Since the advent of the Internet era, online social networks have become an indispensable part of our lives. Platforms such as Twitter, Facebook, and Sina Weibo, which focus on social networking or possess social networking attributes, have become primary channels for people to access and share information on a daily basis. However, the exponential growth of content on social media platforms has been accompanied by a proliferation of rumors (fake information), which has had a detrimental impact on the online social environment [1]. The widespread dissemination of rumors distorts facts, leading individuals toward erroneous positions and thereby undermining the public opinion within social networks and posing a serious threat to society [2].

Detection methods and intervention strategies for rumors on social networking platforms have received considerable attention. Facebook encourages users to actively flag suspicious information, while Sina Weibo has established a dedicated Weibo Community Management Center to handle user reports of fake information. However, these existing approaches rely solely on manual verification which, although typically accurate, is limited in effectiveness due to the complexity of the identification process and the constraints of human resources in practical application. Consequently, an increasing number of researchers have been dedicating efforts to developing algorithms for detecting rumors, with the aim of automatically identifying rumors on the internet and addressing the challenges posed by the overwhelming volume of rumors that surpass the capacity of manual verification.

Early automatic rumor detection methods primarily relied on traditional machine learning. Researchers utilized feature engineering to model information from various dimensions of rumor events, followed by supervised training of classifiers to classify rumors and non-rumors. For instance, the authors of [3] employed decision trees, those for [4] utilized random forests, and the authors of ^[5][6][5,6] employed support vector machines (SVMs). These methods demonstrated certain rumor detection capabilities but heavily depended on feature engineering, thus exhibiting noticeable limitations. In recent years, however, deep neural networks (DNNs) have gained popularity, eliminating the need for intricate feature engineering. By training on raw data alone, DNNs can achieve optimal performance, making them widely applicable in the field of rumor detection. For instance, the authors of [7] employed recurrent neural networks (RNNs) to learn the textual content of rumors, while those for [8] utilized convolutional neural networks (CNNs) to extract textual information. Furthermore, with the development of graph neural networks (GNNs) and RNNs, effective modeling of the spatiotemporal features of rumors has become feasible. For instance, the authors of [9] used a tree-structured recursive neural network for propagation feature extraction, and those for ^[10][11][10,11] employed graph convolutional neural networks (GCNs) for structural feature extraction. The authors of [12] captured text and propagation features using a graph encoder and decoder model, and those for [13] utilized gated recurrent units (GRUs) to extract a rumor’s temporal and propagation features. These methods have demonstrated excellent performance in rumor detection tasks.

Despite the effective progress made in previous work, several issues still remain. First, existing methods primarily utilize GNNs based on message-passing frameworks to learn the structural features of rumors. However, the propagation of rumors follows a tree structure that unfolds based on time and interaction relationships, with the information source serving as the root node. The connectivity within rumor propagation graphs is relatively simple, but the depth exceeds that of typical graphs. Figure 1 illustrates the distinction between typical graphs and rumor propagation graphs. The characteristics of rumor propagation graphs pose limitations on the application of existing GNN frameworks for rumor detection tasks. Specifically, when faced with deeper node relationships, traditional GNNs can only aggregate information from high-order neighboring nodes by stacking multiple layers. However, this approach leads to issues such as oversmoothing of node features and gradient vanishing, resulting in performance degradation of the network ^[14][15][14,15]. Therefore, a challenge lies in how to better extract the interaction relationships among multi-hop neighbors within rumor propagation graphs.

Furthermore, the current approaches for extracting the temporal features of rumors, whether based on dynamic graphs ^{[16][17][18][19]}[16,17,18,19] or time series ^[20][21][20,21], only focus on learning features at the level of the original semantic information. However, some studies have indicated that word embeddings, which are used to represent semantic information, possess certain distinctive properties. By analyzing arithmetic operations on word embeddings, the authors of [22] discovered that certain word embedding models can encode linguistic relational patterns. Moreover, the authors of ^[23][24][25][23,24,25] conducted case studies on the meanings of individual neurons in word embeddings and found systematic distributions of different linguistic attributes within the embeddings. This allows to consider word embeddings as relatively stable signals and obtain their changing information through differential operation. The advantage of modeling differential time series explicitly is that rumor features can be extracted from a perspective that varies in time series.

2. Text-Based Rumor Detection

Text is the core feature of rumors. In the era of traditional machine learning, researchers primarily relied on a series of feature engineering techniques to extract information such as lexical features, symbolic features, and sentiment features from text. For instance, the authors of [3] subdivided text features into string length, presence of emoticons, and personal pronouns. The authors of ^[26][27] incorporated the word distribution ratios of rumor and non-rumor information as text features. The authors of ^[27][28] included features such as tags, links, and questions present in the text. However, these methods only utilized shallow information and had limited generalization capabilities. Subsequent deep learning algorithms overcame the limitations of traditional machine learning and enabled modeling of deep semantic information. The authors of [7] used RNNs to capture long-range dependencies in text. The authors of [8] employed CNNs to extract deep features from text. The authors of ^[28][29] introduced a word-sentence-document structure to extract hierarchical text features while preserving the text’s structural hierarchy. Attention mechanisms automatically capture the dependencies between words, giving them a significant advantage over CNNs or RNNs in modeling text content. Consequently, attention mechanisms have been employed to model tweet information in rumors by researchers, such as those in ^[29][30][30,31]. Building upon this, some scholars ^[31][32][32,33] recognized that different domains have distinct linguistic expression forms and incorporated domain-specific terminologies into text features. Additionally, other researchers ^{[33][34][35][36]}[34,35,36,37] noted the presence of emotional and thematic information in rumor events and extracted features such as sentiment and topics from text for rumor detection tasks.

3. Propagation-Based Rumor Detection

In order to effectively capture the multidimensional features of rumors and achieve better detection performance, recent works have focused on exploring the differences between rumors and non-rumors in the propagation process. They have modeled events from the perspective of propagation, including structural and temporal aspects, to achieve more accurate identification. The authors of ^[37][38] modeled rumor propagation as a propagation tree and employed kernel learning to extract features from the propagation tree. Similarly, the authors of [9] modeled rumor propagation as a tree and used recursive units to learn propagation features in a top-down and bottom-up manner. The authors of [13] considered the influence of temporal relationships based on the tree structure and proposed a deep spatiotemporal network to simultaneously learn the structural and temporal features of rumors. The authors of [20] combined an RNN with attention mechanisms to capture the contextual changes of semantic information over time in events. The authors of [21] modeled rumors as dynamic time series over time and used GRU units to learn temporal information. Apart from using RNN architectures to extract propagation features, another mainstream approach is to model events as graphs and utilize graph neural networks under the message-passing framework to learn the propagation features of rumors. For instance, the authors of ^[38][39] proposed a GNN-based semi-supervised method for fake news detection. the authors of [11] employed a bidirectional graph convolutional network to learn the propagation and aggregation structures of rumors and included a root node enhancement mechanism in each GCN layer to strengthen the influence of the rumor source on the entire rumor event. The authors of ^[39][43] proposed source identification based on graph convolutional networks, using spectral domain convolution to obtain the multi-hop neighbor information of nodes and locate multiple rumor sources without prior knowledge of the underlying propagation model. In addition to the aforementioned methods based on homogeneous graphs, the authors of [40] modeled the global relationships among all source tweets, retweets, and users as a heterogeneous graph to capture richer structural information. The authors of [41] constructed a word-user heterogeneous graph based on the textual content of rumors and the propagation of source tweets, and they proposed a heterogeneous graph attention network framework based on metapaths to capture the global semantic relationships of text content and global structural information of source tweet propagation. The authors of [42] introduced the concept of a joint graph to integrate the propagation structure of all tweets and mitigate sparsity issues, and they utilized network embeddings to learn the representations of nodes in the joint graph. Considering that static graph structures cannot model the temporal features of rumor propagation, recent research has extended events to dynamic graph structures. The authors of [16] represented rumor posts and their response posts as discrete dynamic graphs and used graph snapshot representation learning with attention mechanisms to capture the structural and temporal information of rumor propagation. The authors of [17] introduced a novel framework for fake news detection based on temporal propagation, modeling the temporal evolution patterns of real-world news as graph evolution patterns under continuous time dynamic diffusion network settings. The authors of [18] modeled each news propagation graph as a series of graph snapshots recorded at discrete time steps and used GCN and attention mechanisms to extract temporal information. The authors of [19] proposed a dual dynamic graph convolutional network that models the dynamic information in message propagation and the dynamic information in the knowledge graph background, learning the two types of structural information in a unified framework.