Article Preview
TopIntroduction
From the 20th century to the present, the world industrial pattern has gradually tilted toward Internet-related fields. Many IT companies, such as Microsoft, Google, and Alibaba, began to rise rapidly. They do not hesitate to invest huge sums of money and recruit a large number of researchers to seize new fields. At present, the information dissemination carrier represented by Twitter has become the main tool for people to communicate. Users can communicate through social software without leaving home and learn about major events in the world. However, people with ulterior motives have begun to spread rumors with the help of social networks, making it difficult for users to distinguish between true and rumor without their knowledge. Since rumors cover a very wide range and users who publish rumors are very concealed, it is very difficult to supervise them. At present, Baidu, Tencent, Weibo, and other well-known Internet companies have established rumor-refuting platforms. Various Internet platforms organize researchers to explore efficient rumor detection methods that can adapt to the big data environment. Text features (Azri et al., 2021; GuangJun et al., 2020; Li et al., 2022; Ma et al., 2022; Shelke & Attar, 2022; Xu et al., 2021), image features (Azri et al., 2021; Li et al., 2022), user features (Shelke & Attar, 2022), and spread features (Ma et al., 2022) have become mainstream research directions.
To adapt to the environmental requirements of big data, the related methods of rumor detection are gradually transferred from manual-based related methods to machine learning-based related methods. Since the related methods based on machine learning cannot model the social relations of users, this method cannot effectively extract the high-level and abstract features of rumors. In 2016, Kipf & Welling (2016) proposed graph convolutional neural networks. The related methods of graph neural networks gradually entered the field of view of many scholars and achieved good performance. Since the graph convolutional neural networks(GCN) needs to introduce an adjacency matrix, out-degree nodes and in-degree nodes need to participate in the node aggregation process at the same time, which limits the aggregation direction. At this stage, the related methods of rumor detection mainly consist of related methods based on machine learning and related methods based on graph neural networks.
Related technologies based on machine learning have been very mature. Most scholars use classifiers or classification functions to determine whether tweets are rumors by extracting relevant features and inputting them into trained models. Ma et al. (2016) captured the time-varying contextual features through a recurrent neural network and proposed a rumor detection model that fuses temporal feature information; Shi et al. (2018) not only improved the detection efficiency but also solved the problem of data sparseness by fusing the recurrent neural network with the topic features of emergencies; Min et al. (2016) combined a momentum model and temporal analysis-based method to filter fake microblogs. These methods demonstrate the importance of temporal features in rumor detection by extracting temporal relationships between tweets or keywords. Gao et al. (2020) used task-specific features based on bidirectional language models to learn contextual embedded textual information and event sequence information; Liu et al. (2020) used deep learning to extract the text features of tweets, image features and text information in images. However, these kinds of methods only stay at the most basic surface features and cannot extract high-level, abstract global features of rumors.