Scholarlink

<p>Product reviews on e-commerce platforms strongly influence consumer purchasing decisions but are increasingly targeted by fraudulent actors seeking to manipulate public perception. This paper presents COVERT-GNN, a novel graph-based framework that integrates metadata, multi-relational structures, and Graph Neural Networks (GNNs) to detect misinformation in user reviews. Unlike existing approaches that rely on static or handcrafted features, COVERT-GNN introduces a vertex-aware covertness mechanism that quantifies both feature-level mimicry and relation-level concealment to reveal fraudsters who intentionally imitate benign users. The framework further incorporates relation-aware neighborhood aggregation and a reinforcement-driven neighbor selection strategy, enabling the model to adaptively capture latent behavioral dependencies across users, items, and reviews. Extensive experiments on the Yelp and Amazon datasets demonstrate that COVERT-GNN achieves strong and consistent performance, obtaining an AUC of 75.52%, Recall of 69.28%, and F1-Macro of 63.70% on Yelp, and an AUC of 93.12%, Recall of 88.03%, and F1-Macro of 92.46% on Amazon. These results confirm the effectiveness of COVERT-GNN in identifying covert misinformation patterns that mimic genuine reviews, offering a robust and scalable solution for reliable fraud detection in e-commerce environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Covertness-aware graph neural networks for misinformation detection in e-commerce

  • 期刊论文

  • 发表日期：

    2026年3月8日

查看全文
  • Abhay Narayan,
  • Dameera Tharun,
  • Madhu Kumar SD,
  • Anu Mary Chacko

摘要

Product reviews on e-commerce platforms strongly influence consumer purchasing decisions but are increasingly targeted by fraudulent actors seeking to manipulate public perception. This paper presents COVERT-GNN, a novel graph-based framework that integrates metadata, multi-relational structures, and Graph Neural Networks (GNNs) to detect misinformation in user reviews. Unlike existing approaches that rely on static or handcrafted features, COVERT-GNN introduces a vertex-aware covertness mechanism that quantifies both feature-level mimicry and relation-level concealment to reveal fraudsters who intentionally imitate benign users. The framework further incorporates relation-aware neighborhood aggregation and a reinforcement-driven neighbor selection strategy, enabling the model to adaptively capture latent behavioral dependencies across users, items, and reviews. Extensive experiments on the Yelp and Amazon datasets demonstrate that COVERT-GNN achieves strong and consistent performance, obtaining an AUC of 75.52%, Recall of 69.28%, and F1-Macro of 63.70% on Yelp, and an AUC of 93.12%, Recall of 88.03%, and F1-Macro of 92.46% on Amazon. These results confirm the effectiveness of COVERT-GNN in identifying covert misinformation patterns that mimic genuine reviews, offering a robust and scalable solution for reliable fraud detection in e-commerce environments.