Social network rumor early detection system based on level-aware attention gru network

Abstract

The application discloses a social network rumor early detection system based on a level-aware attention GRU network, comprising a news event acquisition module, a forwarding level measurement module, a basic feature extraction module, a tweet feature representation module and a rumor detection module; the application proposes a new system based on an attention mechanism, the system pays attention to more important tweet content according to the forwarding level of a node, and learns the overall features of the tweets from a time sequence through a classifier based on a GRU neural network to predict a label (rumor or non-rumor).

Description

Technical Field

[0001] The present invention relates to the field of social networks, and in particular to an early detection system for social network rumors based on a level-aware attention GRU network. Background Art

[0002] Social networks have grown rapidly in recent years, providing people with access to more information. However, this has also facilitated the mass production and dissemination of rumors at low cost. Rumors not only seriously mislead people and damage their spiritual well-being, but also pose a further threat to public safety. Early detection of rumors and cleaning up the online environment are urgent.

[0003] On social networks, detection latency is approximately linear with the lifecycle of a rumor. The longer a rumor spreads, the easier it is for users to detect it. However, by the time a rumor has spread widely and even caused significant damage, it's already too late. Therefore, detecting rumors in their early stages (i.e., early rumor detection) is crucial and necessary.

[0004] Most existing research on rumor detection in social networks relies on text or images. However, some rumors are carefully fabricated to mimic real news in order to be more convincing. Therefore, these detection methods that extract features solely from text or images have limitations. Recently, some researchers have discovered rumors based on social content, but they have overlooked the topological structure of rumor propagation. Rumors in social networks form a specific propagation structure due to user interactions (such as forwarding, following, and commenting).

[0005] While existing rumor detection methods achieve high detection accuracy after a rumor has been widely forwarded, they suffer from a significant problem: they ignore the varying contributions of tweet nodes across different time spans and forwarding hops. Specifically, first, tweets that directly forward the original tweet are often more relevant to the news topic, while the content of indirect forwarding may deviate from the news topic. Second, social platforms' recommendation systems ensure that users who engage with a topic within a short period of time are often more interested in or knowledgeable about it, thus often having a greater voice. Therefore, it is necessary to give varying degrees of attention to tweet content. Summary of the Invention

[0006] The purpose of the present invention is to provide a social network rumor early detection system based on a level-aware attention GRU network, which includes a news event acquisition module, a forwarding level measurement module, a basic feature extraction module, a tweet feature representation module and a rumor detection module.

[0007] The news event acquisition module acquires the dissemination information of the news event i to be detected in the social network and transmits it to the forwarding level measurement module.

[0008] The forwarding level measurement module establishes the propagation structure of news event i based on its propagation information and calculates the forwarding level of each node in the propagation network. The forwarding level measurement module transmits the propagation structure and the forwarding level of each node to the basic feature extraction module and the tweet feature representation module.

[0009] The basic feature extraction module extracts the basic features of each node in the propagation structure and propagates them to the tweet feature representation module.

[0010] The tweet feature representation module updates the node features according to the basic features and forwarding level of each node and transmits them to the rumor detection module.

[0011] The rumor detection module stores a rumor detection model.

[0012] The rumor detection model detects node features and obtains node tweet detection results.

[0013] Furthermore, the propagation structure is denoted as G i =<V i ,E i >,V i is a set of nodes, E i is the set of edges between interacting nodes.

[0014] Furthermore, the nodes in the propagation structure Forwarding level As shown below:

[0015]

[0016] Where, Representation node The forwarding level. l is the dimension of the forwarding level feature vector. It is a join operation. is the transformation matrix that transforms the fusion features into a high-dimensional vector space, is the bias. Indicates the timespan of the tweets. Indicates the number of forwarding hops.

[0017] Furthermore, the basic features include user features and text features

[0018] The text features As shown below:

[0019]

[0020] Where, is the text feature vector. c is the dimension of the vector. u is the number of columns. The first singular vector of the matrix .

[0021] Among them, from the node Sentence vector As shown below:

[0022]

[0023] Where S i From the node a is a hyperparameter. p(w) is the estimated probability of word vector w.

[0024] Furthermore, the tweet feature representation module updates node features according to the basic features and forwarding level of each node, including:

[0025] 1) Combine text features and user features to obtain node feature vectors Right now:

[0026]

[0027] Where, is the node feature vector. t is the dimension of the node feature vector.

[0028] 2) Based on the node's forwarding level, calculate the node feature vector Attention score Right now:

[0029]

[0030] Where, is the weight. For bias.

[0031] 3) Based on attention score Optimize the node feature vector to obtain the optimized node feature Right now:

[0032]

[0033] Where n is the number of all nodes in the event.

[0034] Furthermore, the rumor detection model includes a node forward propagation model and a node detection model.

[0035] The node features are input into the node forward propagation model, and the output node The hidden vector

[0036] node The hidden vector The node prediction model is input and the detection result of the node tweet is output. The node tweet detection result includes whether the node tweet is a rumor or not.

[0037] Furthermore, the node forward propagation model is as follows:

[0038]

[0039]

[0040]

[0041]

[0042] Where, [W * ,U * ] is the weight connection inside GRU. Representation node The latent vector of . ⊙ represents element-wise multiplication. i j Represents the reset gate. represents the update gate.

[0043] Furthermore, the node detection model is as follows:

[0044]

[0045] Where, is the hidden vector when the threshold is k. i and b i are the weights and biases of the output layer. is the output.

[0046] Furthermore, the loss function L(θ, k) of the node detection model is as follows:

[0047]

[0048] Where, is the rumor classifier. Y is the news label set.

[0049] It is worth noting that the present invention proposes a rumor early detection model (GLA4RD) with a level-aware attention mechanism based on a GRU neural network. The system first extracts the forwarding time span and forwarding hop count of each node in the propagation structure, and then combines these two features to measure the priority level of the node. At the same time, GLA4RD extracts text feature vectors and user feature vectors from text content and user profiles, and combines them into tweet feature vectors. In order to promote the early detection of rumors, the present invention also proposes a new level-aware attention mechanism to update the feature vectors of nodes. These features are then used as input to each unit of the GRU classifier in time series, and the final output vector is used to classify news as rumors or non-rumors.

[0050] The technical effects of the present invention are undoubted, and the beneficial effects of the present invention are as follows:

[0051] 1) This paper proposes a rumor early detection model (GLA4RD) with a level-aware attention mechanism based on a GRU neural network, which integrates text features and user features, and measures the propagation nodes according to the propagation process to give different attention.

[0052] 2) This paper proposes a hierarchical awareness attention mechanism to mine the hidden features of the propagation structure from the time span and forwarding hops, and update the node features.

[0053] 3) The present invention extracts textual and user features from a given tweet, generating a unified tweet node representation for each tweet. Simultaneously, the present invention simulates the propagation structure of news dissemination, calculating the time span and number of forwarding hops between the source tweet and the forwarding, and encoding this as the level of the tweet forwarding node. In summary, the present invention proposes a new system based on an attention mechanism that focuses on more important tweet content based on the forwarding level of the node. A GRU-based neural network classifier learns the overall characteristics of tweets from the time series to predict the label (rumor or non-rumor). BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 For system workflow;

[0055] Figure 2 For the propagation diagram;

[0056] Figure 3 (a) Comparison results of GLA4RD and baseline when the number of retweets is less than 80 in Weibo; Figure 3 (b) Comparison results of GLA4RD and the baseline when the number of retweets is less than 80 on Twitter;

[0057] Figure 4 Comparison results between GLA4RD-LA and GLA4RD under different k settings. DETAILED DESCRIPTION

[0058] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.

[0059] Example 1:

[0060] See also Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 ,The social network rumor early detection system based on level-aware attention GRU network,,including news event acquisition module, forwarding level measurement module,,basic feature extraction module, tweet feature representation module and rumor detection module.

[0061] The news event acquisition module acquires the dissemination information of the news event i to be detected in the social network and transmits it to the forwarding level measurement module.

[0062] The forwarding level measurement module establishes the propagation structure of news event i based on its propagation information and calculates the forwarding level of each node in the propagation network. The forwarding level measurement module transmits the propagation structure and the forwarding level of each node to the basic feature extraction module and the tweet feature representation module.

[0063] The basic feature extraction module extracts the basic features of each node in the propagation structure and propagates them to the tweet feature representation module.

[0064] The tweet feature representation module updates the node features according to the basic features and forwarding level of each node and transmits them to the rumor detection module.

[0065] The rumor detection module stores a rumor detection model.

[0066] The rumor detection model detects node features and obtains node tweet detection results.

[0067] The propagation structure is denoted as G i =<V i ,E i >,V i is a set of nodes, E i is the set of edges between interacting nodes.

[0068] Nodes in the propagation structure Forwarding level As shown below:

[0069]

[0070] Where, Representation node The forwarding level. l is the dimension of the forwarding level feature vector. It is a join operation. is the transformation matrix that transforms the fusion features into a high-dimensional vector space, is the bias. Indicates the timespan of the tweets. Indicates the number of forwarding hops.

[0071] The basic features include user features and text features

[0072] The text features As shown below:

[0073]

[0074] Where, is the text feature vector. c is the dimension of the vector. u is the number of columns. The first singular vector of the matrix .

[0075] Among them, from the node Sentence vector As shown below:

[0076]

[0077] Where S i From the node a is a hyperparameter. p(w) is the estimated probability of word vector w.

[0078] The step of the tweet feature representation module updating the node features according to the basic features and forwarding level of each node includes:

[0079] 1) Combine text features and user features to obtain node feature vectors Right now:

[0080]

[0081] Where, is the node feature vector. t is the dimension of the node feature vector.

[0082] 2) Based on the node's forwarding level, calculate the node feature vector Attention score Right now:

[0083]

[0084] Where, is the weight. For bias.

[0085] 3) Based on attention score Optimize the node feature vector to obtain the optimized node feature Right now:

[0086]

[0087] Where n is the number of all nodes in the event.

[0088] The rumor detection model includes a node forward propagation model and a node detection model.

[0089] The node features are input into the node forward propagation model, and the output node The hidden vector

[0090] node The hidden vector The node prediction model is input and the detection result of the node tweet is output. The node tweet detection result includes whether the node tweet is a rumor or not.

[0091] The node forward propagation model is as follows:

[0092]

[0093]

[0094]

[0095]

[0096] Where, [W * ,U * ] is the weight connection inside GRU. Representation node The latent vector of . ⊙ represents element-wise multiplication. i j Represents the reset gate. represents the update gate.

[0097] The node detection model is as follows:

[0098]

[0099] Where, is the hidden vector when the threshold is k. i and b i are the weights and biases of the output layer. is the output.

[0100] Furthermore, the loss function L(θ, k) of the node detection model is as follows:

[0101]

[0102] Where, is the rumor classifier. Y is the news label set.

[0103] The model GLA4RD proposed in this embodiment mainly consists of four modules: basic feature extraction, forwarding level measurement, tweet feature representation and rumor detection (such as Figure 1 (As shown in the figure). The early rumor detection task of this invention is based on the first k retweets. The threshold for measuring the detection deadline is the number of retweets, rather than the traditional absolute time. Because many hot topics experience explosive growth in the number of tweets within a short period of time, generating significant social impact, the early rumor detection model strives for high efficiency when the spread range is small.

[0104] Example 2:

[0105] See also Figure 1 and Figure 2 Using the social network rumor early detection system disclosed in Example 1 as a tool, a social network rumor early detection method based on a level-aware attention GRU network includes the following contents:

[0106] 1) Problem Formulation

[0107] Let ε = {E1, E2, ..., E |ε |} is a series of news events in social networks, each event is associated with a label y i ∈{0, 1}, where when E i Rumors and rumors i =0, it is a rumor i =1. is the jth retweet in chronological order, where Represents text content and user portrait files (for example: user name, location, followers, etc.). Refers to the first tweet of news event i (i.e., source tweet). G i Refers to the propagation structure of news event i. Given a dataset, the goal of rumor detection is to learn a classifier:

[0108] f:(ε,k)→Y, (1)

[0109] Where Y is a set of labels for news events, and k is the maximum detection threshold that can be set.

[0110] 2) Forwarding level metrics

[0111] The spread of news events is forwarded and promoted on social networks, which can be simulated as a spread graph (see Figure 2 The depth of the background color indicates the passage of time. The arrow points from the responding node to the responded node, and the depth of the node color indicates the level of the forwarding microblog node. The task of this stage is to generate the forwarding level of the node.

[0112] The propagation graph is represented as graph G i = <V i , E i >, V i is a set of nodes, E i is the set of edges between interacting nodes. in represents the timespan of the tweet, Indicates the number of forwarding hops.

[0113] The forwarding level of each node is calculated as follows:

[0114]

[0115] in Represents a tweet node Forwarding level, d l is the dimension of the forwarding level feature vector. is a join operation, is the transformation matrix that transforms the fusion features into a high-dimensional vector space, is the bias.

[0116] 3) Basic feature extraction

[0117] Every tweet in a social network contains textual content and is associated with the user who posted it.

[0118] Text features:

[0119] Word vectors It is extracted from the text sentence by word2vec, where d w The dimension of the pointing quantity. In previous rumor detection models, neural network models such as CNNs were often used to learn text features. However, most word embedding methods attempt to obtain word co-occurrence probabilities through vector inner products, resulting in unnecessarily large inner products for word pairs. The sentence vectors in this invention are extracted using a simple yet effective method, Sentence2Vec, which emphasizes the characteristics of low-frequency words while retaining the content of high-frequency words:

[0120]

[0121]

[0122] in is the text feature vector, where d c is the dimension of the vector, S i Refers to the tweet node a is a parameter, p(w) is the estimated probability of the word, and u is the number of columns. The first singular vector of the matrix .

[0123] User characteristics:

[0124] The method of embedding user feature vectors is the method mentioned in FNED. All user features are normalized to the range of [0,1] using Z-score to obtain the user feature vector of the user who posted the tweet

[0125] 4) Tweet Feature Representation

[0126] Combining text features and user features, the basic features are expressed as follows:

[0127]

[0128] is the basic eigenvector, where d x is the dimension of the vector, It is a join operation.

[0129] Level-aware attention mechanism

[0130] The system needs to know the importance of each tweet node and assign different weights to them based on their rank. To address this problem, a rank-aware attention mechanism is proposed. It is an extension of the basic attention mechanism.

[0131] For each basic feature vector obtained in Formula 5 Its attention weight and transformed vector are calculated as follows:

[0132]

[0133]

[0134] in Concatenates the basic feature vector and retweet level of the tweet. w is a weight Bias The attention score function of . Relu is the linear rectification function.

[0135] 5) Rumor Detection

[0136] GRU Neural Network:

[0137] To obtain enhanced feature vectors, a gated recurrent unit (GRU) is used to learn the characteristics of news events. In a tweet sequence, indicative features are aggregated along with the propagation history. Extending the GRU to simulate long-distance interactions between nodes reduces the required parameters while improving efficiency. The forward propagation process for node j is as follows:

[0138]

[0139] in is the node representation enhanced by the level-aware attention mechanism, [W * , U * ] is the weight connection inside GRU, Representation node The hidden state of . As with the standard GRU, ⊙ represents element-wise multiplication; reset gate Decide how to combine the current input with the memory of the previous tweet, update gate Defines the memory size for concatenating the previous tweet to the current node; Represents the candidate activation of the hidden state of the current node. This allows the model of this embodiment to merge representations from different tweet nodes.

[0140] After recursive aggregation of the time series, the final output state can be regarded as a representation of the news event for classification. Dropout is applied to the GRU layer to avoid overfitting.

[0141] Classification:

[0142] Finally, a multi-layer perceptron (MLP) block is used to predict the class label of the news event, which can be simply written as:

[0143]

[0144] in is the hidden vector when the threshold is k, W i and b i are the weights and biases of the output layer.

[0145] optimization:

[0146] set up is a rumor classifier based on a GRU neural network with a level-aware attention mechanism, and θ represents all parameters involved. Let Y be the news label set. The cross entropy function measures the detection loss:

[0147]

[0148] The optimization problem is solved using an optimization method based on stochastic gradient descent.

[0149] Example 3:

[0150] A computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for early detection of social network rumors based on a level-aware attention GRU network as described in Example 2 is implemented.

[0151] Example 4:

[0152] An electronic device comprising: one or more processors, a storage device;

[0153] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the social network rumor early detection method based on the level-aware attention GRU network described in Example 2.

[0154] Example 5:

[0155] Experiments to verify the early detection system or method for social network rumors described in Examples 1-4:

[0156] This embodiment experiments on two real datasets and performs ablation studies on each proposed component. This embodiment implements the system using Tensorflow. When preprocessing the text, the jieba tool is used to segment and regularize the characters. The effectiveness of the model is verified by five-fold cross-training and testing. In each round of cross-validation, the entire dataset is randomly divided into five parts of equal size. This embodiment trains the model 1000 times to minimize its loss. This embodiment uses stochastic gradient descent and Adadelta update to update the weights and biases.

[0157] Specifically:

[0158] 1) Dataset

[0159] This example conducts experiments on two real-world datasets: Twitter and Weibo. Weibo contains 4,664 news events, while Twitter contains 680 news events. Both Weibo and Twitter provide user profiles, including name length, number of followers, number of friends, city ID, and verification status. The Twitter dataset does not provide the text content of retweets, so this example uses the content of the source tweets as the text content of each retweet.

[0160] This example constructs a propagation graph based on the message ID and source ID of tweets and designs a simple program to obtain the forwarding hop count of each tweet node by traversing all nodes on the graph path. Table 1 shows the statistics of the two datasets:

[0161] Table 1 Statistics of the dataset

[0162]

[0163] 2) Baseline

[0164] This example compares GLA4RD with the following set of baseline models:

[0165] CSI: A deep hybrid model based on three features: text content, social content, and time series. It identifies rumors through article and user characteristics.

[0166] RvNN: A model based on tree-structured RNN as a rumor detection framework, which includes bottom-up and top-down models.

[0167] FNED: A deep learning framework that proposes a novel attention mechanism and CNN-based pooling method to improve the effectiveness of rumor detection by leveraging user characteristics of news disseminators.

[0168] RDLNP: A rumor detection framework based on linear and nonlinear propagation domains that integrates various characteristics of news.

[0169] 3) Compared with baseline

[0170] Table 2 Hyperparameter Set

[0171]

[0172] Table 2 lists the hyperparameters of the models. This example uses standard effectiveness metrics, including accuracy, precision, recall, and F1 score, all of which are used to evaluate all models. This example uses the number of retweets as the detection threshold. Table 3 shows the performance comparison of GLA4RD and the baseline model when the number of retweets reaches 10, that is, when k=10. GLA4RD outperforms the baseline model in all evaluation metrics. This example then compares GLA4RD with the best baseline FNED when the number of retweets is less than 80 (that is, k∈[10,80]), and obtains an overall effect (see Figure 3 (a) Figure 3 (b)).

[0173] Table 3 Performance comparison of GLA4RD and baseline models when k=10

[0174]

[0175] Through these experiments, this example makes the following observations: GLA4RD performs well in rumor detection, especially in the early stages of a rumor. More specifically, GLA4RD improves accuracy by 2.75% and 1.21% on Twitter and Weibo, respectively. In terms of precision, GLA4RD outperforms the strongest baselines by 3.21% and 2.55%, respectively. This suggests that by assigning different levels of attention to tweet nodes, as measured by time span and number of retweets, rumor detection can be effective.

[0176] 4) Ablation analysis

[0177] This example also evaluates several simplified variants of GLA4RD in Weibo, each of which removes a key component; the purpose of this ablation study is to investigate the impact of each key component on performance. The following is a list of simplified internal models:

[0178] GLA4RD-LA: does not include the level-aware attention layer. Table 4 shows the comparison between the simplified model and the full GLA4RD model when k = 10.

[0179] Table 4 Comparison of the simplified model and the complete GLA4RD model when k = 10

[0180]

[0181] GLA4RD-TS: Only considers the forwarding hop count and ignores the time span.

[0182] GLA4RD-RH: Only consider the time span and ignore the forwarding hops.

[0183] In addition, this example also compares GLA4RD-LA and GLA4RD under different k settings (see Figure 3 ). The ablation experiment results show that:

[0184] When a key component is deleted, the performance of the model proposed in this embodiment will degrade.

[0185] Among the two key factors considered by GLA4RD, time span and forwarding hops, the impact of forwarding hops on detection accuracy is more significant.

[0186] When the detection threshold is large, GLA4RD-LA can achieve good detection results, while when the detection threshold is small, the effect is significantly reduced. This further shows that measuring tweet nodes based on time span and forwarding hop count is effective for early detection of rumors.

[0187] 5) Conclusion

[0188] This paper proposes a new early rumor detection model based on a GRU neural network with a level-aware attention mechanism. The system is designed to be used for early rumor detection in social networks by measuring each tweet node according to the propagation structure. The idea is to reveal the hidden features of the propagation structure. GLA4RD integrates the propagation structure, text content, and user portrait files, and proposes a level-aware attention mechanism to give different attention to tweet nodes. This allows the model to maintain the same detection performance at different stages of propagation and also performs well in the early stages of propagation. Finally, a GRU-based classifier is used to predict the label. Extensive experiments on two real datasets also demonstrate the effectiveness of the method of this embodiment, and an ablation study quantitatively verifies that each component makes an important contribution. In the future, optimizing the measurement method of tweet nodes may be a promising way to further improve early detection performance.

Claims

1. A social network rumor early detection system based on level-aware attention GRU network, characterized by: It includes news event acquisition module, forwarding level measurement module, basic feature extraction module, tweet feature representation module and rumor detection module; The news event acquisition module acquires the dissemination information of the news event i to be detected in the social network and transmits it to the forwarding level measurement module; The forwarding level measurement module establishes the propagation structure of news event i based on the propagation information of news event i and calculates the forwarding level of each node in the propagation network; the forwarding level measurement module transmits the propagation structure and the forwarding level of each node to the basic feature extraction module and the tweet feature representation module; The basic feature extraction module extracts the basic features of each node in the propagation structure and propagates them to the tweet feature representation module; The tweet feature representation module updates the node features according to the basic features and forwarding level of each node and transmits them to the rumor detection module; The rumor detection module stores a rumor detection model; The rumor detection model detects node features and obtains node tweet detection results; Nodes in the propagation structure Forwarding level As shown below: Where, Representation node Forwarding level; d l is the dimension of the forwarding level feature vector; It is a join operation; is the transformation matrix that transforms the fusion features into a high-dimensional vector space, is the bias; node represents the time span of the tweet; Indicates the forwarding hop count; The step of the tweet feature representation module updating the node features according to the basic features and forwarding level of each node includes: 1) Combine text features and user features to obtain node feature vectors Right now: Where, is the node feature vector; d t is the dimension of the node feature vector; 2) Based on the node's forwarding level, calculate the node feature vector Attention score Right now: Where, is the weight; is bias; 3) Based on attention score Optimize the node feature vector to obtain the optimized node feature Right now: Where n is the number of all nodes in the event; The rumor detection model includes a node forward propagation model and a node detection model; The node features are input into the node forward propagation model, and the output node The hidden vector node The hidden vector The node prediction model is input and the detection result of the node tweet is output; the node tweet detection result includes whether the node tweet is a rumor or not.

2. The social network rumor early detection system based on level-aware attention GRU network according to claim 1 is characterized by: The propagation structure is denoted as G i =<V i ,E i >,V i is a set of nodes, E i is the set of edges between interacting nodes.

3. The social network rumor early detection system based on level-aware attention GRU network according to claim 1 is characterized by: The basic features include user features and text features The text features As shown below: Where, is the text feature vector; d c is the dimension of the vector; u is the number of columns The first singular vector of the matrix; Among them, from the node Sentence vector As shown below: Where S i From the node a is a hyperparameter; p(w) is the estimated probability of word w.

4. The social network rumor early detection system based on level-aware attention GRU network according to claim 1 is characterized in that: The node forward propagation model is as follows: Where, [W * ,U * ] is the weight connection inside GRU; Representation node The latent vector of ;⊙ represents element-wise multiplication; r i j Reset gate. represents the update gate.

5. The social network rumor early detection system based on level-aware attention GRU network according to claim 1 is characterized in that: The node detection model is as follows: Where, is the hidden vector when the threshold is k; W i and b i are the weights and biases of the output layer; is the output.

6. The social network rumor early detection system based on level-aware attention GRU network according to claim 5 is characterized in that: The loss function L(θ,k) of the node detection model is as follows: Where, is the rumor classifier; Y is the news label set.

Images