A continuous-time dynamic graph processing method for rumor identification

Abstract

This invention discloses a continuous-time dynamic graph processing method for rumor identification, comprising the following steps: Step 1, constructing a temporal propagation graph; Step 2, constructing a temporal knowledge graph; Step 3, using the constructed temporal propagation graph and temporal knowledge graph as input, and fusing them respectively using a time-space feature fusion module to generate two node embedding sets; Step 4, using the two node embedding sets generated in Step 3 as input, and using a subgraph fusion module to generate a node embedding set that fuses the structural information of the two subgraphs; Step 5, using the node embedding set generated in Step 4 as input, and using a classification module to classify the input to obtain the final prediction result. This technique can obtain better node embeddings and has high efficiency in rumor detection.

Description

Technical Field

[0001] This invention relates to rumor identification technology, specifically a continuous-time dynamic graph processing method for rumor identification. Background Technology

[0002] With the rapid development of the Internet, we have witnessed the emergence of many online social media tools in the past decade, such as Twitter, Facebook, Weibo, and Instagram. These online social media tools have gradually become the main source of information in people's daily lives and have fundamentally changed the way people share information [1–4]. However, online social media tools are also a double-edged sword. On the one hand, they have created social connections during social distancing and facilitated the dissemination of knowledge in various contexts. On the other hand, they may lead to people sharing quick and superficial ideas and spreading rapidly, such as rumors. Rumors are defined as unverified or proven false information [5,6]. The explosive spread of rumors threatens the credibility of the Internet and has serious adverse effects on individuals and society. The severity of rumor generation has spurred the need for effective detection of misinformation using online social media tools [8]. However, the short and rapid nature of rumors makes them difficult to detect automatically. Therefore, automatic rumor detection has attracted increasing attention from researchers.

[0003] Early rumor detection methods mainly relied on manual feature selection. The selected features focused on extracting various lexical and semantic features from the original and forwarded messages [5], and relied on traditional machine learning methods to learn models from labeled data [5]. Machine learning-based methods require a lot of labor and time to select appropriate features. Therefore, researchers have introduced deep learning methods into the problem of social media rumor detection. Deep learning methods have powerful feature learning capabilities. They can learn more accurate and representative features of the original data than traditional machine learning algorithms obtained through feature engineering, thus achieving better classification results.

[0004] Recent studies have shown that modeling propagation as a propagation tree or graph can effectively identify rumors. Huang et al. [9] used a tree-like recurrent neural network (RNN) as the architecture for rumor detection and designed two ways to encode the propagation structure: bottom-up and top-down. Bian et al.

[10] modeled propagation as a graph structure and used a bidirectional graph convolutional network (Bi-GCN) to encode the propagation. However, these spatial structure-based methods often ignore or simplify the temporal structure of rumor propagation.

[0005] The temporal structure of rumor propagation is characterized by the time series of comments or reposts, which can further distinguish the propagation form [11,12]. Ma et al.

[13] input the source post and reposts into an RNN in chronological order to capture the contextual information of the relevant posts. Sun et al.

[14] treated the propagation process as multiple ordered static snapshots and encoded the temporal information by inputting these snapshots sequentially into a graph convolutional network (GCN).

[0006] Meanwhile, another parallel research approach has introduced external knowledge related to rumor identification [15–18]. These studies typically utilize knowledge related to the content of the source post to expand its semantics, thereby improving the quality of the post representation. The basic logic is that posts are usually short and contain many abbreviations, proper nouns, nicknames, etc., which traditional classification models often struggle to understand. Knowledge bases include background knowledge corresponding to these unique terms, and adding this knowledge to the text representation can greatly help the model determine the credibility of the post.

[0007] Despite significant progress in existing deep learning-based rumor detection methods, these methods still face several limitations:

[0008] Lack of joint learning of spatiotemporal structure: Most rumor detection methods tend to focus only on the spatial structure of propagation [9,10] or the temporal structure [11–13], while few models can effectively learn the spatial and temporal structural features together.

[0009] Coarse temporal granularity: Existing time-based rumor detection methods either input posts into an RNN in chronological order [9,13], or treat the dynamic propagation process as a discrete-time dynamic graph (DTDG), i.e., a set of ordered static graph snapshots, and input each static graph into a graph neural network (GNN) to simulate temporal features

[14] . However, these methods do not represent temporal information in a sufficiently specific or detailed way. Since the propagation of posts has a definite time and is a constantly changing process, the propagation process of posts can be modeled as a continuous-time dynamic graph (CTDG). Figure 3 As shown, CTDG can represent the propagation process of a post with precise time, thus providing accurate temporal information. In contrast, DTDG can only extract the static propagation structure over time intervals, resulting in a relatively coarse-grained time series.

[0010] Lack of background knowledge on comments: User comments often contain a wealth of useful information, and semantic expansion of user comments can help detect rumors. Most existing knowledge augmentation methods only use background knowledge related to the source post, ignoring background knowledge related to user comments [15–18]. At the same time, knowledge information also spreads with dissemination, and existing methods ignore the dynamic structure of knowledge. Summary of the Invention

[0011] The purpose of this invention is to address the shortcomings of existing technologies by proposing a continuous-time dynamic graph processing method for rumor identification. This technique can achieve better node embedding and has high efficiency in rumor detection.

[0012] The technical solution to achieve the technical objective of this invention is:

[0013] A continuous-time dynamic graph processing method for rumor identification includes the following steps:

[0014] Step 1: Using the source post and the text of the reply posts of an event as input, construct a time-series propagation graph;

[0015] Step 2: Using the text of the source post and reply posts of an event as input, use TagMe to link entities to the input, linking the entities mentioned in the source post and reply posts to the relevant entities in YAGO, obtaining an entity set. For each identified entity in the entity set, obtain the corresponding concept of the identified entity in the existing knowledge graph through conceptualization, obtaining a concept set. Based on the entity set and concept set obtained above, construct a temporal knowledge graph.

[0016] Step 3: Using the temporal propagation graph and temporal knowledge graph constructed in Step 1 and Step 2 as input, the temporal-spatial feature fusion module is used to fuse the input temporal propagation graph and temporal knowledge graph respectively to generate two node embedding sets;

[0017] Step 4: Using the two node embedding sets generated in Step 3 as input, use the subgraph fusion module to generate a node embedding set that fuses the structural information of the two subgraphs.

[0018] Step 5: Using the node embedding set generated in Step 4 as input, the classification module is used to classify the input to obtain the final prediction result.

[0019] The specific process of step 1 is as follows: for an event ε i Its time-series propagation diagram is represented as follows Among them, node set Includes source posts and reply posts, edge set In the set of edges, the edges represent interactions between the source post and the reply post, or interactions between reply posts. For simplicity, the direction of the edges in the edge set is not considered. Represented as an undirected graph, and initialized with embedding vectors. Each node in the representation.

[0020] The specific process of step 2 is as follows: for an event ε i Its time-series knowledge graph representation is as follows Among them, node set It is a time-series propagation graph node set Entity set E i and concept set C i The union of the sets, initialized with embedding vectors. Each node in the representation follows the following rules for edge construction: (1) Post-entity edge: if One of the posts contains links to E i (1) For words in the entity, an edge will be added between the post node and the entity node; (2) Entity-entity edge, entity-concept edge, concept-concept edge: Use a fixed-size sliding window to count the relevance of nodes in the global corpus, then calculate the PMI score between node pairs, keep the edges with positive PMI values, and remove the edges with non-positive PMI values.

[0021] In step 3, when using the time-space feature fusion module, for event ε i Based on ε i The constructed temporal propagation graph and temporal knowledge graph will be processed through two time-space feature fusion modules to obtain the final time t. m Two sets of node embeddings: and

[0022] In step 3, for the process of generating node embeddings by the time-space feature fusion module, the time-space feature fusion module has the following core components:

[0023] Memory component: The memory component records the historical information of the nodes. At time t, the vector s of each node constitutes the memory. The memory of the node will be updated after the event occurs. In this way, the time-space feature fusion module can remember the long-term dependencies of each node in the temporal propagation graph and the temporal knowledge graph. If a new node appears, the memory of the new node is initialized to a 0 vector. When an event involving the new node occurs, the memory of the new node will be updated.

[0024] Post function: Each interaction event containing nodes i and j corresponds to two posts.

[0025] p i (t)=psg s (s i (t - ),s j (t - ),Δt,e ij (t))(1)

[0026] p j (t)=psg d (s j (t- ),s i (t - ),Δt,e ij (t))(2)

[0027] An event containing only node i will correspond to one post:

[0028] p i (t)=psg n (s i (t - ),t,v i (t))(3)

[0029] Where s i (t - ) represents the memory of node i before time t, psg s ,psg d and PSG n Each layer consists of three fully connected layers;

[0030] Post aggregator: Uses aggregation functions to aggregate t1,…,t b p at time ≤ t i (t1),…,p i (t b ):

[0031]

[0032] And average pooling is chosen as the agg(·) function;

[0033] Post updater: The post updater updates itself based on the node's memory and event messages.

[0034]

[0035] When an interaction event occurs between node i and node j, the memories of both nodes will be updated after the event occurs. For a single node event, only the memory of the involved node is updated, and the memory of other nodes is not updated.

[0036] Embedding Component: The embedding component generates the temporal embedding of node i at time t. The embedding component contains a series of graph attention layers that compute the embedding of node i by aggregating information from its L-hop temporal neighborhood. The input to the l-th graph attention layer includes the representation of node i in the upper layers. Neighbor representation of node i at current time step t Timestamps t1, ..., t N And the feature e of the edge of node i at each time step. i1 (t1),…,e iN (t N ):

[0037]

[0038] Here, || represents the concatenation operation, and φ(·) represents the time encoder; multi-head attention is performed at each layer, where the query q (l) (t) represents the reference node, and the key K (l) (t) and value V (l) (t) represents the query q (l) The neighbors of (t), the final MLP is used to fuse the features of node i in the previous layer. Features of node i after aggregating neighbor nodes The initial input for each node is represented by the current memory s. j (t) and time node features v j (t):

[0039]

[0040] Original Post Storage: The post embedding for each batch updates the information in memory using posts from previous batches stored in the original post storage, and generates node embeddings.

[0041] In step 4, the subgraph fusion module generates a set of node embeddings that fuse the structural information of two subgraphs using a subgraph-level attention mechanism. To learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert it into a low-dimensional vector. Then, the similarity between each node and the subgraph attention vector q is calculated, and the similarity is used as the importance score of the node in the current subgraph. The average importance score of all nodes in the current subgraph is taken as the importance score of each subgraph. The average importance score of all nodes in the same subgraph is used as the importance score of the subgraph. The importance w of the temporal propagation graph and the temporal knowledge graph is calculated. TP(TK) The formula is as follows:

[0042]

[0043] Among them W sub This is the weight matrix, where b is the bias. At the subgraph level, pay attention to the vector q and the weight W. sub The bias b of the linear layer is shared between the two subgraphs, and a subgraph importance score w is obtained. TP(TK) Then, these scores are normalized into weighting coefficients:

[0044]

[0045] Finally, using the learned subgraph weight coefficients, a set of node embeddings that integrates information from both subgraphs is obtained.

[0046]

[0047] In step 5, the classification module first applies mean pooling to fuse the data. Structural information representation of nodes:

[0048]

[0049] Then, BERT is applied to encode the source post information to obtain p. i0 and p i0 With X T Then, stitch the images together and apply a linear layer for fusion.

[0050] S i =linear(BERT(p) i0 ||X T ))(16)

[0051] Finally, S i The input is fed into the classifier to obtain the prediction result:

[0052]

[0053] Training is guided by cross-entropy loss and regularization, as shown in the following formula:

[0054]

[0055] Where y i Let be the baseline truth label for the i-th event, λ represent the tradeoff coefficient, and θ represent the set of all parameters. This represents the L2 regularization term, used to prevent overfitting.

[0056] This technical solution employs a temporal-spatial feature fusion module to capture the semantic, spatial structure, and temporal information of content during post propagation. Instead of using a discrete-time dynamic graph (a snapshot of the time graph), this solution uses a continuous-time dynamic graph (precise temporal information) to model post propagation because post propagation in social networks is inherently continuous and evolves over time. Furthermore, this solution extracts conceptual knowledge related to the source post and individual comments from an external knowledge base and constructs a temporal-series knowledge graph that evolves over time. Similarly, this solution uses a temporal-spatial feature fusion module to encode the dynamic knowledge propagation process. Finally, this solution uses a subgraph-level attention mechanism to adaptively fuse dynamic post information and dynamic knowledge information, thereby achieving better rumor detection.

[0057] The main contributions of this technical solution to address the aforementioned limitations in the field of rumor detection are as follows:

[0058] Joint temporal and spatial structure: Node features are encoded using a temporal-spatial feature fusion module framework. A graph attention network (GAT) is used to fuse node neighbor information to encode the content semantics of posts and the spatial structure of propagation. Furthermore, temporal information is introduced as a feature during node updates, resulting in a node representation that integrates content semantics, spatial, and temporal information.

[0059] Combining fine-grained temporal features: Compared with existing methods, this technical solution uses a continuous-time dynamic graph to represent the propagation process of posts, utilizing precise temporal information rather than static graph snapshots. This makes the representation of the temporal propagation process more detailed.

[0060] Dynamic knowledge information: This technical solution not only focuses on the propagation process of posts, but also on the propagation process of knowledge. It utilizes an external knowledge base to construct a time-series knowledge graph that expands over time, and encodes it using a time-space feature fusion module framework to obtain dynamic knowledge information.

[0061] This technology captures spatial structure, continuous temporal information, external knowledge, and textual semantics through a unified framework. First, propagation graphs and knowledge graphs are constructed based on the dynamic evolution of post diffusion and background knowledge. Then, a temporal graph network framework is used to capture the continuous temporal and spatial structure during the propagation process. Finally, a subgraph-level attention mechanism is used to adaptively integrate the different information in the two subgraphs to create a representation for each post, resulting in better node embedding and high efficiency in rumor detection. Attached Figure Description

[0062] Figure 1 This is a structural diagram of an embodiment;

[0063] Figure 2 This is a flowchart illustrating the operation of the time-space feature fusion module in the embodiment.

[0064] Figure 3 The diagram illustrates how to model the propagation process of a post in the prior art as a continuous-time dynamic graph and a discrete-time dynamic graph. The left side shows the propagation of a post as a continuous-time dynamic graph, and the right side shows the propagation of a post as a discrete-time dynamic graph. Each Twitter icon represents a post.

[0065] Figure 4 The following are schematic diagrams of ablation experiments for the embodiments. Figure a shows the ablation experiment of Twitter 15, and Figure b shows the ablation experiment of Twitter 16.

[0066] Figure 5The following are schematic diagrams illustrating the rumor detection on Twitter 15 and Twitter 16 in the early stages of the embodiment, as the time taken or the number of posts increases. Figure a shows the rumor detection based on the time taken on Twitter 15, Figure b shows the rumor detection based on the number of posts on Twitter 15, Figure c shows the rumor detection based on the time taken on Twitter 16, and Figure d shows the rumor detection based on the number of posts on Twitter 16. Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0068] Example:

[0069] Problem definition:

[0070] The rumor detection dataset provides an event set ε = {ε1, ε2, ..., ε...} n}, where ε i Let represent the i-th event. Each event contains a source post and related reply posts. Define the source post s in the i-th event. i and replying to posts p i The set is Where m represents ε i The number of replies to posts, s i Can be used as p i0 Use it. You can also obtain information related to the event ε. i Related release time series Where t i0 =0. and They are combined to obtain ε i ={(p i0 , t i0 ), (p i1 , t i1 ), ..., (p im , t im We need to learn a model f: ε → y to represent event ε. i The event is categorized into a predefined category y = {0, 1}, where 0 represents non-rumor and 1 represents rumor.

[0071] Reference Figure 1 A continuous-time dynamic graph processing method for rumor identification includes the following steps:

[0072] Step 1: Using the source post and the text of the reply posts of an event as input, construct a time-series propagation graph;

[0073] Step 2: Using the text of the source post and reply posts of an event as input, use TagMe to link entities mentioned in the source post and reply posts to relevant entities in YAGO, obtaining an entity set. For each identified entity in the entity set, obtain the corresponding concept of the identified entity in an existing knowledge graph (such as YAGO and Microsoft concept graph) through conceptualization, obtaining a concept set. Based on the entity set and concept set obtained above, construct a temporal knowledge graph.

[0074] Step 3: Using the temporal propagation graph and temporal knowledge graph constructed in Step 1 and Step 2 as input, the temporal-spatial feature fusion module is used to fuse the input temporal propagation graph and temporal knowledge graph respectively to generate two node embedding sets;

[0075] Step 4: Using the two node embedding sets generated in Step 3 as input, use the subgraph fusion module to generate a node embedding set that fuses the structural information of the two subgraphs.

[0076] Step 5: Using the node embedding set generated in Step 4 as input, the classification module is used to classify the input to obtain the final prediction result.

[0077] The specific process of step 1 is as follows: for an event ε i Its time-series propagation diagram is represented as follows Among them, node set This represents the set of source post embeddings and reply post embeddings, and the edge set. The edges included represent the interactions between the source post and the reply posts, as well as the interactions between the reply posts themselves. For simplicity, the edge set is not considered. The direction of the edges in the set will Represented as an undirected graph, and initialized with embedding vectors. Each node in the representation.

[0078] The specific process of step 2 is as follows: for an event ε i Its time-series knowledge graph representation is as follows Among them, node set It is a time-series propagation graph node set Entity set E i and concept set C i The union of the sets, initialized with embedding vectors. Each node in the representation follows the following rules for edge construction: (1) Post-entity edge: if One of the posts contains links to E i(1) For words in the entity, an edge will be added between the post node and the entity node; (2) Entity-entity edge, entity-concept edge, concept-concept edge: Use a fixed-size sliding window to count the relevance of nodes in the global corpus, then calculate the PMI (Point Mutual Information) score between node pairs, retain the edge with a positive PMI value, and remove the edge with a non-positive PMI value.

[0079] In step 3, when using the time-space feature fusion module, for event ε i Based on ε i The constructed temporal propagation graph and temporal knowledge graph will be processed through two time-space feature fusion modules to obtain the final time t. m Two sets of node embeddings: and

[0080]

[0081] like Figure 2 As shown, in step 3, for the process of generating node embeddings by the time-space feature fusion module, the time-space feature fusion module has the following core components:

[0082] Memory component: The memory component records the historical information of nodes. The vector s of each node at time t constitutes the memory. The memory of a node will be updated after an event occurs (e.g., interaction with other nodes or changes at the node level). In this way, the time-space feature fusion module can remember the long-term dependencies of each node in the temporal propagation graph and the temporal knowledge graph. If a new node appears, the memory of the new node is initialized to a 0 vector. When an event involving the new node occurs, the memory of the new node will be updated.

[0083] Post function: Each interaction event containing nodes i and j corresponds to two posts.

[0084] p i (t)=psg s (s i (t - ),s j (t - ),Δt,e ij (t))(1)

[0085] p j (t)=psg d (s j (t - ),s i (t - ),Δt,e ij (t))(2)

[0086] An event containing only node i will correspond to one post:

[0087] p i (t)=psg n (s i (t - ),t,v i (t))(3)

[0088] Where s i (t - ) represents the memory of node i before time t, psg s ,psg d and PSG n Each layer consists of three fully connected layers;

[0089] Post Aggregator: To improve efficiency, batch processing is used to reduce computation time. However, this can lead to multiple events involving node i within the same batch. To address this, aggregation functions are used to aggregate t1,…,t b p at time ≤ t i (t1),…,p i (t b ):

[0090]

[0091] And average pooling is chosen as the agg(·) function;

[0092] Post updater: The post updater updates itself based on the node's memory and event messages.

[0093]

[0094] When an interaction event occurs between node i and node j, the memories of both nodes will be updated after the event occurs. For a single node event, only the memory of the involved node is updated, and the memory of other nodes is not updated.

[0095] Embedding Component: The embedding component generates the temporal embedding of node i at time t. The embedding component contains a series of graph attention layers that compute the embedding of node i by aggregating information from its L-hop temporal neighborhood. The input to the l-th graph attention layer includes the representation of node i in the previous layer. Neighbor representation of node i at current time step t Timestamps t1, ..., t N And the feature e of the edge of node i at each time step. i1 (t1),…,e iN (t N ):

[0096]

[0097] Here, || represents the concatenation operation, and φ(·) represents the time encoder; multi-head attention is performed at each layer, where the query q (l) (t) represents the reference node (e.g., the target node or an L-1 hop neighbor), while the key K (l) (t) and value V (l) (t) represents the query q (l) The neighbors of (t), the final MLP is used to fuse the features of node i in the previous layer. Features of node i after aggregating neighbor nodes The initial input for each node is represented by the current memory s. j (t) and time node features v j (t):

[0098]

[0099] Original Post Storage: Since modules associated with memory do not directly affect the loss or receive gradients, to address this issue, the information in the memory is updated using posts from previous batches stored in the original post storage, and node embeddings are generated. The updated memory is then used to compute node embeddings, and new interactions generated after this interaction are stored in the original post storage.

[0100] In step 4, the subgraph fusion module generates a set of node embeddings that fuse the structural information of two subgraphs using a subgraph-level attention mechanism. To learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert it into a low-dimensional vector. Then, the similarity between each node and the subgraph attention vector q is calculated, and this similarity is used as the importance score of the node in the current subgraph. The average importance score of all nodes in the current subgraph is taken as the importance score of each subgraph. The average importance score of all nodes in the same subgraph is used as the importance score of the subgraph. The importance w of the time propagation graph and the time knowledge graph is calculated. TP(TK) The formula is as follows:

[0101]

[0102] Where W sub This is the weight matrix, where b is the bias. At the subgraph level, pay attention to the vector q and the weight W. sub The bias b of the linear layer is shared between the two subgraphs, and a subgraph importance score w is obtained. TP(TK) Then, these scores are normalized into weighting coefficients:

[0103]

[0104] Finally, using the learned subgraph weight coefficients, a set of node embeddings that integrates information from both subgraphs is obtained.

[0105]

[0106] In step 5, the classification module first applies mean pooling to fuse the data. Structural information representation of nodes:

[0107]

[0108] Then, BERT is applied to encode the source post information, and the source post information is compared with X. T To splice and merge:

[0109] s i =linear(BERT(p) i0 ||X T ))(16)

[0110] Finally, S i The input is fed into the classifier to obtain the prediction result:

[0111]

[0112] Training is guided by cross-entropy loss and regularization, as shown in the following formula:

[0113]

[0114] Where y i Let be the baseline truth label for the i-th event, λ represent the tradeoff coefficient, and θ represent the set of all parameters. This represents the L2 regularization term, used to prevent overfitting.

[0115] The entire training process in this example is outlined in Algorithm 1.

[0116]

[0117]

[0118] The performance of this example in rumor detection is compared with state-of-the-art baseline methods. Furthermore, an ablation experiment was designed to compare several variants of this example to investigate the impact of different components. The focus of this example is quantified as follows:

[0119] RQ1: How does this example compare to state-of-the-art rumor detection baseline methods?

[0120] RQ2: What is the impact of each component in this example on the overall performance?

[0121] RQ3: Is it possible to detect this in the early stages of rumor dissemination?

[0122] Experimental setup:

[0123] Dataset:

[0124] This example constructs experiments on two real datasets: Twitter 15 and Twitter 16

[35] . The labels for each event on Twitter 15 and Twitter 16 are based on the real labels of tweets on rumor websites. The Twitter dataset contains four labels: non-rumor (NR), false rumor (FR), true rumor (TR), and unverified rumor (UR). In this example, the two datasets are divided into training, validation, and test sets in a constant ratio of 7:1:2. Table 1 shows the statistics for these datasets.

[0125] Table 1 Statistical information of the dataset

[0126]

[0127] This example is compared with feature-based methods, deep learning methods, and recently proposed state-of-the-art methods.

[0128] DTR

[36] : A ranking model based on decision trees, which clusters posts and selects the top k cluster results as rumors.

[0129] DTC

[37] : A decision tree-based model that combines manually designed post features to distinguish rumors.

[0130] RFC

[38] : A random forest classifier that combines manual features such as content, structural attributes and propagation user attributes for detection.

[0131] SVM-TK

[22] : A support vector machine classifier based on Tree Kernel, which attempts to capture the structural features of propagation through kernel learning.

[0132] GRU-RNN: An RNN with GRU units that models the response time characteristics of comments.

[0133] TD-RvNN and BU-RvNN

[39] : A tree-structured recurrent neural network with GRU units, wherein the RvNN obtains its representation from a bottom-up (BU-RvNN) or top-down (TD-RvNN) propagation structure.

[0134] PLAN

[40] : A transformer-based rumor detection model for modeling long-distance post-to-post interactions.

[0135] Bi-GCN

[10] : A rumor detection method based on GCN, using a two-way propagation structure (propagation and diffusion structure) and text content (source posts and comments).

[0136] STS-NN

[11] : A model based on RNN that models spatial and temporal structures by modifying RNN.

[0137] GCRES

[41] : A deep residual graph convolutional network that detects rumors by fusing content and propagation structure and utilizing the interactive features of heterogeneous graphs to improve accuracy.

[0138] RA

[42] : A method for detecting errors based on a large language model (LLMS), which identifies whether a post matches any of nine predefined error message features.

[0139] LeRuD

[43] : An LLM-based approach that uses specific cues to guide the model to focus on key clues in news content and comments, dividing the dissemination of information into “dissemination chains” for stepwise analysis and reasoning.

[0140] Implementation settings:

[0141] This example uses the PyTorch framework, with parameters optimized using the Adam algorithm. BERT-base is used as the encoder for the source code and pre-trained on the dataset. The learning rate for BERT is set to 5e-5, while the learning rates for other parameters are set to 1e-3, the batch size is 16, and the training epochs are 5. This example selects the optimal parameter settings on the validation set and evaluates its performance on the test set. Accuracy(Acc) and the F1 score for each class are used to evaluate the model's performance across all categories.

[0142] RQ1: How does this example compare to state-of-the-art rumor detection baseline methods?

[0143] Table 2 lists the performance comparisons between this example and other baselines, from which the following observations were made:

[0144] Table 2 compares the performance of the Twitter 15 and Twitter 16 datasets. The best results are shown in bold, and the second-best results are shown underlined.

[0145]

[0146] Feature-based models, such as DTR, DTC, and RFC, perform poorly because they use hand-developed features based on overall post statistics. However, these features are too coarse and lack generality. RFC performs better than the other two models because it uses additional structural and temporal features.

[0147] Deep learning models, by using neural networks, can automatically extract effective features, significantly improving performance compared to feature-based methods. GRU-RNN performs relatively poorly because it only utilizes the content of posts, ignoring other useful information such as the propagation structure and temporal information of rumors. BU-RVNN and TD-RVNN outperform GRU-RNN, indicating that simulating spatiotemporal structure is effective for rumor detection. PLAN performs well on both datasets, demonstrating its effectiveness in capturing interactions between posts. STS-NN and Bi-GCN models capture the propagation structure and temporal information between source posts and comments. Bi-GCN outperforms STS-NN because it can distinguish spatiotemporal structural patterns. Furthermore, Bi-GCN does not include linguistic features, focusing instead on feature mining for rumor propagation and diffusion. As the number of nodes in the propagation process decreases, the information that can be provided decreases accordingly, leading to a decline in model performance. The strong performance of GCRES on baseline models demonstrates the effectiveness of using deep residual graph networks to learn content and structural interaction information in network propagation models.

[0148] LLM-based Relational Analysis (RA) and LeRuD methods offer a flexible and interpretable approach to rumor detection. They utilize cue engineering, guiding the model to focus on specific features, making it adaptable to various situations. However, since LLMs are not specifically trained for rumor detection, they may not be as effective at capturing certain subtle features as deep learning methods fine-tuned for this task. While RA and LeRuD excel at leveraging language understanding and providing insights through interpretability, they sometimes struggle with complex temporal and contextual information, especially when relying on predefined cues. In contrast, deep learning models, particularly those using graph neural networks, excel at extracting dynamic features in both time and space, benefiting from specialized training on structured data. This enables them to learn complex patterns within propagation networks.

[0149] This example achieves state-of-the-art results on both datasets for several reasons. First, it uses dynamic graphs to fully model the structural and temporal features of rumor propagation, which is crucial for distinguishing rumors. Second, it models the temporal propagation structure as a continuous-time dynamic graph, rather than a discrete-time one, allowing for a more granular representation of temporal propagation structures with specific temporal information. Finally, it models the dynamic contextual knowledge information of the source posts and comments, contributing to its state-of-the-art performance. Lastly, it uses a subgraph-level attention mechanism to dynamically fuse the representations of two subgraphs, rather than simply concatenating them. This allows the model to adjust the proportion of each type of information in the representation vector.

[0150] RQ2: What is the impact of each component in this example on the overall performance?

[0151] This example constructs ablation experiments on two datasets to comprehensively analyze the key components of the method used in this example. In particular, comparisons are constructed by removing certain components from the following variant of this example.

[0152] -source: Removes source post content enhancement functionality, i.e. only uses aggregated graphical representations of SS input to the classification module. -structure: Ignores the structural features of Gulu propagation, but retains content, time, and knowledge features.

[0153] -time_GAT: Use GAT to remove time information, i.e., the Koep rumor propagation structure and aggregated gate graph structure information. -time_GCN: Remove time information, i.e., retain the rumor propagation structure and usc vanills GCN aggregated graph structure information.

[0154] -CTDE: Remove the CTDE module and replace CTDE with TGAT

[44] .

[0155] - Knowledge: Remove dynamic knowledge structures, i.e., only retain the time propagation graph.

[0156] -sub_Atten: Removes subgraph-level Atten, which connects the time propagation graph with the time knowledge graph representation.

[0157] The performance of these variants is summarized in Figure 4 The following observations were made:

[0158] As Figure 4 The results show that this example outperforms all other variations, and removing these components degrades the performance of rumor detection. Among the various features, the content of the source post (-source) remains the most important signal for identifying rumors. However, relying solely on content-based features is insufficient for effective rumor identification.

[0159] The significant decrease in structure indicates that structured information is crucial for detecting rumors. Both -time GAT and -time GCN use static graph neural network methods, but -time GCN performs worse than -time GAT, reflecting the negative focus on reducing the weights of noisy nodes. Static graph methods are much more accurate than dynamic graph methods (-CTDE), with much lower accuracy on both datasets. This indicates that temporal information is an important feature for detecting rumors. This result also explains why the method proposed by Bian et al.

[45] is not as effective as this example. Since the former focuses on simulating temporal information, while the latter mainly explores the diffusion process, this example uses dynamic graph neural networks, which can make full use of temporal information and capture structured features.

[0160] This example outperforms -knowledge. This shows that the dynamic diffusion structure of knowledge can also be an important feature for identifying rumors. The lack of subgraph-level attention prevents this example from achieving peak performance, because subgraph-level attention can assign weights to subgraphs and adaptively adjust the weight sizes to obtain the optimal representation.

[0161] Figure 5 This example demonstrates a comparison of early detection performance between this case and the baseline.

[0162] RQ3: Is it possible to detect this in the early stages of rumor dissemination?

[0163] Early detection of misinformation is a key objective in rumor detection. This example now examines the performance of early rumor detection, comparing the detection accuracy of different modules at different time intervals or number of posts

[45] . The earlier a model detects a rumor, the less information it needs to process. This example evaluates its performance based on the accuracy of specific checkpoints.

[0164] First, the accuracy of all models improved as the deadline or number of posts increased, with this example showing the highest accuracy over other models for each deadline or number of posts.

[0165] Secondly, compared to deep learning-based methods, DTR and SVM-TK have poorer early detection performance because they only use hand-created features. In contrast, TD-RVNN and BU-RVNN, which only capture unidirectional propagation or diffusion structures, use a bidirectional GCN to capture bidirectional structures. Therefore, they are more advantageous for the early detection of rumors.

[0166] This example performs exceptionally well when only a limited sample of observations is available. It identifies misinformation in a very short time because it captures the dynamic structure of rumor propagation and contextual information. From Figure 5 The study also noted that the accuracy of this example fluctuated in the early stages on the Twitter15 and Twitter16 datasets. This may be due to noise in the early stages of propagation interfering with the detection process, while structural and temporal information becomes more valuable in later stages. This example significantly outperformed other models across all evaluation points. This further demonstrates the effectiveness of this example for early rumor detection.

[0167] in conclusion:

[0168] This method aims to capture spatial structure, continuous temporal information, external knowledge, and textual semantics within a unified framework. First, a propagation graph and a knowledge graph are constructed based on the dynamic evolution of post diffusion and background knowledge. Specifically, for temporal information, this example uses precise time, rather than simply treating the temporal graph as a discrete snapshot of an ordered graph. Then, a Temporal Graph Networks framework is used to capture the continuous temporal and spatial structure of propagation. Finally, a subgraph-level attention mechanism adaptively integrates the different information from the two subgraphs to create a representation for each post. Experimental results show that this method is highly effective in rumor detection on two real-world datasets, significantly outperforming other state-of-the-art models in early rumor detection-related tasks. References:

[0169] [1]Chen,

[0170] [2]Rahimi,M.,Roayaei,M.:A multi-view rumor detection framework using dynamic propaga-tion structure,interaction network,and content.lEEETransactions on Signal and Infor-mation Processing over Networks(2024).[3]Chen,X.,Zhou,F.,Trajcevski,G.,Bon-sangue,M:Multi-view learning with distinguishable feature fusion for rumor detec-tion.Knowledge-Based Systems240,108085(2022).

[0171] [4]Sun,M,Zhang,X,Ma,J,Xie,s.,Liu,Y,Philip,S.Y.:Inconsistent matters:Aknowledge-guided dual-consistency network for multi-modal rumordetection.EEETransactions on Knowledge and Data Engineering(2023).

[0172] [5]GAO,Y.-j.,LIANG,G.,JIANG,F.-t.,XUC.,YANG,J,CHEN,J-.,WANG,H.:Socialnetwork rumor detection:asurveyActa Electonica Sinica 48(7),1421(2020).

[0173] [6]Lin,H.,Yi,P.,Ma,1.,Jiang,H.,Luo,Z.,Shi.S.,iu,R.:Zero-shot rumordetection withpropagation structure via prompt learningIn:Proceedings of theAA AI Conference onArtificial Intelligence,vol.37,pp.5213-5221(2023).

[0174] [8]Zheng,P,Huang,Z.,Dou,Y..Yan,Y..Rumor detection on social mediathroughmining the social circles with high homogeneity.Information Sciences642,119083(2023).

[0175] [9]Huang,Q.,Zhou,C..Wu,J.Wang,M.Wang,B.:Deep structure learning forrumordetection on twitter.In:Proceedings of the2019 International JointConference on Neural Networks,pp.1-8.1EEE,Budapest,Hungary(2019).

[0176]

[10] Bian,T.,Xiao,X.,Xu,T.,Zhao,P.,Huang.W.,Rong,Y.,Huang,J.:Rumordetection on social media with bi-directional graph con-volutionalnetworks.In:Proceedings of theAAAl Conference on Artificial Intelligence,vol.34,Pp.549 556(2020).

[0177]

[11] Huang,Q..Zhou,C.,Wu,J.,Liu,..,Wang.B.:Deep spatial-temporalstructure learn-ing for rumor detection on twitter.NeuralComputing andApplications 35(18),1299513005(2023).

[0178]

[12] Xia,R,Xuan,K,Yu,J:A stateindependent and time-evolving networkforearly rumor detection in social media.In:Proceedings ofthe 2020ConferenceonEmpirical MMethods in Natural Language Pro-cessing,pp.90429051.Associationfor Com-putational Linguistics,Online(2020).

[0179]

[13] Ma,J.,Gao,W.,Mitra,P,Kwon,S.Jansen,B.1.,Wong,K.-F.,Cha,M.:Detect.ing rumors from microblogs with recurrentneural networks.In:Proceedings of the 25thInternational Joint Conference onArtificialIntelligence,pp.38183824.AAAI Press,New York,USA(2016).

[0180]

[14] Sun,M.,Zhang,X.,Zheng,J.,Ma,G.:DDGCN:dual dynamic graphconvolutiona!networks for rumor detection on social mediaIn:Proceedings ofthe AAAI Conference onArtificial Intelligence,vol.36,pp.4611-4619AAAI Press,Online(2022).

[0181] 15]Zhang,H.,Fang,Q.,Qian,$.,Xu,C.:Multi-modal knowledge-aware eventmemory net-work for social media rumor detection.In:Proceedings of the 27thAOM InternationalConference on Multimedia,pp.1942-1951(2019).

[0182]

[16] Cui,L.,Seo,H.,Tabar,M.,Ma,F.,Wang,s.Lee,D.:Deterrent:Knowledgeguided graphattention network for detecting healthcaremisinformation.In:Proceedings of the 26thACM SIGKDD Interational Conference onKnowledgeDiscovery&Data Mining,pp492-502(2020).

[0183]

[17] Wang,Y.,Qian,s.,Hu,J..Fang,Q..XuC.:Fake news detection viaknowledge-drivenmultimodal graph convolutional networksIn:Proceedings of the2020 InternationalConference on Multimedia Retrieval,pp.540-547(2020).

[0184]

[18] Dun,y..Tu,K..Chen,C.,Hou,C..Yuan,X.:Kan:Knowledge-aware attentionnetwork forfake news detection.In:Proceedings of theAA AI Conference onArtificial Intelligence.vol.35,pp.81-89(2021).

[0185]

[19] Rossi,E.,Chamberlain,B..Frasca,F..Eynard,D.,Monti,F..Bronstein,M.:Temporal graph networks for deep learn-ing on dynamic graphs.arXivpreprintarXiv:2006.10637(2020).

[0186]

[20] Xu,D.,Ruan,C.,Korpeoglu,E.,Kumar,S.,Achan,K.:Inductiverepresentation learning on temporal graphs.arXiv preprintarXiv:2002.07962(2020).

[0187]

[21] Song,C.,Shu,K.,Wu,B.:Temporally evolv.ing graph neural networkfor fake news detec,tion.Information Processing&Management58(6),102712(2021).

[0188]

[22] Ma,1.,Gao,W.,Wong,K.-F.:Detect rumorsin microblog posts usingpropagation structure via kernel learning.In:Proceedings ofthe 55th AnnualMeeting of the Association for Computational Linguistics(Volume1:LongPapers),pp.708-717.Association for Computational inguistics,VancouverCanada(2017).

[0189]

[23] Wei,.,Hu,D..Zhou,w..Yue,Z..HuS.:Towards propagation uncertainty:Edgeenhanced bayesian graph convolutional networks for rumor detection.arXivpreprintarXiv:2107.11934(2021).

[0190]

[24] Hisano,R.:Semi-supervised graph embedding approach to dynamiclink prediction.In:Proceedings of the Complex Networks lX,pp.109-121.Springer,Cham(2018).

[0191]

[25] Yu,w..Cheng,w..Aggarwal,C.C..chenH.,Wang,W:Link prediction withspatial and temporal consistency in dynamicnetworks.In:ICAl,pp.33433349(2017).

[0192]

[26] Ma,Y.,Tresp,V.,Daxberger,E.A.:Embedding models for episodicknowledge graphs.Journal of Web Semantics 59,100490(2019).

[0193]

[27] Sankar,A.,Wu,Y.,Gou,L,Zhang,W.Yang,H.:Dysat:Deep neuralrepresentationlearning on dynamic graphs via self-attentionnetworks.In:Proceedings of the 13th International Conference on Web Search andDataMining,pp.519-527(2020).

[0194]

[28] Pareja,A.,Domeniconi,G.,Chen,J..Ma.T.,Suzumura,T.,Kanezashi,H.,Kaler.T.,Schardl,T.,Leiserson,C.:Evolvegcn:Evolving graph convolutionalnetworks fordynamic graphs.In:Proceedings of the A A A1Conference onArtifcial Intelligence,vol.34.Pp.5363-5370(2020).

[0195]

[29] Nguyen,G.H.,Lee,J.B.,Rossi,R.A.,Ahmed,N.K..Koh,E.,Kim,S.:Dynamicnetworkembeddings:From random walks to temporal random walks.In:Proceedingsof the2018 1EEE International Conference on BigData,pp.1085-1092.EEE,Seattle,WA.USA(2018).

[0196]

[30] Bastas,N.,Semertzidis,T.,Axenopoulos,A.Daras,P.:evolve2vec:earning network representations using temporal unfolding.In:Proceedings ofthe InterationalConterenceon Multimedia Modeling,vol.11295,pp.447458.Springer,Thessaloniki,Greece(2019).

[0197]

[31] Kazemi,$.M.,Goel,R.,Jain,K.,Kobyzev,1.,Sethi,A.,Forsyth,P.,Poupart,P.:Representation learning for dynamic graphs:A sur-vey.Journal ofMachine Learning Research 21(70),1-73(2020).

[0198]

[32] Pellissier Tanon,T..Weikum,G.,SuchanekF.:Yago 4:A reason-ableknowledge base.In:Proceedings of the European Semantic Web Conference,vol.12123,pp.583-596.Springer,Heraklion,Crete,Greece(2020)

[0199]

[33] Ji,...Wang,Y.,Shi,B.,Zhang,D.,Wang,Z..Yan,1.:Microsoft conceptgraph:Miningsemantic concepts for short text understand.ing.Data Intelligence1(3),238-270(2019).

[0200]

[34] Kenton,J.D.M.-W.C.,Toutanova,L.K.Bert:Pre-training of deepbidirectional transformers for language understanding.In:Proceedings ofnaacL-HLT,vol.1,p.2(2019).

[0201]

[35] Ma,1.,Gao,W.,Wong,K.-F.:Detect rumorsin microblog posts usingpropagation structure via kernel learning.In:Proceedings othe 55th AnnualMeeting of the Association for Computational Linguistics(Volume 1:LongPapers),pp.708-717.Associationfor Computational inguistics,Vancouver,Canada(2017).

[0202]

[36] Zhao,Z.,Resnick,P.,Mei,Q.:Enquiringminds:Early detection ofrumors in socialmedia from enquiry posts.In:Proceedings ofthe 24thInternational Conference on WorldWide Web,pp.1395-1405(2015).

[0203]

[37] Castillo,C..Mendoza,M..Poblete,B.:Information credibility ontwitter.In:Proceed.ings of the 20th International Conference on World WideWeb,pp.675684(2011).

[0204]

[38] Kwon,S.,Cha,M:Jung:K.:Chen,w.38Wang,Y.:Prominent features ofrumor propagation in online social media.In:Proceed.ings of lEEE 13thInternational Conferenceon Data Mining,pp.1103-1108.1EEE,Dallas,TX,USA(2013).

[0205]

[39] Ma,J.,Gao,w.,Wong,K.-F.:Rumor detec-tion on Twitter with tree-structured recursive neural networks.In:Proceedings of the56th Annual Meetingof the Association forComputational Linguistics(Volume 1:LongPapers),pp.1980-1989.Association for Com.putational Linguistics,Melbourne,Australia(2018).

[0206]

[40] Khoo,L.M.S.,Chieu,H...,Qian,Z..Jiang,1:Interpretable rumordetection inmicroblogs by attending to user interactions.In:Proceedings ofthe AA Al Conference onArtificial Intelligence,vol.34,pp.8783 8790(2020).

[0207]

[41] Ye,N.,Yu,D.,Zhou,Y.,shang,K.-k..Zhang,S.:Graph convolutional-based deepresidual modeling for rumor detection onsocial media.Mathematics 11(15),3393(2023).

[0208]

[42] Yue Huang,LS:Fake detection:Fake news gen-eration,explanation anddetection of largelanguage models.In:Proceedings of the ACMWeb Conference2024,pp.1-1(2024).

[0209]

[43] Liu,Q.,Tao,X.,Wu,J.,Wu,S.,Wang,L.:Can largelanguage models detect rumors on social media?arXiv preprintarXiv:2402.03916(2024).

[0210]

[44] Huang,$.,Poursafaei,F.,Danovitch,J.,Fey,M.,Hu,W.,Rossi,E.,Leskovee,J.,Bron-stein,M.,Rabusseau,G.,Rabbany,R.:Tem-poral graph benchmarkfor machine learning on temporal graphs.Advances in NeuralInformationProcessing Systems 36(2024).

[0211]

[45] Bian,T.,Xiao,X.,Xu,T.,Zhao,P.,HuangW.,Rong,Y.,Huang,J.:Rumor detection onsocial media with bi-directional graph con-volutional networks.

Claims

1. A continuous-time dynamic graph processing method for rumor identification, characterized in that, Includes the following steps: Step 1: Using the source post and the text of the reply posts of an event as input, construct a time-series propagation graph; Step 2: Using the text of the source post and reply posts of an event as input, use TagMe to link entities to the input, linking the entities mentioned in the source post and reply posts to the relevant entities in YAGO, obtaining an entity set. For each identified entity in the entity set, obtain the corresponding concept of the identified entity in the existing knowledge graph through conceptualization, obtaining a concept set. Based on the entity set and concept set obtained above, construct a temporal knowledge graph. Step 3: Using the temporal propagation graph and temporal knowledge graph constructed in Step 1 and Step 2 as input, the temporal-spatial feature fusion module is used to fuse the input temporal propagation graph and temporal knowledge graph respectively to generate two node embedding sets; The time-space feature fusion module includes the following core components: Memory component: The memory component records the historical information of nodes in time. Vector of each node This constitutes a memory, and the memory of a node is updated after an event occurs; in this way, the time-space feature fusion module can remember the long-term dependencies of each node in the temporal propagation graph and the temporal knowledge graph. If a new node appears, the memory of that new node is initialized. A vector whose memory is updated when an event involving the new node occurs; Post function: Each containing node and The interactive event will correspond to two posts: ； ； Includes only nodes Each event corresponds to a post: ； in Represents a node In time Previous memories, , and Each layer consists of three fully connected layers; Post aggregator: Uses aggregation functions to aggregate posts. Moment : ； And select average pooling as function; Step 4: Using the two node embedding sets generated in Step 3 as input, use the subgraph fusion module to generate a node embedding set that fuses the structural information of the two subgraphs. Step 5: Using the node embedding set generated in Step 4 as input, the classification module is used to classify the input to obtain the final prediction result.

2. The continuous-time dynamic graph processing method for rumor identification according to claim 1, characterized in that: The specific process of step 1 is as follows: for an event Its time-series propagation diagram is represented as follows , where the node set Includes source post and reply post, edge set In the set of edges, the edges represent interactions between the source post and the reply post, or interactions between reply posts. For simplicity, the direction of the edges in the edge set is not considered. Represented as an undirected graph, and initialized with embedding vectors. Each node in the representation.

3. The continuous-time dynamic graph processing method for rumor identification according to claim 2, characterized in that: The specific process of step 2 is as follows: for an event Its time-series knowledge graph representation is as follows , where the node set It is a time-series propagation graph node set Entity set and concept set The union of the sets, initialized with embedding vectors. Each node in the representation follows the following rules for edge construction: (1) Post-entity edge: if One of the posts contains links to (2) Entity-entity edge, entity-concept edge, concept-concept edge: Use a fixed-size sliding window to count the relevance of nodes in the global corpus, then calculate the PMI score between node pairs, retain edges with positive PMI values, and remove edges with non-positive PMI values.

4. The continuous-time dynamic graph processing method for rumor identification according to claim 2, characterized in that: In step 3, when using the time-space feature fusion module, for events... ,based on The constructed temporal propagation graph and temporal knowledge graph will be processed through two time-space feature fusion modules to obtain the final time. Two sets of node embeddings: and .

5. The continuous-time dynamic graph processing method for rumor identification according to claim 1, characterized in that: In step 3, the time-space feature fusion module, for the process of generating node embeddings, also includes the following core components: Post updater: The post updater updates itself based on the node's memory and event messages. ； When node and nodes When an interaction event occurs between the nodes, the memories of both nodes will be updated after the event occurs. For a single node event, only the memory of the involved node is updated, and the memory of other nodes is not updated. Embedded Components: Nodes generated by embedded components In time Temporal embedding, the embedded component contains a series of graph attention layers through nodes of Nodes are calculated by aggregating information in the time-series neighborhood. The embedding, the first The input to the graph attention layer includes nodes. In the upper layer representation Current time step ,node The neighbors said timestamp and nodes Features of the edges at each time step : ； in, This indicates a splicing operation. This represents a time encoder; multi-head attention is performed at each layer, where queries... The reference node is represented, and the key is... Sum Indicates query The neighbor, the last For fusion nodes Features at the upper layer and the nodes after aggregating neighbor nodes Features The initial input for each node is represented by the current memory. and time node features : ； Original Post Storage: The post embedding for each batch updates the information in memory using posts from previous batches stored in the original post storage, and generates node embeddings.

6. The continuous-time dynamic graph processing method for rumor identification according to claim 1, characterized in that: In step 4, the subgraph fusion module generates a set of node embeddings that fuse the structural information of two subgraphs using a subgraph-level attention mechanism. To learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert it into a low-dimensional vector. Then, the attention vector of each node and the subgraph are calculated. The similarity between nodes is used as the importance score of the nodes in the current subgraph. The average importance score of all nodes in the current subgraph is used as the importance score of each subgraph. The average importance score of all nodes in the same subgraph is used as the importance score of the subgraph. The importance of the time-series propagation graph and the time-series knowledge graph are calculated. The formula is as follows: ； in It is a weight matrix. It's a bias, a subgraph-level attention vector. Weight and the bias of the linear layer Shared between the two subgraphs, a subgraph importance score is obtained. Then, these scores are normalized into weighting coefficients: ； final Using the learned subgraph weight coefficients, a set of node embeddings that integrates information from both subgraphs is obtained. : 。 7. The continuous-time dynamic graph processing method for rumor identification according to claim 1, characterized in that: In step 5, the classification module first applies mean pooling to fuse the data. Structural information representation of nodes: ； Then, apply Encode the source post information to obtain , and and Then, stitch the images together and apply a linear layer for fusion. ； Finally, The input is fed into the classifier to obtain the prediction result: ； Training is guided by cross-entropy loss and regularization, as shown in the following formula: ； in It is the first The baseline truth label for each event Represents the trade-off coefficient. Represents the set of all parameters. This represents the L2 regularization term, used to prevent overfitting.

Images