Improving fake news detection with domain-adversarial and graph-attention neural network

摘要

With the widespread use of online social media, we have witnessed that fake news causes enormous distress and inconvenience to people's social life. Although previous studies have proposed rich machine learning methods for identifying fake news in social media, the task of detecting fake news in emerging news events/domains remains a challenging problem due to the wide range of news topics on social media as well as the evolution and variation of fake news contents in the web. In this study, we propose an approach which we term “domain-adversarial and graph-attention neural network” (DAGA-NN) model to address the challenge. Its main advantage is that, in a text environment with multiple events/domains, only partial domain sample data are needed to train the model to achieve accurate cross-domain fake news detection in those domains with few (or even no) samples, which makes up for the limitations of traditional machine learning in fake news detection tasks due to news content evolution or cross-domain identification (where there is no sample data). Extensive experiments were conducted on two multimedia datasets of Twitter and Weibo, and the results showed that the proposed model was very effective in detecting fake news across events/domains.