Scholarlink

<p>Consumer choices, market dynamics, and brand reputation are all greatly influenced by customer reviews. However, the increasing number of manipulated or fake reviews damages digital platforms’ integrity and erodes trust. This study introduces a hybrid model for detecting fraudulent reviews that combines graph-based methods and metaheuristic optimization with supervised and unsupervised learning approaches. Graph Autoencoders are used to identify anomalies in unlabeled datasets by learning latent graph structures, whereas GraphSAGE is applied to labeled review data to produce contextual embeddings through neighborhood aggregation. By fine-tuning model parameters, the Whale Optimization Algorithm (WOA) improves generalization and reduces overfitting. The last classifier is LightGBM. BERT embeddings, sentiment scores (VADER), and PCA for dimensionality reduction are all used in feature extraction. The approach exhibits improved capacity and adaptability compared to conventional models such as Graph Convolutional Networks (GCNs), Graph Attention Networks (GAT), and Recurrent CNNs. Four datasets—40K, Yelp, openSnippet, and Google Maps—are adopted to test the model. Its results show 98% accuracy, 0.99 precision, 0.95 recall, and an F1-score of 0.98. The outcomes demonstrate how well structural, semantic, and optimization methods work together to determine fraudulent reviews in dynamic, large-scale settings.</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

Synergizing graph embeddings and metaheuristic optimization for fraudulent review detection

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Kuldeep Vayadande,
  • Amit Mishra,
  • Yogesh Bodhe,
  • Ninad Kale,
  • Anish Katariya,
  • Amey Kharade,
  • Parth Supekar,
  • Lalit Patil

摘要

Consumer choices, market dynamics, and brand reputation are all greatly influenced by customer reviews. However, the increasing number of manipulated or fake reviews damages digital platforms’ integrity and erodes trust. This study introduces a hybrid model for detecting fraudulent reviews that combines graph-based methods and metaheuristic optimization with supervised and unsupervised learning approaches. Graph Autoencoders are used to identify anomalies in unlabeled datasets by learning latent graph structures, whereas GraphSAGE is applied to labeled review data to produce contextual embeddings through neighborhood aggregation. By fine-tuning model parameters, the Whale Optimization Algorithm (WOA) improves generalization and reduces overfitting. The last classifier is LightGBM. BERT embeddings, sentiment scores (VADER), and PCA for dimensionality reduction are all used in feature extraction. The approach exhibits improved capacity and adaptability compared to conventional models such as Graph Convolutional Networks (GCNs), Graph Attention Networks (GAT), and Recurrent CNNs. Four datasets—40K, Yelp, openSnippet, and Google Maps—are adopted to test the model. Its results show 98% accuracy, 0.99 precision, 0.95 recall, and an F1-score of 0.98. The outcomes demonstrate how well structural, semantic, and optimization methods work together to determine fraudulent reviews in dynamic, large-scale settings.