Co-attention Based Deep Model with Domain-Adversarial Training for Spam Review Detection

Douban has become one of the most popular Chinese film review platforms and has been attacked by various spammers. In our research based on real data from Douban, we found that many spammers spontaneously form different groups, which have different targets and have different effects on movies. In this paper, by dividing reviews into four categories, namely, true positive reviews, spam positive reviews, true negative reviews, and spam negative reviews, we design a co-attention based neural network to fuse multiple features to classify reviews. In order to improve the robustness of the detection model, we adopt the idea of domain-adversarial training. In the domain-adversarial training method, we use real data and noisy data for model training, and the model is indiscriminative for whether the data source is real data or noisy data. The experimental results show that our proposed domain-adversarial training method can improve both the best classification performance and robustness.