Continuous time dynamic graph processing method for rumor recognition

Abstract

The invention discloses a continuous time dynamic graph processing method for rumor recognition. The continuous time dynamic graph processing method comprises the following steps: step 1, constructing a time sequence propagation graph; 2, constructing a time sequence knowledge graph; 3, taking the constructed time sequence propagation graph and time sequence knowledge graph as input, and fusing the input time sequence propagation graph and time sequence knowledge graph by using a time-space feature fusion module to generate two node embedding sets; 4, taking the two node embedding sets generated in the step 3 as input, and generating a node embedding set fused with two sub-graph structure information by using a sub-graph fusion module; and step 5, taking the node embedding set generated in the step 4 as input, and classifying the input by using a classification module to obtain a final prediction result. According to the technology, better node embedding can be obtained, and the rumor detection efficiency is high.

Description

Technical Field

[0001] The invention relates to rumor identification technology, in particular to a continuous-time dynamic graph processing method for rumor identification. Background Art

[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, Instagram, etc. 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 create social connections during social distancing and promote the dissemination of knowledge in various contexts. On the other hand, they may lead people to share quick and superficial ideas and spread rapidly, such as rumors. Rumors are defined as unconfirmed or proven false news [5,6]. The explosive spread of rumors threatens the credibility of the Internet and has serious adverse effects on individuals and society. The seriousness of rumor generation has stimulated the need for effective detection of false information on online social media tools [8]. However, the short and fast-spreading nature of rumors makes them difficult to detect automatically. Therefore, automatic rumor detection has attracted more and more attention from researchers.

[0003] Early rumor detection methods mainly relied on manual feature selection. The selected features focused on extracting a variety of lexical and semantic features from 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 introduced deep learning methods to the problem of social media rumor detection. Deep learning methods have powerful feature learning capabilities. They are able to learn more accurate and representative raw data features than those obtained by traditional machine learning algorithms through feature engineering, thereby achieving better classification results.

[0004] Recent studies have shown that modeling diffusion 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 the 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 sequence of comments or reposts, which can further distinguish the propagation forms [11,12]. Ma et al.

[13] input the source posts and reposts into RNN in time order to capture the contextual information of related posts. Sun et al.

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

[0006] At the same time, another parallel research line introduced external knowledge related to rumor identification [15–18]. These studies usually use knowledge related to the content of the source post to extend the semantics and thus improve the quality of the post representation. The basic logic is that posts are usually short and contain many abbreviations, proper nouns, nicknames, etc., and traditional classification models often have difficulty understanding their meaning. The knowledge base includes background knowledge corresponding to these unique words. Adding this knowledge to the text representation can greatly help the model determine the credibility of the post.

[0007] Although existing deep learning-based rumor detection methods have made significant progress, they still face several limitations:

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

[0009] Coarse temporal granularity: Existing time-based rumor detection methods either input posts into 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 specific or detailed manner. Since the propagation of posts has a definite time and is an ever-changing process, the propagation process of posts can be modeled as a continuous time dynamic graph (CTDG). Figure 3 As shown in Figure 2, CTDG can represent the propagation process of posts with exact time, so it can accurately represent temporal information. In contrast, DTDG can only extract the static propagation structure of interval time, so the granularity of the time series is relatively coarse.

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

[0011] The purpose of the present invention is to address the deficiencies in the prior art and to propose a continuous-time dynamic graph processing method for rumor identification. This technology can obtain better node embedding and has high efficiency in rumor detection.

[0012] The technical solution for achieving the technical purpose of the present invention is:

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

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

[0015] Step 2: Take the text of the source post and reply post of an event as input, use TagMe to perform entity linking on the input, link the entities mentioned in the source post and reply post to the relevant entities in YAGO, and obtain 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 to obtain a concept set. Based on the entity set and concept set obtained above, construct a time series knowledge graph;

[0016] Step 3: Taking the temporal propagation graph and temporal knowledge graph constructed in steps 1 and 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, a subgraph fusion module is used to generate a node embedding set that fuses the structural information of the two subgraphs.

[0018] Step 5: Take the node embedding set generated in step 4 as input, use the classification module to classify the input, and get the final prediction result.

[0019] The specific process of step 1 is as follows: for an event ε i , and its timing propagation diagram is expressed as Among them, the node set Contains source posts and reply posts, side collection The edges in represent the interaction between the source post and the reply post or the interaction between the reply posts. For simplicity, the direction of the edges in the edge set is not considered. Represented as an undirected graph and initialized with an embedding vector Each node in represents .

[0020] The specific process of step 2 is as follows: for an event ε i , whose temporal knowledge graph is represented as Among them, the node set is the set of nodes in the time series propagation graph Entity Set E i and concept set C i The union of , using the embedding vector initialization Each node in represents that the construction of the edge follows the following rules: (1) Post-entity edge: If A post in contains a link to E i For words containing entities in the post, an edge will be added between the post node and the entity node; (2) Entity-entity edge, entity-concept edge, concept-concept edge: a fixed-size sliding window is used to count the relevance of nodes in the global corpus, and then the PMI score between node pairs is calculated, edges with positive PMI values ​​are retained, and edges with non-positive PMI values ​​are removed.

[0021] In step 3, when the time-space feature fusion module is used, for event ε i , based on ε i The constructed temporal propagation graph and temporal knowledge graph are processed by two time-space feature fusion modules to obtain the final time t m The two nodes of the embedding set are: and

[0022] In step 3, for the process of generating node embedding 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 node. The vector s of each node at time t 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 dependency relationship 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 involving nodes i and j will correspond 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] Events that only contain node i will correspond to a 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 They are composed of three fully connected layers respectively;

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

[0031]

[0032] And choose average pooling 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 are updated after the event. For a single node event, only the memory of the involved nodes is updated, while 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 of the l-th graph attention layer includes the representation of node i in the upper layer. Neighbor representation of node i at current time step t Timestamp t1,…,t N , and the edge feature e of node i at each time step i1 (t1),…,e iN (t N ):

[0037]

[0038] where ‖ represents the concatenation operation and φ(·) represents the temporal 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) (t) neighbors, the final MLP is used to fuse the features of node i in the previous layer And the characteristics of node i after aggregating neighbor nodes The initial input of each node is represented as the current memory s j (t) and time node feature v j (t):

[0039]

[0040] Raw Post Store: The post embeddings of each batch update the information in the memory using the posts from the previous batch, which are stored in the raw post store, and generate node embeddings.

[0041] In step 4, the subgraph fusion module generates a set of node embeddings that fuse the structural information of the two subgraphs by using the subgraph-level attention mechanism. In order to learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert the subgraph representation 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 importance scores of all nodes in the current subgraph are averaged 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, and the importance w of the temporal propagation graph and the temporal knowledge graph is calculated. TP(TK) The formula is as follows:

[0042]

[0043] Where W sub is the weight matrix, b is the bias, sub-graph level attention vector q, weight W sub The bias b of the linear layer is shared between the two subgraphs to obtain the subgraph importance score w TP(TK) Finally, these scores are normalized into weight coefficients:

[0044]

[0045] Finally, using the learned subgraph weight coefficients, we obtain a node embedding set that combines the information of the two subgraphs.

[0046]

[0047] In step 5, the classification module first applies a mean pooling operation to fuse The structural information of the midpoint is represented as follows:

[0048]

[0049] Then, BERT is applied to encode the source post information and obtain p i0 , and p i0 With X T Concatenate and then apply a linear layer blend:

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

[0051] Finally, S i Input into the classifier and get the prediction result:

[0052]

[0053] The cross entropy loss and regularization term are used to guide the training, and the formula is as follows:

[0054]

[0055] where y i is the ground truth label of the ith event, λ represents the trade-off coefficient, θ represents the set of all parameters, Represents the L2 regularization term, which is used to prevent overfitting.

[0056] This technical solution uses a time-space feature fusion module to capture the content semantics, spatial structure, and temporal information in the post propagation process. This technical solution uses a continuous-time dynamic graph (precise time information) to model post propagation instead of using a discrete-time dynamic graph (time graph snapshot), because the propagation of posts in social networks is essentially continuous and evolves over time. On the other hand, this technical solution extracts conceptual knowledge related to source posts and individual comments from an external knowledge base and constructs a temporal knowledge graph that evolves over time. Similarly, this technical solution uses a time-space feature fusion module to encode the dynamic knowledge propagation process. Finally, this technical solution uses a subgraph-level attention mechanism to adaptively fuse dynamic post information and dynamic knowledge information to achieve better rumor detection.

[0057] The main contributions of this technical solution to the above 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 as well as the spatial structure of propagation. In addition, temporal information is introduced as a feature during the node update process to obtain a node representation that integrates content semantics, spatial, and temporal information.

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

[0060] Dynamic knowledge information: This technical solution focuses not only on the dissemination process of posts, but also on the dissemination process of knowledge. It uses an external knowledge base to build a temporal 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 text semantics through a unified framework. First, the propagation graph and knowledge graph are constructed according to the dynamic evolution of post diffusion and background knowledge. Then, the temporal graph network framework is adopted to capture the continuous temporal and spatial structure in the propagation process. Finally, the subgraph-level attention mechanism is used to adaptively integrate the different information in the two subgraphs, create representations for each post, obtain better node embedding, and achieve high efficiency in rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0063] Figure 2 This is an operation flow chart of the time-space feature fusion module in the embodiment;

[0064] Figure 3 A schematic diagram of modeling the post propagation process in the prior art as a continuous time dynamic graph and a discrete time dynamic graph, the left side is modeling the post propagation as a continuous time dynamic graph, the right side is modeling the post propagation as a discrete time dynamic graph, and each Twitter logo represents a post;

[0065] Figure 4 Schematic diagram of ablation experiment of the embodiment, FIG. a is an ablation experiment of Twitter 15, and FIG. b is an ablation experiment of Twitter 16;

[0066] Figure 5Schematic diagrams of rumor detection on Twitter 15 and Twitter 16 as the time spent or the number of posts increases in the early stage of the embodiment, FIG. a is the rumor detection based on the time spent on Twitter 15, FIG. b is the rumor detection based on the number of posts on Twitter 15, FIG. c is the rumor detection based on the time spent on Twitter 16, and FIG. d is the rumor detection based on the number of posts on Twitter 16. DETAILED DESCRIPTION

[0067] The content of the present invention is further described below in conjunction with the drawings and embodiments, but the present invention is not limited thereto.

[0068] Example:

[0069] Problem definition:

[0070] The rumor detection dataset gives the event set ε = {ε1,ε2,...,ε n}, where ε i Represents 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 reply to the post p i The collection is Where m represents ε i The number of replies in s i Can be used as p i0 You can also get the event ε using i Related release time series where t i0 =0. and are combined to get ε i ={(p i0 , t i0 ), (p i1 , t i1 ),…,(p im , t im )}. We need to learn a model f:ε→y to transform the event ε i Classify into predefined categories y = {0, 1}, where the event label is 0 for non-rumor and 1 for rumor.

[0071] Reference Figure 1 , a continuous time dynamic graph processing method for rumor identification, comprising the following steps:

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

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

[0074] Step 3: Taking the temporal propagation graph and temporal knowledge graph constructed in steps 1 and 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, a subgraph fusion module is used to generate a node embedding set that fuses the structural information of the two subgraphs.

[0076] Step 5: Take the node embedding set generated in step 4 as input, use the classification module to classify the input, and get the final prediction result.

[0077] The specific process of step 1 is as follows: for an event ε i , and its timing propagation diagram is expressed as Among them, the node set Represents the source post embedding and reply post embedding set, edge set The edges included in represent the interaction between the source post and the reply post, as well as the interaction between the reply posts. For simplicity, the edge set is not considered. The direction of the edge in the set will Represented as an undirected graph and initialized with an embedding vector Each node in represents .

[0078] The specific process of step 2 is as follows: for an event ε i , whose temporal knowledge graph is represented as Among them, the node set is the set of nodes in the time series propagation graph Entity Set E i and concept set C i The union of , using the embedding vector initialization Each node in represents that the construction of the edge follows the following rules: (1) Post-entity edge: If A post in contains a link to E iFor words containing entities in the post, an edge will be added between the post node and the entity node; (2) Entity-entity edge, entity-concept edge, concept-concept edge: a fixed-size sliding window is used to count the relevance of nodes in the global corpus, and then the PMI (Point Mutual Information) score between node pairs is calculated, edges with positive PMI values ​​are retained, and edges with non-positive PMI values ​​are removed.

[0079] In step 3, when the time-space feature fusion module is used, for event ε i , based on ε i The constructed temporal propagation graph and temporal knowledge graph are processed by two time-space feature fusion modules to obtain the final time t m The two nodes of the embedding set are: and

[0081] like Figure 2 As shown, in step 3, for the process of generating node embedding 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 the node. The vector s of each node at time t constitutes the memory. The memory of the node will be updated after an event occurs (for example, interaction with other nodes or changes in 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 involving nodes i and j will correspond 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] Events that only contain node i will correspond to a 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 They are composed of three fully connected layers respectively;

[0089] Post Aggregator: To improve efficiency, batch processing is used to reduce computation time. However, this results in multiple events involving node i in the same batch. To handle this problem, an aggregation function is used to aggregate t1,…,t b ≤ p at time t i (t1),…,p i (t b ):

[0090]

[0091] And choose average pooling 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 are updated after the event. For a single node event, only the memory of the involved nodes is updated, while 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 of the l-th graph attention layer includes the representation of node i in the upper layer. Neighbor representation of node i at current time step t Timestamp t1,…,t N , and the edge feature e of node i at each time step i1 (t1),…,e iN (t N ):

[0096]

[0097] where ‖ represents the concatenation operation and φ(·) represents the temporal 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), and the key K (l) (t) and value V (l) (t) represents the query q (l) (t) neighbors, the final MLP is used to fuse the features of node i in the previous layer And the characteristics of node i after aggregating neighbor nodes The initial input of each node is represented as the current memory s j (t) and time node feature v j (t):

[0098]

[0099] Raw Post Storage: Since the modules associated with memory do not directly affect the loss and do not receive gradients, in order to solve the above problem, the information in the memory is updated with posts from the previous batch, which are stored in the raw post storage and generate node embeddings; the updated memory is then used to calculate the embedding of the node, and the new interactions generated after this interaction will be stored in the raw post storage.

[0100] In step 4, the subgraph fusion module generates a set of node embeddings that fuse the structural information of the two subgraphs by using a subgraph-level attention mechanism. In order to learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert the subgraph representation 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 importance scores of all nodes in the current subgraph are averaged 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, and the importance w of the temporal propagation graph and the temporal knowledge graph are calculated. TP(TK) The formula is as follows:

[0101]

[0102] Where W sub is the weight matrix, b is the bias, sub-graph level attention vector q, weight W sub The bias b of the linear layer is shared between the two subgraphs to obtain the subgraph importance score w TP(TK) Finally, these scores are normalized into weight coefficients:

[0103]

[0104] Finally, using the learned subgraph weight coefficients, we obtain a node embedding set that combines the information of the two subgraphs.

[0105]

[0106] In step 5, the classification module first applies a mean pooling operation to fuse The structural information of the midpoint is represented as follows:

[0107]

[0108] Then, BERT is applied to encode the source post information and combine it with X T To perform splicing and fusion:

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

[0110] Finally, S i Input into the classifier and get the prediction result:

[0111]

[0112] The cross entropy loss and regularization term are used to guide the training, and the formula is as follows:

[0113]

[0114] where y i is the ground truth label of the ith event, λ represents the trade-off coefficient, θ represents the set of all parameters, Represents the L2 regularization term, which is used to prevent overfitting.

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

[0116]

[0117]

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

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

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

[0121] RQ3: Can this case detect rumors at an early stage of spreading?

[0122] Experimental setup:

[0123] Dataset:

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

[35] . The labels of each event on Twitter15 and Twitter16 are labeled according to 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). This example divides the two datasets into training set, validation set, and test set in a constant ratio of 7:1:2. Table 1 shows the statistics of these datasets.

[0125] Table 1 Statistics 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 decision tree-based ranking model that clusters posts and selects the top k clustering 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 properties, and propagating user attributes for detection.

[0131] SVM-TK

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

[0132] GRU-RNN: An RNN with GRU units to model the response time characteristics of reviews.

[0133] TD-RvNN and BU-RvNN

[39] : A tree-structured recurrent neural network with GRU units, where 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 pair interactions.

[0135] Bi-GCN

[10] : A GCN-based rumor detection method that uses a bidirectional propagation structure (propagation and diffusion structure) as well as text content (source posts and comments).

[0136] STS-NN

[11] : A RNN-based model that modifies RNN to model spatial and temporal structures.

[0137] GCRES

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

[0138] RA

[42] : A misinformation detection method based on large language models (LLMs), which identifies whether a post meets any of nine predefined misinformation features.

[0139] LeRuD

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

[0140] Implementation settings:

[0141] This example is implemented using the Pytorch framework, and the parameters are optimized using the Adam algorithm. BERT-base is used as the encoder of the source code and pre-trained on the dataset. The learning rate of BERT is set to 5e-5, the learning rate of other parameters is set to 1e-3, the batch size is set to 16, and the training epoch is set to 5. This example selects the best parameter settings on the validation set and evaluates the performance of this example on the test set. This example uses Accuracy (Acc) in all categories and the F1 score for each category to evaluate the performance of the model.

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

[0143] Table 2 lists the performance comparison between this example and other baselines and makes the following observations:

[0144] Table 2 shows the performance comparison on Twitter 15 and Twitter 16 datasets. The best results are shown in bold and the second best results are underlined.

[0145]

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

[0147] Due to the use of neural networks, deep learning models can automatically extract effective features, and the performance is significantly improved compared to feature-based methods. The performance of GRU-RNN is relatively poor because it only utilizes the content of the post and ignores other useful information such as the propagation structure and temporal information of the rumor. BU-RVNN and TD-RVNN perform better than GRU-RNN, which shows that modeling spatiotemporal structure is effective for rumor detection. PLAN performs well on both datasets, indicating that capturing the interaction between posts is effective. STS-NN and Bi-GCN model the propagation structure and temporal information between source posts and comments. Bi-GCN is better than STS-NN because it can distinguish spatiotemporal structural patterns. In addition, Bi-GCN does not include language features, but focuses on feature mining for rumor propagation and diffusion. As the number of nodes in the propagation process decreases with the tree, the information that can be provided decreases accordingly, resulting in a decrease in model performance. The strong performance of GCRES over the baseline model demonstrates the effect of using deep residual graph networks to learn content and structural interaction information in network propagation models.

[0148] The RA and LeRuD methods based on LLMs provide a flexible and interpretable approach to rumor detection. They use cue engineering, which can guide the model to focus on specific features, making it adaptable in a variety of situations. However, since LLMs are not specifically trained for rumor detection, they may not be as effective in capturing certain subtle features as deep learning methods fine-tuned for this task. While RA and LeRuD excel in leveraging language understanding and providing insights through interpretability, they sometimes face challenges when dealing with complex temporal and contextual information, especially when relying on predefined cues. In contrast, deep learning models, especially those using Graph Neural Networks, excel in extracting dynamic features in time and space, benefiting from specialized training on structured data. This enables them to learn complex patterns in propagation networks.

[0149] There are several reasons why this example achieved the best results on both datasets. First, this example uses dynamic graphs to fully model the structural and temporal features of rumor propagation, which is crucial for distinguishing rumors. Second, this example models the temporal propagation structure as a continuous-time dynamic graph instead of a discrete-time dynamic graph, which can represent the temporal propagation structure with specific temporal information in a more fine-grained manner. Finally, this example models the dynamic background knowledge information of the source posts and comments, which helps this example achieve the best results. Finally, this example uses a subgraph-level attention mechanism to dynamically fuse the representations of the two subgraphs instead of simply concatenating them. This enables the model to adjust the proportion of the representation vector occupied by the two types of information.

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

[0151] We construct ablation experiments on two datasets to fully analyze the key components of our method. In particular, we construct comparisons by removing certain components of the following variants of our method.

[0152] -source: remove the source post content enhancement, i.e. only use the pooled graph representation to input into the classification module -structure: ignore the structural features of Guruk-propagation but keep the content, time and knowledge features.

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

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

[44] .

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

[0156] -sub_Atten: Remove the sub-graph level Atten, i.e. connect the temporal propagation graph with the temporal 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 variants, and removing these components will reduce the performance of rumor detection. Among various features, the content of the source post (-source) is still the most important signal for identifying rumors. However, relying solely on content-based features is not enough to effectively identify rumors.

[0159] The structure is significantly reduced, indicating that structured information is crucial for detecting rumors. Both -Temporal GAT and -Temporal GCN use static graph neural network methods, while -Temporal GCN performs worse than -Temporal GAT, reflecting the negative focus on reducing the weights of noisy nodes. The static graph method is much more accurate than the dynamic graph method (-CTDE), which has much lower accuracy on both datasets. This shows 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 modeling temporal information and the latter mainly explores the diffusion process, this example uses a dynamic graph neural network, which can make full use of temporal information and capture structured features.

[0160] The performance of this example is better than -knowledge. This shows that the dynamic diffusion structure of knowledge can also be an important feature for identifying rumors. The lack of sub-graph level attention (-sub Atten) prevents this example from achieving peak performance, because sub-graph level attention can assign weights to sub-graphs and adaptively adjust the weight size to obtain the best representation.

[0161] Figure 5 The comparison of this example with the baseline in terms of early detection performance is shown.

[0162] RQ3: Can this case detect rumors at an early stage of spreading?

[0163] Early detection of misleading information is a key goal in rumor detection. This example now examines the performance of early rumor detection, that is, this example compares the detection accuracy of different modules at different time or number of posts

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

[0164] First, as the deadline or number of posts increases, the accuracy of all models improves, and in particular, this example has the highest accuracy compared to other models at each deadline or number of posts.

[0165] Second, compared with deep learning-based methods, DTR and SVM-TK have poor early detection performance because they only use hand-crafted features. Compared with TD-RVNN and BU-RVNN that only capture unidirectional propagation or diffusion structures, Bi-GCN uses bidirectional GCN to capture bidirectional structures. Therefore, it is more conducive to the early detection of rumors.

[0166] This example performs particularly well when there are only a limited number of observations. It takes only a short time to identify false information because it captures the dynamic rumor propagation structure and background knowledge. Figure 5 In the , we also noticed that the accuracy of this example on Twitter15 and Twitter16 datasets fluctuates in the early stages. This may be due to the fact that the noise in the early stages of propagation interferes with the detection process, while the structural and temporal information becomes more valuable in the later stages. This example significantly outperforms other models at each evaluation point. This further proves the effectiveness of this example for early rumor detection.

[0167] in conclusion:

[0168] The method in this example aims to capture spatial structure, continuous temporal information, external knowledge, and text semantics in a unified framework. First, based on the diffusion of posts and the dynamic evolution of background knowledge, a propagation graph and a knowledge graph are constructed. In particular, for temporal information, this example uses precise time rather than just treating the temporal graph as a discrete snapshot of an ordered graph. Then, the Temporal Graph Networks framework is used to capture the continuous temporal and spatial structure of propagation. Finally, a subgraph-level attention mechanism is used to adaptively integrate different information in two subgraphs to create a representation for each post. Experimental results show that this example is very effective in rumor detection on two real-world datasets, and significantly outperforms 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: The steps include: Step 1: Take the source post and reply post text of an event as input and construct a time series propagation graph; Step 2: Take the text of the source post and reply post of an event as input, use TagMe to perform entity linking on the input, link the entities mentioned in the source post and reply post to the relevant entities in YAGO, and obtain 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 to obtain a concept set. Based on the entity set and concept set obtained above, construct a time series knowledge graph; Step 3: Taking the temporal propagation graph and temporal knowledge graph constructed in steps 1 and 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; Step 4: Using the two node embedding sets generated in step 3 as input, a subgraph fusion module is used to generate a node embedding set that fuses the structural information of the two subgraphs. Step 5: Take the node embedding set generated in step 4 as input, use the classification module to classify the input, and get 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 ε i , and its timing propagation diagram is expressed as Among them, the node set Contains source posts and reply posts, side collection The edges in represent the interaction between the source post and the reply post or the interaction between the reply posts. For simplicity, the direction of the edges in the edge set is not considered. Represented as an undirected graph and initialized with an embedding vector Each node in represents .

3. The continuous-time dynamic graph processing method for rumor identification according to claim 2 is characterized in that: The specific process of step 2 is as follows: for an event ε i , whose temporal knowledge graph is represented as Among them, the node set is the set of nodes in the time series propagation graph Entity Set E i and concept set C i The union of , using the embedding vector initialization Each node in represents that the construction of the edge follows the following rules: (1) Post-entity edge: If A post in contains a link to E i For words containing entities in the post, an edge will be added between the post node and the entity node; (2) Entity-entity edge, entity-concept edge, concept-concept edge: a fixed-size sliding window is used to count the relevance of nodes in the global corpus, and then the PMI score between node pairs is calculated, edges with positive PMI values ​​are retained, and edges with non-positive PMI values ​​are removed.

4. The continuous-time dynamic graph processing method for rumor identification according to claim 2 is characterized in that: In step 3, when the time-space feature fusion module is used, for event ε i , based on ε i The constructed temporal propagation graph and temporal knowledge graph are processed by two time-space feature fusion modules to obtain the final time t m The two nodes of the embedding set are: 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 generates a process of node embedding, and the time-space feature fusion module has the following core components: Memory component: The memory component records the historical information of the node. The vector s of each node at time t 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 dependency relationship 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. Post function: Each interaction event involving nodes i and j will correspond to two posts: p i (t)=psg s (s i (t - ),s j (t - ),Δt,e ij (t))(1) p j (t)=psg d (s j (t - ),s i (t - ),Δt,e ij (t))(2) Events that only contain node i will correspond to a post: p i (t)=psg n (s i (t - ),t,v i (t))(3) where s i (t - ) represents the memory of node i before time t, psg s ,psg d and psg n They are composed of three fully connected layers respectively; Post Aggregator: Uses an aggregation function to aggregate t1,…,t b ≤ p at time t i (t1),…,p i (t b ): And choose average pooling as the agg(·) function; Post Updater: The post updater updates itself based on the node's memory and event messages: When an interaction event occurs between node i and node j, the memories of both nodes are updated after the event. For a single node event, only the memory of the involved nodes is updated, while the memory of other nodes is not updated. 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 of the l-th graph attention layer includes the representation of node i in the upper layer. Neighbor representation of node i at current time step t Timestamp t1,…,t N , and the edge feature e of node i at each time step i1 (t1),…,e iN (t N ): K (l) (t)=V (l) (t)=C (l) (t)(9) where ‖ represents the concatenation operation and φ(·) represents the temporal 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) (t) neighbors, the final MLP is used to fuse the features of node i in the previous layer And the characteristics of node i after aggregating neighbor nodes The initial input of each node is represented as the current memory s j (t) and time node feature v j (t): Raw Post Store: The post embeddings of each batch update the information in the memory using the posts from the previous batch, which are stored in the raw post store, and generate 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 the two subgraphs by using a subgraph-level attention mechanism. In order to learn the weights of the two subgraphs, a linear layer is first applied to the subgraph representation to convert the subgraph representation 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 importance scores of all nodes are averaged in the current subgraph 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 to calculate the importance of the temporal propagation graph and the temporal knowledge graph. w TP(TK) The formula is as follows: Where W sub is the weight matrix, b is the bias, sub-graph level attention vector q, weight W sub The bias b of the linear layer is shared between the two subgraphs to obtain the subgraph importance score w TP(TK) Finally, these scores are normalized into weight coefficients: Finally, using the learned subgraph weight coefficients, we obtain a node embedding set that combines the information of the two subgraphs.

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 a mean pooling operation to fuse The structural information of the midpoint is represented as follows: Then, BERT is applied to encode the source post information and obtain p i0 , and p i0 With X T Concatenate and then apply a linear layer blend: S i =linear(BERT(p i0 ‖X T ))(16) Finally, S i Input into the classifier and get the prediction result: The cross entropy loss and regularization term are used to guide the training, and the formula is as follows: where y i is the ground truth label of the ith event, λ represents the trade-off coefficient, θ represents the set of all parameters, Represents the L2 regularization term, which is used to prevent overfitting.

Images