a heavily negative influence on the public, timely detection and response are of great significance. To this end, effective fake review detection has become an emerging research area that attracts increasing attention from various disciplines like network science, computational social science, and data science. An important line of research in fake review detection is to utilize graph learning methods, which incorporate both the attribute features of reviews and their relationships into the detection process. To further compare these graph learning methods in this paper, we conduct a detailed survey on fake review detection. The survey presents a comprehensive taxonomy and covers advancements in three high-level categories, including fake review detection, fake reviewer detection, and fake review analysis. Different kinds of fake reviews and their corresponding examples are also summarized. Furthermore, we discuss the graph learning methods, including supervised and unsupervised learning approaches for fake review detection. Specifically, we outline the unsupervised learning approach that includes generation-based and contrast-based methods, respectively. In view of the existing problems in the current methods and data, we further discuss some challenges and open issues in this field, including the imperfect data, explainability, model efficiency, and lightweight models.
虚假评论在电子商务和社交媒体平台等各种社交网络上已变得十分普遍。由于虚假评论对公众造成了严重的负面影响，及时检测和应对具有重要意义。为此，有效的虚假评论检测已成为一个新兴的研究领域，吸引了网络科学、计算社会科学和数据科学等各学科越来越多的关注。虚假评论检测中的一个重要研究方向是利用图学习方法，该方法将评论的属性特征及其关系都纳入检测过程。为了在本文中进一步比较这些图学习方法，我们对虚假评论检测进行了详细的调研。该调研提出了一个全面的分类法，并涵盖了三个高级类别中的进展，包括虚假评论检测、虚假评论者检测和虚假评论分析。还总结了不同类型的虚假评论及其相应示例。此外，我们讨论了图学习方法，包括用于虚假评论检测的监督学习和无监督学习方法。具体来说，我们概述了无监督学习方法，该方法分别包括基于生成和基于对比的方法。鉴于当前方法和数据中存在的问题，我们进一步讨论了该领域的一些挑战和开放性问题，包括不完美的数据、可解释性、模型效率和轻量级模型。
