融合注意力机制与图神经网络的电商虚假评论识别研究

Research on the Identification of False Reviews in E-Commerce by Integrating Attention Mechanism and Graph Neural Network

摘要: 针对现有图神经网络模型在电商虚假评论识别任务中对关键信息聚集不足的问题，本文通过对CARE-GNN模型引入节点基于点积的注意力聚合，提出一种融合注意力机制的Att-CARE-GNN改进模型。相较于原始模型，本文所改进的模型能够在聚合阶段为目标节点的邻居节点动态赋予聚合权重，从而优化特征选择过程，进而增强模型对电商等虚假评论的识别效果。本文基于YelpChi以及Amazon数据集进行实验，对比CARE-GNN等现有基准模型，结果表明，Att-CARE-GNN在F1值、准确率上表现优异，在YelpChi数据集上F1值与准确率分别提升1.7%、4.9%以上，在Amazon数据集上，F1值与准确率分别提升0.3%、1.1%以上，验证了注意力机制在抑制噪声干扰、提升关键特征权重分配方面的有效性。本文为电商平台的虚假评论识别提供了更具鲁棒性和可解释性的解决方案。

Abstract: Aiming at the problem of insufficient aggregation of key information in the task of identifying false reviews in e-commerce by the existing graph neural network models, this paper introduces node attention aggregation based on dot product to the CARE-GNN model and proposes an improved Att-CARE-GNN model integrating the attention mechanism. Compared with the original model, the improved model proposed in the paper can dynamically assign aggregation weights to the neighbor nodes of the target node in the aggregation stage, thereby optimizing the feature selection process and further enhancing the model’s recognition effect on false reviews in e-commerce and other fields. This paper conducts experiments based on the YelpChi and Amazon datasets and compares existing benchmark models such as CARE-GNN. The results show that Att-CARE-GNN performs excellently in F1 value and accuracy. On the YelpChi dataset, the F1 value and accuracy increase by more than 1.7% and 4.9% respectively. On the Amazon dataset, the F1 value and accuracy rate have increased by more than 0.3% and 1.1% respectively, verifying the effectiveness of the attention mechanism in suppressing noise interference and improving the distribution of key feature weights. This article provides a more robust and interpretable solution for identifying false reviews on e-commerce platforms.