Fake Review Detection Using Deep Neural Networks with Multimodal Feature Fusion Method

In order to enhance brand benefits or discredit competitors, some merchants hire fake reviewers to post large amounts of fake reviews on e-commerce platforms. This behavior inevitably harms consumers’ interests and causes unfair market competition for other merchants. Researches on fake review detection mainly focus on mining the content of the reviews, the behavioral features of the reviewers, or building models using deep learning. However, most existing research have not taken into the differences in motivation between fake positive and fake negative reviews, the review time distribution features of true reviewers post reviews, and how to effectively integrate multi-modal features. In this paper, we collect restaurant review datasets from Yelp.com in three different regions, and propose a fake review detection method based on a neural network model called BERT-Multi feature-TextCNN-BiGRU-Attention(BMTBA). Firstly, we use the BERT pre-training model to train a restaurant review language model. Then, we propose to use a multimodal fusion method to combine the BERT pre-trained word vector sequences with extracted multidimensional statistical features as input(including a newly proposed reviewer feature called Review weekday). Finally, considering that the motivation for fake positive and fake negative reviews is different, we construct fake positive and fake negative model separately to detect them. Multiple ablation experiments are conducted on the three datasets mentioned above, and the results show that the proposed BMTBA model outperformed the baseline model (BERT-TextCNN-BiGRU-Attention) with a higher classification detection accuracy of 94.68%.