一种改进图神经网络的虚假评论检测方法

A Fake Reviews Detection Method Based on Improved Graph Neural Network

电商平台中的虚假评论存在误导消费者购买决策、损害消费者和合法商家权益的问题。现有的虚假评论识别方法难以发现虚假评论之间的隐含关联。为了提高虚假评论检测的分类效果,提出一种基于TrustRank算法和图神经网络的虚假评论检测方法。首先,通过构建虚假评论相关特征,计算出特征的重要性分数;其次,结合改进Trust⁃Rank方法计算评论的可疑值,利用自适应邻域采样策略对图中节点进行有偏向和自适应地选择并生成目标节点的邻域,以此改进GraphSAGE的随机采样算法;最后,使用Yelp数据集对该模型进行验证。结果表明,TR-GraphSAGE模型相比于LSTM、TextCNN、GCN和GraphSAGE,在准确率、召回率与F13个方面分别提升了5.86%、15.01%和10.12%。TR-GraphSAGE模型能够降低噪音对预测的影响,保证邻域信息的质量和数量,从而提高关联特征表示质量,为虚假评论检测提供了一种新方法。

Fake reviews in e-commerce platforms mislead consumers'purchase decisions,and damage the rights and interests of consumers and legitimate businesses.The existing methods are difficult to find the implicit association between fake reviews.In order to improve the clas⁃sification accuracy of fake reviews detection,a fake review detection method based on the TrustRank algorithm and graph neural network was proposed.Firstly,the features associated with the fake reviews were constructed,and the importance scores of these features were calculated.Secondly,to improve the random sampling algorithm of GraphSAGE,the suspicious values of fake reviews were calculated by the improved TrustRank method,which combined the adaptive neighborhood sampling strategy was used to select nodes in the graph with bias and adaptive and generate the neighborhood of target nodes.Finally,Yelp data set was used to verify the proposed model.The accuracy,recall and F1 of TR-GraphSAGE model were approximately 5.86%,15.01%,and 10.12%better on average,respectively,than LSTM,TextCNN,GCN,and GraphSAGE.The TR-GraphSAGE model can eliminate the noise that affects the prediction,ensure the quality and quantity of the neighbor⁃hood,and thus improve the quality of the associated feature representation,which provides a new method for fake review detection.