As social media spreads globally, the challenge of detecting misinformation and rumors has intensified. Temporal-domain-based methods, particularly those using time series analysis, have gained significant attention for identifying rumors by analyzing the propagation patterns of information.

For instance, one approach [17] uses Recurrent Neural Networks (RNNs) to capture the erratic, burst-like patterns typical of rumor spread, contrasting with the steady progression of truthful information. This innovation lies in applying RNNs to model the specific temporal dynamics of rumor propagation, offering an adaptive detection mechanism.

Another method [18] combines time series data with user behavior analysis, examining reposting rates and the speed of information spread. Rumors tend to have faster, more volatile lifespans in early propagation stages, which serves as a key indicator. The novelty of this approach is integrating both temporal data and user behavior to improve detection robustness.

Additionally, integrating social network features with time series analysis has proven effective for early-stage rumor detection [19]. This approach analyzes both temporal behavior and social interactions, offering a more nuanced understanding of rumor propagation. Its contribution lies in combining these two dimensions to enhance early detection accuracy.

Furthermore, multi-modal models, such as the Capture, Score, and Integrate (CSI) framework [7], combine content, social context, and temporal features for comprehensive detection. By incorporating emotional tone, user behavior, and timing, these models provide more robust mechanisms for rumor identification. The novelty here is the multi-dimensional approach, leveraging a wide range of features for enhanced detection.

Overall, temporal-domain-based methods offer a powerful tool for identifying rumors by analyzing the dynamic patterns of information propagation within social networks.

Spatial-domain-based rumor detection methods focus on analyzing the structural features of information dissemination networks. These methods use graph theory and network topology to examine the relationships between nodes (users) and identify potential rumors.

One approach [20] constructs a graph of information propagation and applies graph embedding techniques to capture the network’s topological features. By analyzing node connectivity and subgraph structures, this method can distinguish between rumors and truthful information. Research shows that rumors often form densely connected subgraphs, which are key indicators of misinformation. The novelty lies in focusing on these subgraphs to track rumor spread.

Another approach [21] uses social network graph models to analyze node interactions, particularly the role of influential nodes (e.g., opinion leaders) in rumor spread. Studies show that rumors tend to spread rapidly among these key nodes, with concentrated dissemination paths. The novelty of this method is its focus on influential node dynamics, which helps predict and control rumor spread more effectively.

Additionally, Graph Convolutional Networks (GCNs) have been used [14] to analyze spatial features of social networks. GCNs examine the relationships between neighboring nodes to detect rumors by analyzing information flow and interaction patterns. This method is particularly effective in detecting subtle signals of rumor spread in complex networks. The novelty of GCNs lies in their ability to handle large, intricate networks, making them highly effective for real-time rumor detection.

Overall, spatial-domain-based methods use graph theory and network structures to identify rumors by analyzing how information propagates and how key nodes interact, offering a powerful tool for detecting misinformation.

Large language models (LLMs), such as GPT-3 and its successors, have revolutionized applications in natural language processing and conversational AI. However, their ability to generate highly realistic, human-like text poses significant challenges for misinformation detection [22]. As LLMs produce content that closely mimics human writing, traditional detection methods, which rely on linguistic features like structure and syntax, often fail to distinguish between machine-generated and human-written misinformation [23,24].

Several studies have explored using LLMs in misinformation detection. Yang et al. [25] utilize GPT-3.5 to extract entities and build relational graphs, enhancing the identification of false information. While effective, it still faces challenges in accuracy and scope. Hu et al. [26] highlight a key limitation: fine-tuned, task-specific models for fake news detection outperform general-purpose LLMs, suggesting that LLMs’ versatility may hinder their efficiency in detecting misinformation.

The growing sophistication of LLMs has spurred interest in hybrid models that combine their capabilities with specialized detection mechanisms. These models aim to improve detection at scale, but LLMs still struggle with capturing nuanced context, addressing adversarial manipulation, and adapting to new domains without fine-tuning. As LLMs continue to evolve, further research is needed to enhance their effectiveness in misinformation detection.

In conclusion, while LLMs offer advances in natural language generation, their use in misinformation detection remains complex. Current efforts leverage LLMs alongside other models, but overcoming their limitations will require further innovation.

The overall structure of the method in this paper is shown in Fig. 2. The proposed rumor detection model consists of five main components, dynamic graph construction, encoding text contents, spatial feature extraction, temporal feature extraction and classification with MLP.

First, we construct a dynamic propagation structure graph of events, and then use graph convolutional neural networks to extract spatial features of propagation structure. Then we use the spatial domain features as input and use the recurrent neural network to extract the temporal domain features. At this time, we get a representation of the fusion of spatial domain and temporal domain features. Then, we integrate the text features of the source microblog to enhance the final feature representation. Finally, MLP is used to classify microblog events.

In this subsection, we construct a dynamic propagation structure graph of events. Unlike ordinary propagation structure graph models, which only build a single complete propagation structure graph, we construct multiple propagation graphs based on time series.

The choice of dynamic graph construction is motivated by the need to capture the evolving nature of rumor propagation over time. Rumors spread through social networks are not static; they evolve as new comments, reposts, and interactions occur. By modeling this progression dynamically, we can better understand how rumors unfold in real-time, allowing us to capture more nuanced patterns of rumor spread.

As shown in Fig. 3, the nodes represent the source microblog and its comments in the event, and the edges between the nodes represent their comment relationships. For example, r1 commented on r0 at time t2, so there is a directed edge between them. As time progresses, the propagation structure graph evolves dynamically, with the network’s structural features becoming increasingly rich. For instance, at time t0, the propagation graph contains only the source tweet as a single node, whereas by time tn, the graph incorporates the complete set of propagation structure features.

In this view, we model each event as a set of directed graphs G={G0,G1,…,Gn}, where n represents the total number of time periods. Each graph Gi=(Vi,Ei,Ai), where Vi={r0,r1,…,rm} is the node set with r0 as the root node. Each node represents a microblog in the process of the event propagation, and m is the total number of microblogs involved at time ti. Ei={est|s,t=0,…,m} is the edge set. If there is a forwarding or commenting relationship between the two microblogs rs and rt, there is an edge est. For example, if r2 is the reposted microblog of r1, there is a directed edge, r1→r2, i.e., e12. The adjacency matrix of Gi can be defined as follows: (2)Ai=(0ai(1,2)⋯ai(1,m)ai(2,1)0⋯ai(2,m)⋮⋮⋱⋮ai(m,1)ai(m,2)⋯0)

The element ai(s,t) of the s-th row and t-th column in Ai is defined as follows: (3)ai(s,t)={0,else1,est∈Ei

That is, when est∈Ei, the corresponding element value of the s-th row and t-th column of the adjacency matrix Ai is 1, otherwise it is 0.

We use text content features as initialization features of nodes in G, so in this section we will describe the method of text content encoding. We adopt Sentence-BERT (SBERT) [27] to encode the text content feature. SBERT is a pre-trained model for encoding text content features. This method is fine-tuned on Bidirectional Encoder Representations from Transformers (BERT) using Siamese and three-level network structure, which improves the traditional sentence embedding method and achieves very good results in multiple application scenarios.

SBERT adds a pooling operation to the output of BERT to derive a fixed sized sentence embedding. The pooling strategy in this paper is MEAN pooling.

The propagation structure graph at time ti is Gi, the node set is Vi={r0,r1,…,rm}, each node rj is a microblog, and the corresponding text content is wj, the feature pj∈R1∗d of each node is extracted by the SBERT model: (4)pj=SBERT(wj)

d is the dimension of the initialization feature vector, and all the node features in the set Vi are combined to obtain the initialization feature matrix Ti of the propagation structure graph Gi: (5)Ti=(p0p1⋮pm−1pm)

We employ a two-layer graph convolutional network to extract spatial feature from the dynamic graphs. Graph Convolutional Network (GCN) is an extension of CNN on graph data, which can effectively capture graph features by aggregating the neighborhood information of nodes in the graph.

For each graph in different time period ti, we use two-layer graph convolutional networks to extract the spatial features. The calculation formula is as follows: (6)H1i=σ(A^iTiW0i) (7)H2i=σ(A^iH1iW1i)where A^i=Ai+I, I is the identity matrix, H1i∈Rm∗v0, H2i∈Rm∗v1 represent the outputs of the first and second graph convolutional layers (GCL), namely the hidden state. m is the total number of nodes at time ti, v0 is the output vector dimension of the first layer, and v1 is the output vector dimension of the second layer. W0i∈Rd∗v0 and W1i∈Rv1∗v1 are parameter matrices GCL. σ(⋅) is the activation function, we use the ReLU function here.

Finally, through average pooling, as shown in Eq. (13), we get the final spatial features. (8)Si=meanpooling(H2i)

In this part, we use a recurrent neural network to extract the temporal features of rumor propagation. We take the spatial features {S0,S1,…,Sn} as input and pass the spatial feature representation St to an RNN unit.

This article compares three different RNN units: basic RNN, LSTM, and GRU through experiments, and finally finds that GRU has the best effect. Its calculation formula is as follows: (9)rt=σ(WirSt+bir+Whrh(t−1)+bhr) (10)zt=σ(WizSt+biz+Whzh(t−1)+bhz) (11)nt=tanh(WinSt+bin+rt⊙(Whnh(t−1)+bhn)) (12)ht=(1−zt)⊙nt+zt⊙h(t−1)where ht is the hidden state at time t, h(t−1) is the hidden state at time t−1. rt, zt, nt are the reset, update, and new gates, respectively. σ is the sigmoid function, and ⊙ is the Hadamard product.

The output hn of the last unit is the result of fusing spatial and temporal features. Finally, we concatenate the source Weibo text feature representation with hn to enhance the feature representation of rumors: (13)D=concat(SBERT(w0),hn)

The predicted label y^ of the event is calculated by multiple fully connected layers and a softmax layer: (14)y^=softmax(FC(D))where y^∈R(1∗K) is a probability vector, K is the number of classes, and the value y^i of each element of y^ represents the probability that the event belongs to the corresponding class.

We train all parameters in the model by minimizing the cross-entropy between the predicted result and the ground truth Y of all events, and add the L2 regular term during the training process to avoid the overfitting problem. The loss function L is defined as: (15)L=∑|C|∑i∈{0,1}−yilogyi^+βL2where β is the coefficient of the regular term and |C| is the total number of events.

Although our rumor detection module can accurately identify rumors, effective curbing of their spread is difficult without providing reasonable explanations. To address this, we utilize a LLM to offer coherent explanations for the rumors. However, in real social scenarios, the number of comments can be quite large, while LLMs have limitations on input length. Furthermore, LLMs often struggle to handle excessively long or redundant information [28,29].

In recent research on LLMs [30–32], complex tasks are typically decomposed into a series of simpler tasks, which significantly enhances the ability of LLMs to handle complexity. Accordingly, we delineate the prompts into a Chain-of-Propagation (CoP), facilitating easier reasoning for the LLMs. Given that comments within a certain timeframe often share similar sentiments, we construct the CoP based on the chronological order of comment timestamps, as illustrated in Fig. 4.

We divide all the comments r1,r2,…,rn in a given event ci into m groups, each containing k comments. In the inference prompt for the LLMs, we select only one group at a time as the input (as shown by the green text in the prompt of Fig. 4). After m rounds of inference, we obtain the final result. The inference steps within the CoP operate in the same session of the LLM, allowing subsequent steps to reference the results of prior steps. We utilize the output from the final reasoning step as the conclusive result, as it aggregates information from all preceding steps. In each round of the inference prompt, the source tweet r0 of the event is also included (as shown by the blue text in the prompt of Fig. 4), along with the rumor detection result (as shown by the red text in the prompt of Fig. 4).

We benchmark our proposed method against state-of-the-art baselines using two public datasets: WEIBO [33] and PHEME [34]. The WEIBO dataset, compiled from Xinhua News Agency and Weibo, comprises a rich collection of data. Similarly, the PHEME dataset is an aggregation of content centered around five distinct breaking news stories, with each story featuring a collection of associated posts. Adhering to the methodology outlined in [35], both the WEIBO and PHEME datasets are segmented into training and testing subsets with an 8:2 ratio. Table 1 shows the statistics of the datasets.

Our hardware environment is configured with a Hygon C86 3285 8-core processor, 128 GB of memory, and an NVIDIA RTX A6000 graphics card. We implemented our rumor detection method using the PyTorch framework, with parameter optimization using the Adam algorithm. In our model, we utilized the pre-trained SBERT: all-MiniLM-L6-v2 and selected the Llama-3-8B [36] large language model as the rumor result interpreter.

In terms of model parameter settings, the output feature dimension of SBERT is 512. The GCL layer uses two layers of GCN, with the first layer having an input feature dimension of 512 and an output feature dimension of 256, while the second layer has an output feature dimension of 50. In the RNN layer, we select a single GRU layer as the temporal feature extraction model, with a hidden layer feature dimension of 100. The classifier has an input feature dimension of 100 and an output dimension corresponding to the number of classes, which is 2.

We frequently employ Accuracy as the primary evaluation metric for binary classification tasks, including the detection of fake news. Nonetheless, the reliability of Accuracy is significantly undermined when dealing with datasets that exhibit class imbalance. To address this limitation, our experimental framework incorporates a suite of complementary metrics alongside Accuracy. Specifically, we introduce Precision, Recall, and the F1 score to provide a more nuanced and comprehensive assessment of our model’s performance in the context of rumor detection.

We compare the following baseline models with our model:

SVM-TS [37]: SVM-TS detects fake news using heuristic rules and a linear SVM classifier.

Table 2 presents the experimental outcomes for our approach and the baseline methods. Key observations include:

(1) Across all datasets, SVM-TS underperforms, suggesting that manually engineered features may be inadequate for fake news detection.

(2) Deep learning models (CNN, GRU, RumorGAN, GACL) surpass SVM-TS, highlighting their advantages over conventional techniques.

(3) Our method is superior to other methods in both data sets, which proves the effectiveness of feature fusion in temporal domain and spatial domain.

We did an ablation study to see whether each module contributes to the model and which modules contribute more. Our model’s main components include SBERT, GCN, and GRU. Based on the complete model, we systematically remove the module and compare their changes in accuracy, precision, recall, F1 value. Fig. 5 shows the experiment results.

The ablation study reveals that the performance of the model is significantly reduced when any of the components (SBERT, GCN, or GRU) are removed. This confirms that each module makes a meaningful contribution to the model’s overall ability to detect rumors accurately.

Full Model. The complete model, which integrates SBERT, GCN, and GRU, achieved the highest performance across all metrics. This demonstrates that combining textual, spatial, and temporal features results in a robust and accurate rumor detection system.

Next, we evaluated the impact of several key parameters in our method, with validation results shown in Fig. 6. We selected three parameters for validation: RNN type, learning rate, and number of iterations.

The choice of these hyperparameters was guided by several considerations. First, the RNN type (such as LSTM, GRU, or Simple-RNN) was chosen based on its ability to capture the temporal dynamics of rumor propagation, with LSTM and GRU generally offering better performance in handling long-term dependencies in time series data. Second, the learning rate was selected as a crucial factor influencing the convergence speed and stability of the model. Lastly, the number of iterations was determined by balancing model training time and accuracy. Too few iterations may result in underfitting, while too many could lead to overfitting or unnecessary computational overhead.

For these three parameters, we used cross-validation to assess performance on the Weibo dataset, which served as a case study in our experiments. To ensure the robustness of the results, we kept the other parameters fixed at their optimal values while adjusting one parameter at a time. This allowed us to isolate the effects of each parameter and find the best configuration for the task at hand. The final hyperparameter values were selected based on the validation performance, ensuring the model achieved the best possible balance between accuracy and computational efficiency.

In Fig. 6, the x-axis represents different types of RNNs, learning rates, and iteration epochs, while the y-axis displays the values of accuracy, precision, recall, and F1 score. The curves in different colors represent the variation trends of different evaluation metrics.

After obtaining the model’s classification results, we used a LLM to explain these results, particularly for events classified as rumors by the model. We chose Llama-3-8B as the interpreter and selected a real tweet from the Weibo dataset as an example. The tweet’s main content discussed China’s supposed plan to implement a “polyandry” system. Table 3 shows the prompts and the outputs.

We employed CoP method, where multiple steps of prompts were used to obtain the final output. The selected example contained 143 reposts or comments, with 113 remaining after removing empty entries. We set the number of comments included in each reasoning step to K = 5, resulting in a total of 23 reasoning steps. Table 3 includes the first two reasoning steps’ prompts and outputs, as well as the final model output.

In Table 3, the sentences highlighted in blue represent the tweet, the green sentences represent comments, and the red sentences provide evidence that the event is a rumor. From the red text, we can observe that as the reasoning steps progressed, the amount of evidence in the output increased, while irrelevant sentences decreased.

In the first reasoning step, only one sentence was able to explain why the tweet was a rumor, and it didn’t consider the comment information. However, in the second reasoning step, there were two sentences serving as evidence for the tweet being a rumor, and the model began to take comments into account, using them to explain why the tweet was classified as a rumor. After 23 reasoning steps, the final output provided a comprehensive explanation of the rumor, covering four key points: 1) It did not follow proper policy-making procedures; 2) It lacked legal basis; 3) There was limited reaction in the comments; 4) It contradicted China’s social values.

The advantage of this method lies in its ability to automatically generate highly persuasive explanatory text for rumors, without requiring manual refinement of language. Additionally, it is simple to use, requiring only carefully designed prompts. However, its limitation is that it cannot reveal the internal mechanisms of the model’s prediction algorithm.