A metadata-aware detection model for fake restaurant reviews based on multimodal fusion

Nowadays, online reviews have a significant impact on customers who purchase products or evaluate services, as well as merchants who do business. Unfortunately, there are a lot of fake reviews included in them, which badly affect consumers’ decision-making. Fake review detection has been a significant focus of many researchers in the last few years. Existing approaches mostly focus on the review text, the behavior of reviewers, and the relationship between review entities, while ignoring the abundant metadata in reviews. In this research, we construct the first publicly available dataset of fake restaurant review detection with rich metadata. Through detailed analysis of our dataset, we extract 30 metadata features and text features to help improve the detection performance of models, including 10 brand-new features as well as 9 redefined features. Moreover, a novel metadata-aware model for fake restaurant review detection is proposed. In this model, review text is fed to a transformer model and a convolutional neural network based on pre-training strategies, and the extracted features are fed to a multi-layer perceptron with an attention mechanism. Such a structure enables an effective combination of text and rich metadata. Experimental results show that our proposed model has good stability and robustness and achieves an accuracy of 93.12%, outperforming the baselines by 1.73% in generalization ability test. The fake review dataset and innovative framework proposed in this study can provide ideas and solutions for future research.