Uncovering fake reviews through graph-embedded hierarchical attention networks

User-generated reviews strongly influence consumer decisions and product strategies on e-commerce platforms; however, fake reviews significantly undermine their credibility. Existing fake review detection (FRD) approaches struggle with limited context and sparse features, thus restricting their ability to capture deeper linguistic and relational patterns. Consequently, this paper proposes a multi-component hierarchical neural architecture, GEmHAttenFake, that integrates Graph Convolutional Networks (GCNs) embedded representation with a Hierarchical Attention Network (HAN) to model both the linguistic and relational dependencies within reviews for FRD. At the word and sentence levels, a Bi-LSTM encoder coupled with a graph-based representation captures syntactic and contextual relationships among words, aspects, and sentiment, which helps detect unnatural and fabricated patterns in fake reviews. Experiments on benchmark Amazon review datasets show that the proposed GCN-HAN with verified purchase achieves 85.35% accuracy and 84.55% F1-score. GEmHAttenFake outperforms non-GCN-HAN with identical dependency features by + 4.59% accuracy and + 3.23% F1-score. It also surpasses the GCN-HAN variant without verified purchase by + 9.32% accuracy and + 5.97% F1-score. These results demonstrate the effectiveness of combining relational modeling with verified purchase signals while improving prediction interpretability. Additional assessment on the OPSpam dataset further validates the effectiveness of the proposed model.