A rumor propagation control method across domains based on graph structure migration

Abstract

The application belongs to the field of network public opinion analysis, and particularly relates to a cross-domain rumor propagation control method based on graph structure migration, comprising the following steps: acquiring data to be propagated in real time, and extracting relevant attributes of the data to be propagated; processing the relevant attributes by using a URR2vec algorithm to obtain user forwarding behavior driving force; calculating user forwarding behavior influence by using a trained graph convolutional neural network according to the user forwarding behavior driving force; determining a rumor propagation trend according to the user forwarding behavior influence, and controlling the rumor according to the propagation trend; the application extracts relevant features such as rumor topic importance, rumor topic heat, user activity, user emotion influence index and user intimacy of the data to be propagated, analyzes the data to be propagated through the extracted features, and makes the predicted rumor propagation trend more accurate.

Description

Technical Field

[0001] The present invention belongs to the field of network public opinion analysis, and specifically relates to a cross-domain rumor propagation control method based on graph structure migration. Background Art

[0002] With the rapid development of the internet, social media has become a convenient online platform for users to access information, express opinions, and communicate with one another. As more and more people engage in discussions and exchanges on hot topics on social media, various rumors have also emerged. Online rumors can not only spread rapidly through social networks, but can also have a significant impact, posing a threat to the stability and maintenance of a harmonious society. Compared to traditional mainstream media such as newspapers and television, the freedom, interactivity, diversity, speed, and mass reach of social media make rumors more likely to arise, spread more quickly, and cause greater harm to society.

[0003] Furthermore, the complexity of social network structures and the difficulty of tracing large groups present significant challenges for researchers in conducting in-depth research on the patterns of online rumor propagation, influencing factors, and user behavior. The spread of rumors can cause serious social harm, especially within social networks. Traditional rumor research has not considered the impact of topic network structures and the diverse domains to which topics belong on the overall evolution of a topic. Therefore, research on cross-domain rumor propagation prediction models based on transfer learning can better uncover the underlying motivations behind the spread of newly emerged or less data-intensive rumor topics, enabling predictions of rumor trends and helping public opinion agencies effectively regulate rumor-related sentiment, address issues, and appease public opinion.

[0004] In recent years, numerous scholars have conducted extensive research on rumor propagation prediction models, focusing on the relationship between user behavior and the sentiment value of rumor topics, the rumor propagation spatial network, and the applicability of these models. These models are primarily based on graph structure transfer, machine learning, and deep learning algorithms. Graph structure transfer algorithms fully account for the non-Euclidean structure of the rumor propagation space and the significant differences between the graph structures of different rumor topics. They refine the intrinsic geometric information learned in the source topic propagation space and fully apply it to the target domain, thereby minimizing the common assumption that training and test data should come from the same feature space and distribution. Machine learning-based models primarily extract user and text features that influence user propagation, transforming the problem into a classification or regression problem. Machine learning algorithms have the advantage of processing massive amounts of data and are suitable for addressing complex problems in social networks.

[0005] However, all of the above methods have some problems, including: (1) the complexity and high dimensionality of rumor message propagation space, and the content of topic information is relatively fragmented. Traditional user text analysis cannot accurately obtain the theme of the information, resulting in an inability to accurately predict the attractiveness of the topic to users, thus affecting the accuracy of user forwarding predictions; (2) Rumor messages have relatively obvious emotional tendencies. How to quantify this emotional value and use it to increase the accuracy of rumor forwarding predictions; (3) For different topic areas, such as politics and food safety, the core words used are very different, and the users who may participate in forwarding are also different. This means that the rumor feature distribution and rumor topic structure characteristics in different fields are different, and directly applying the basic model will affect the perception of the rumor propagation situation. Summary of the Invention

[0006] In response to the problems existing in the above-mentioned existing technologies, the present invention proposes a cross-domain rumor propagation control method based on graph structure migration, which includes: obtaining the data to be propagated in real time and extracting relevant attributes of the data to be propagated; using the URR2vec algorithm to process the relevant attributes to obtain the driving force of user forwarding behavior; using a graph convolutional neural network to calculate the influence of user forwarding behavior based on the driving force of user forwarding behavior; determining the rumor propagation trend based on the influence of user forwarding behavior, and controlling the rumor based on the propagation trend.

[0007] Preferably, the relevant attributes of the data to be disseminated include the importance of the rumor topic, the popularity of the rumor topic, user activity, the user emotional impact index and user intimacy; the importance of the rumor topic is composed of the number of people who forwarded the rumor at time i and the proportion of its certified users; the popularity of the rumor topic includes calculating the user's rumor topic popularity using a heat cooling function; user activity reflects whether the user will participate in forwarding the topic in the near future; the user emotional impact index is obtained by analyzing the emotional value of the user's historical forwarding topics; and user intimacy is the trust and activity of two related users.

[0008] Preferably, the process of processing the relevant attributes using the URR2vec algorithm includes:

[0009] Step 1: Obtain the basic attributes of all nodes in the source topic propagation network. The basic attributes include the social platform user-topic forwarding set, the user relationship network, and the basic information set of the rumor topic;

[0010] Step 2: Use the node2vec algorithm to perform a random walk on the user network based on the basic attributes of all nodes to obtain a low-dimensional user relationship network and the implicit relationships between users;

[0011] Step 3: Use the masked language model to process the input topic; the specific process includes: taking the semantic vector representation of the target topic and each word in the context as the input of the model, obtaining the vector representation of the target topic, the vector representation of each word in the context, and the original representation of the target topic and each word in the context through linear transformation; calculating the similarity between the topic vector and each word vector, and using the similarity as the weight; performing weighted fusion on the topic vector of the target word, the word vector of each context word, and the implicit relationship between users according to the weight to obtain a word / sentence vector that conforms to the context and logic; creating a sentiment dictionary based on the word / sentence vector that conforms to the context and logic, and quantifying the topic sentiment value through the dependency relationship between word / sentence vectors; the topic sentiment value is the driving force value of the user's forwarding behavior.

[0012] Preferably, the process of using graph convolutional neural network to calculate the influence of user forwarding behavior includes:

[0013] Step 1: Construct a rumor propagation tree based on the rumor topic, use the rumor propagation tree to process the driving force of user forwarding behavior, and obtain the relevant adjacency matrix and feature matrix;

[0014] Step 2: Input the adjacency matrix and feature matrix into the graph convolutional neural network, and use Drop Edge to randomly delete edges from the input adjacency matrix and feature matrix to obtain the result of the binary classification problem, which includes whether the topic is forwarded or not forwarded;

[0015] Step 3: Use transfer learning to transfer the intrinsic geometric structure and related attribute features of the rumor topic propagation graph learned from the source topic domain to the target domain, use the migrated target domain model to predict the topic, and obtain the influence of user forwarding behavior.

[0016] Beneficial effects of the present invention:

[0017] The present invention extracts relevant features such as the importance of rumor topics, the popularity of rumor topics, user activity, user emotional influence index and user intimacy from the data to be spread, and analyzes the data to be spread through the extracted features, so that the predicted rumor propagation trend is more accurate; the present invention adopts a new URR2vec algorithm to calculate relevant attributes, and obtains the driving force of user forwarding behavior, and more accurately predicts the spread trend of rumors based on the driving force of user forwarding behavior, thereby improving the accuracy of the prediction; the present invention adopts a graph structure migration-based method to train the model, so that the model can adapt to different topic areas, improves the training efficiency of the model, and helps to minimize the problems caused by lack of data and incomplete structural information for new tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1A flowchart of data processing for the cross-domain rumor propagation prediction model based on transfer learning of the present invention;

[0019] Figure 2 This is the user forwarding behavior prediction graph of the present invention;

[0020] Figure 3 A schematic diagram of the present invention using the URR2vec algorithm to represent hidden information;

[0021] Figure 4 This is a graph convolutional neural network structure diagram of the present invention;

[0022] Figure 5 Schematic diagram of the model domain adaptation of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] A cross-domain rumor propagation control method based on graph structure migration includes: acquiring the data to be propagated in real time and extracting relevant attributes of the data to be propagated; processing the relevant attributes using the URR2vec algorithm to obtain the driving force of user forwarding behavior; calculating the influence of user forwarding behavior using a trained graph convolutional neural network based on the driving force of user forwarding behavior; determining the rumor propagation trend based on the influence of user forwarding behavior, and controlling the rumor based on the propagation trend.

[0025] A specific implementation method of a cross-domain rumor propagation control method based on graph structure migration, such as Figure 1 As shown, the method includes:

[0026] S1: Obtain data online and extract relevant attributes. Data can be obtained from public data websites or through mature social network public APIs. What needs to be obtained here is all the relevant participants of the rumor topic during its life cycle and its own information. The topic information that needs to be obtained is the time when the topic is forwarded and commented on, the basic information of the participating users, and the relationship information between the participating users (including follow and follow information); the information of the topic participants includes the information that the user has forwarded and commented on in history. Based on the obtained basic user information, user relationship network, and topic information, attribute extraction is performed from three perspectives: rumor attributes, users, and user relationships, and the influence of user behavior is quantified.

[0027] S2: Building a Basic Model. First, we propose the URR2vec algorithm to mine implicit relationships between users and between users and rumor topics. We then fuse the constructed user and topic features into a feature matrix to calculate the driving force behind user forwarding behavior. Finally, we construct the TSA-GCN using a graph convolutional neural network to analyze the influence of rumor topic sentiment and user relationships on user forwarding behavior, and use this information to study rumor propagation trends.

[0028] S3: Model domain adaptation. Using the knowledge in the source domain and the basic model trained for the target task (user forwarding behavior prediction), the intrinsic geometric information of the source topic domain graph structure learned during the training process is transferred and copied into the convolutional layer and pooling layer of the source domain target task training features; the prediction results of user forwarding are as follows: Figure 2 When completing the target task in the target domain, the model parameters are fine-tuned to minimize the impact of different feature distributions of topics belonging to different domains on the accuracy of rumor propagation prediction.

[0029] In this embodiment, extracting relevant attributes of the data to be disseminated includes: the method of obtaining data can be from a public data website, or by using a mature social network public API. What needs to be obtained here is all the relevant participants of the rumor topic during its life cycle and its own information. The topic information that needs to be obtained is the time when the topic is forwarded and commented on, the basic information of the participating users, and the relationship information between the participating users (including follow and follow information); the information of the topic participants includes the information that the user has forwarded and commented on in history. Based on the obtained basic user information, user relationship network, and topic information, attribute extraction is performed from three perspectives: rumor attributes, users, and user relationships, and the influence of user behavior is quantified. The specific steps include:

[0030] S11: Obtaining raw data: Raw data can be obtained through social network public APIs or public data websites.

[0031] S12: Simple data cleaning. The raw data typically obtained is unstructured and cannot be directly used for data analysis. Simple data cleaning can structure most unstructured data, eliminating outliers or null values ​​and reducing the inconvenience caused by subsequent calculations.

[0032] S13: Extracting relevant attributes. Exploring the rumor propagation process essentially involves extracting relevant features from both the message and user levels and modeling and analyzing these features. This paper addresses the implicit driving force behind user forwarding in rumor information dissemination networks. By examining the rumor's inherent attributes, users, and user relationships, this paper explores the key features of user message forwarding, specifically including:

[0033] S131: Importance of rumor topic

[0034] The main factors affecting the importance of a rumor topic are the number of people who forwarded the rumor at time i and the proportion of verified users. The more people who forwarded the rumor and the more verified users there were, the wider the rumor topic spreads. The higher the importance, the more likely users are to participate in forwarding. Therefore, the importance of a rumor topic is defined as:

[0035]

[0036] in, represents the number of people who are forwarding the rumor topic at the current moment, isVerified(user) represents the number of authenticated users in the forwarding group, and k and b (k+b=1) represent the weights of the two factors respectively.

[0037] S132: Rumor topic popularity

[0038] The popularity of a rumor topic will be affected by the intervention of user behavior and the passage of time, thus decaying or growing. Often, the popularity of a rumor topic will greatly affect user behavior. In order to more accurately quantify the popularity of a rumor topic at time i, a heat cooling function is introduced. The rumor topic heat is defined as:

[0039]

[0040] in, represents the initial heat of the topic, ∝ is the attenuation coefficient, Indicates the importance of the rumor topic at time i.

[0041] S133: User activity Active (user j )

[0042] User activity reflects to some extent whether the user will participate in the recent topic forwarding, which has a positive impact on predicting the user's next action. User activity is:

[0043]

[0044] Among them, Retweet(user j ) indicates the user's historical forwarding record, AcountAge(user j ) represents the existence time of the user account. Dividing the two is to more clearly reflect the user's recent activity level and weaken the impact of forwarding records with a longer time span. TimeDifference(userj ) ave The average value of the difference between the time of all forwarded messages of a user and their original time.

[0045] S134: User emotional impact index EmotionalRate (user j )

[0046] By analyzing the sentiment value of a user's historical forwarding topics, we can more clearly characterize the user's psychology when facing rumors and whether to forward or not, and further predict the user's behavior. The index of user's emotional influence is:

[0047]

[0048] Among them, SentiValue(Retweet(user j )) represents the average sentiment value of the user's historical forwarded topics, SentiValue(topic j ) represents the sentiment value of the current rumor topic. The closer the value is to 1, the more likely the user is to forward it.

[0049] S135: User intimacy Intimate (user i ,user i+1 )

[0050] Whether a user will forward the rumor topic is affected by his trust in the relevant user and his activity, and this influence can be determined by the next user activity (Active i+1 ) and the intimacy between two users Intimate(user i ,user i+1 ) to quantify. Based on UN, Intimate (user i ,user i+1 )Quantification formula:

[0051]

[0052]

[0053] Among them, k represents the total number of messages forwarded by the current user, t i Indicates the time when the next user forwards the current user's message, t k Indicates the release time of the current message, and α is the decay factor.

[0054] In this embodiment, the constructed basic model includes two stages: hidden information representation and basic model algorithm design. In the first stage, starting from the complex relationship between users and users, users and rumor topics in the social network platform, the URR2vec algorithm is proposed to mine the implicit relationship between users and users, users and rumor topic information, and the constructed user and topic features are integrated into a feature matrix to calculate the driving force of user forwarding behavior, such as Figure 3 In the second stage, we use graph convolutional neural networks to build TSA-GCN, analyze the influence of rumor topic sentiment and user relationships on user forwarding behavior, and study the rumor propagation trend based on this, as shown in the following example: Figure 4 shown.

[0055] The first stage is the hidden information representation, which includes structure representation and topic content representation.

[0056] In this embodiment, the structural representation includes: first, based on the advantages of node2vec in processing social network data structure and combining the node attribute characteristics, a new rumor space representation method is proposed. Secondly, the relationship between the network is regarded as a wandering process, and this process is easily affected by the intimacy and activity of the relevant users. For example, at the same time t, the activity of user a is higher than that of user b, then other users are more likely to contact user a. Based on this, a new wandering strategy is formulated: whether the user will forward the rumor topic will be affected by his trust in the relevant users and their activity, and this influence can be determined by the next user activity (Active(c i+1 ) and the intimacy between two users Intimate(c i ,c i+1 ) to quantify.

[0057] In order to better represent the rumor topic network as a vector, given a source node user0 in the current network, and taking it as the origin, simulate a random walk of fixed length L, so that r0 = user0, and let r i To traverse the i-th node in the network, r i+1 Represents the next node. The generation condition of r, that is, the probability of wandering to the next node, is specifically expressed as the formula:

[0058]

[0059] Where α(x,y) is the unnormalized transition probability between two nodes, and z is a normalization constant.

[0060] In this embodiment, the topic content representation includes: the BERT model takes into account the importance of bidirectional pre-training for topic representation, uses a masked language model to achieve pre-trained deep bidirectional representation, and can simultaneously consider the contextual relationship of the input topic, decompose it, and embed the contextual relationship into each word vector. In addition, when decomposing a sentence, the BERT model introduces the relative position of each word in the sentence as a reference point to facilitate a more complete analysis of the relationship between words, thereby obtaining a sentence vector that is more in line with the context and logic.

[0061] The specific process includes: first, based on UInfo, give S={s1,s2,s3,...,s |S|} is the original topic group; the text messages in the rumor topic are segmented and stop words are removed to reduce the impact of the noise generated by stop words on the subsequent model; since the messages in the topic are medium-length texts, and punctuation marks have a certain impact on the severity and authenticity of the content expressed in the text, regular punctuation marks are retained as part of the text features; each s i ∈S is a short text consisting of a word or a character sequence l i For S i The length of the message is finally input into the BERT model for representation, and the BERT model outputs the text feature vector V of the current message. text ∈R d .

[0062] The second stage involves designing the underlying model algorithm. Specifically, the network formed between tweet originators and corresponding responders naturally requires a technique that can leverage network characteristics. Therefore, to capture network properties, a graph neural network (GNN)-based approach is employed to identify potential rumor spreaders. Specifically, a graph convolutional neural network (GCN) is implemented within the framework of graph theory, implementing convolution operations on graphs. Therefore, this paper chooses GCN to process rumor topic data. Furthermore, considering the impact of rumor sentiment (text sentiment) on user forwarding psychology, a TSA-GCN rumor propagation prediction model is proposed. Finally, by integrating user factors, analyzing rumor information, and TSA-GCN results, this paper predicts user forwarding behavior. This paper defines rumor propagation prediction as a binary classification task, classifying forwarding behavior from individual to group, and analyzing rumor propagation trends.

[0063] The process of using graph convolutional neural networks to calculate the influence of user forwarding behavior includes:

[0064] Step 1: Construct a rumor propagation tree based on the rumor topic, use the rumor propagation tree to process the driving force of user forwarding behavior, and obtain the relevant adjacency matrix and feature matrix;

[0065] Step 2: Input the adjacency matrix and feature matrix into the graph convolutional neural network. Use Drop Edge to randomly delete edges from the input adjacency matrix and feature matrix. This randomly removes a certain ratio of edges to prevent overfitting and reduce message passing in the graph convolution to alleviate oversmoothing. The result of the binary classification problem is obtained, which includes the topic forwarding or the topic not forwarding.

[0066] Step 3: Use transfer learning to transfer the intrinsic geometric structure and related attribute features of the rumor topic propagation graph learned from the source topic domain to the target domain, thereby improving the universality of the model and thus improving the prediction accuracy of the model.

[0067] Specifically, we construct a propagation tree based on each rumor topic based on the forwarding relationship, and then let A∈R N×N and They are the adjacency matrix and feature matrix of the rumor propagation tree, and serve as the input of the model. The specific description is as follows:

[0068] Characteristic matrix TUN = N × F 0 , where N represents the number of user nodes in the rumor topic propagation network, F 0 It is the input feature dimension of each node, which is the user's own attribute characteristics after the data is represented by learning, the calculated rumor sentiment value, and the index of the user's emotional influence; the adjacency matrix UN = N × N, which represents the connection information between users in the rumor topic message propagation network.

[0069] In this embodiment, a two-layer graph convolutional network with a DropEdge layer is used as a model for predicting rumor propagation trends. The data processing process of this model includes: assuming that the total number of edges in the graph is N e , the drop rate is P, then the DropEdge that needs to be performed in each training period, the obtained UN′ can be expressed as:

[0070] UN′=UN-UN drop

[0071] Among them, UN drop is to use N randomly sampled from the original edge set e ×P constructed matrix.

[0072] The core calculation formula of TSA-GCN is:

[0073]

[0074] Among them, W(0) With W (1) is a layer-specific trainable weight matrix.

[0075]

[0076]

[0077] Among them, UN′ represents the adjacency matrix after the DropEdge algorithm, which represents the connection information between users in the rumor topic message propagation network, N e represents the total number of edges in the graph, P represents the drop rate in the DropEdge algorithm, and TUN represents the feature matrix composed of the user's own attribute characteristics after representation learning, the calculated rumor sentiment value, and the index of the user's emotional influence. Represents the adjacency matrix after normalization, ReLU represents the activation function, Represents a normalized matrix, each row of which is the degree of the adjacency matrix UN′. Indicates that a self-loop is manually added to each node to prevent the information of the node itself from being lost after obtaining the information of the surrounding nodes. I represents the identity matrix.

[0078] Let Z = RN t+1 (user i )=P(R,N|user i ), the specific situation is as follows:

[0079]

[0080] If the value of Y is 1, it means that the user i The rumor will be forwarded in the next time period, otherwise it will not be forwarded.

[0081] This paper uses the cross entropy loss function to calculate the loss. However, the traditional cross entropy loss function tends to favor the direction with more sample labels during training, causing the model to overfit. Therefore, this paper proposes a new cross entropy loss function, which is expressed as:

[0082] loss i =-(TrueP i ln P(R|user i )+(1-TrueP i )ln(1-P(R|user i )))

[0083] Among them, TrueP i is the actual label value of the current node, P(R|useri ) is the predicted output.

[0084] There are different rumors on the entire social network platform. These rumors have different focuses, and for different topic areas, such as politics and food safety, the core words are very different. This means that the feature distribution of rumors in different fields and the users who participate in forwarding are also different. After the deep learning model completes its training process, the topic feature space, distribution and intrinsic geometric information will be fixed in the model. Based on the above facts, when there is insufficient training data for the rumor topic to be predicted, directly using the previously proposed basic model will inevitably lose prediction accuracy. Therefore, a model domain adaptation solution based on graph structure and parameter migration is proposed, such as Figure 5 shown.

[0085] Transfer learning overcomes the common belief that training and test data should come from the same feature space and distribution. Transfer learning between different rumor topic domains can reduce the burden of training models for new topics. Given the importance of topic structure in graph analysis, the structural features learned by the base model are transferred from the source domain to the target domain, and the parameters of each layer are fine-tuned based on the characteristics of the topic itself.

[0086] The domain in transfer learning is the feature space and probability distribution Composition, of which Given a domain use Represents a task, which has a label space and from the training data The prediction function learned in where x∈X, The general goal of transfer learning is to use the source domain and source tasks to improve the target domain Prediction tasks in The present invention migrates the source topic domain graph learned from the previous step The intrinsic geometric information in and Based on this, the present invention can skip the Zhongwei generate The steps of extracting structural features. and In the case of structural similarity, by copying Included in The convolutional layer and pooling layer of the trained features are used to train middle This transfer learning approach improves learning efficiency and helps minimize problems caused by lack of data and incomplete structural information for new tasks.

[0087] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cross-domain rumor propagation control method based on graph structure migration, characterized by: include: Acquire the data to be disseminated in real time and extract its relevant attributes. Use the URR2vec algorithm to process the relevant attributes and obtain the driving force behind user forwarding behavior. Use a graph convolutional neural network to calculate the influence of user forwarding behavior based on the driving force. Determine the rumor propagation trend based on the influence of user forwarding behavior and control the rumor by spreading the trend. The relevant attributes of the data to be spread are extracted, including the importance of the rumor topic, the popularity of the rumor topic, user activity, the user emotional influence index, and user intimacy. The importance of the rumor topic is composed of the number of people who forwarded the rumor at time i and the proportion of its authenticated users. The popularity of the rumor topic includes calculating the user's rumor topic popularity using the heat cooling function. User activity reflects whether the user will participate in the forwarding of the topic in the near future. The user emotional influence index is obtained by analyzing the emotional value of the user's historical forwarding topics. User intimacy is the trust and activity of two related users. The calculation formula for the importance of rumor topics is: Among them, k and b represent the weights of the two factors respectively. Represents the number of people who forwarded the rumor topic at the current moment, isVerified(user) indicates the number of verified users in the forwarding group, topic j represents the jth rumor topic, user represents the user; The calculation formula for the popularity of rumor topics is: in, represents the initial heat of the topic, ∝ is the attenuation coefficient, represents the importance of the rumor topic at time i, t i Indicates that time i is the current time, and t0 indicates the starting time; The formula for user activity is: Among them, Retweet(user j ) indicates the user's historical forwarding record, AcountAge(user j ) indicates the existence time of the user account, TimeDifference(user j ) ave The average value of the difference between the time when all messages forwarded by the user are first sent and the time when they are first sent; The expression of the user's emotional influence index is: Among them, SentiValue(Retweet(user j )) represents the average sentiment value of the user's historical forwarded topics, SentiValue(topic j ) represents the sentiment value of the current rumor topic; The calculation formula for user intimacy is: Among them, k represents the total number of messages forwarded by the current user, t i Indicates the time when the next user forwards the current user's message, t k Indicates the release time of the current message, α is the decay factor; The process of processing related attributes using the URR2vec algorithm includes: Step 1: Obtain the basic attributes of all nodes in the source topic propagation network. The basic attributes include the social platform user-topic forwarding set, the user relationship network, and the basic information set of the rumor topic; Step 2: Use the node2vec algorithm to perform a random walk on the user network based on the basic attributes of all nodes to obtain a low-dimensional user relationship network and the implicit relationships between users; Step 3: Use the masked language model to process the input topic; the specific process includes: taking the semantic vector representation of the target topic and each word in the context as the input of the model, obtaining the vector representation of the target topic, the vector representation of each word in the context, and the original representation of the target topic and each word in the context through linear transformation; calculating the similarity between the topic vector and each word vector, and using the similarity as the weight; performing weighted fusion on the topic vector of the target word, the word vector of each context word, and the implicit relationship between users according to the weight to obtain a word / sentence vector that conforms to the context and logic; creating a sentiment dictionary based on the word / sentence vector that conforms to the context and logic, and quantifying the topic sentiment value through the dependency relationship between word / sentence vectors; the topic sentiment value is the driving force value of the user's forwarding behavior.

2. A cross-domain rumor propagation control method based on graph structure migration according to claim 1, characterized in that: The process of using graph convolutional neural networks to calculate the influence of user forwarding behavior includes: Step 1: Construct a rumor propagation tree based on the rumor topic, use the rumor propagation tree to process the driving force of user forwarding behavior, and obtain the relevant adjacency matrix and feature matrix; Step 2: Input the adjacency matrix and feature matrix into the graph convolutional neural network, and use Drop Edge to randomly delete edges from the input adjacency matrix and feature matrix to obtain the result of the binary classification problem, which includes whether the topic is forwarded or not forwarded; Step 3: Use transfer learning to transfer the intrinsic geometric structure and related attribute features of the rumor topic propagation graph learned from the source topic domain to the target domain, use the migrated target domain model to predict the topic, and obtain the influence of user forwarding behavior.

Images