An information dissemination prediction method based on user information and topic game relationship

Abstract

The present invention belongs to the field of public opinion analysis, and particularly relates to an information dissemination prediction method based on user information and topic game relationships. The method includes obtaining data and extracting relevant attributes to construct a user information matrix; establishing a full-user relationship network based on the obtained data, mining the influence of neighbor users on message dissemination, and thus constructing the fusion of the user relationship matrix and the user information matrix; designing a dynamic game strategy, calculating the driving force of the original topic and the driving force of the derivative topic in combination with the fusion result, and further obtaining the latest full-user relationship network; constructing a topic dissemination prediction model to predict the dissemination trend of the topic at the next moment. The present invention proposes a method that can more effectively perceive the influence of various factors such as preposed emotions and derivative topics on user behavior during the topic dissemination process and better predict the topic dissemination situation.

Description

Technical Field

[0001] The present invention belongs to the field of public opinion analysis, relates to user forwarding behavior analysis, and specifically relates to an information dissemination prediction method based on the game relationship between user information and topic. Background Art

[0002] The medium through which topics spread is constantly evolving. Traditionally, topic dissemination relies on word of mouth, which has a narrow reach and limited influence. With the rapid development of the internet, people are discussing various topics through social platforms, making information dissemination even faster. This also means that once a topic "explodes," it quickly spreads across major social platforms, and any single netizen's voice has the potential to become a hot topic. To effectively control public opinion, major social platforms such as Sina Weibo, Facebook, and Twitter are focusing on research into topic dissemination.

[0003] Current research on topic diffusion is primarily categorized into macro and micro levels. Macro-level research focuses on further optimizing infectious disease models while also predicting group behavior based on the driving forces of topic diffusion. With the advancement of machine learning and deep learning, the use of neural networks to explore individual behavior at the micro level has become commonplace. While numerous scholars have achieved significant results in topic diffusion research, the following challenges remain:

[0004] 1. The hidden nature of user prior emotions. The emotions a user experiences immediately before engaging with a topic become a potential factor influencing their subsequent behavior, so it is important to consider the driving force behind their prior emotions.

[0005] 2. Users have complex relationships during the spread of hot topics. Topics have derivative behaviors during the spread process, and are complex and multi-type. Topic spread is random, so the relationships between users are complex and diverse, making it difficult to uncover hidden relationships between users.

[0006] 3. A dynamic game of competition exists between topics. As topics spread, they generate multiple new topics. These new topics compete with the original topics, influencing the spread of topics. This driving force needs to be further quantified, and the dynamic timeframe of topic spread needs to be considered.

[0007] More importantly, Chung et al. proposed a theory-based model and a proof-of-concept system in their article "Dissecting emotion and user influence in social media communities: An interaction modeling approach" to dissect emotions and user influence in social media networks. Inspired by the above paper, the present invention proposes a complex topic information dissemination model. A game dissemination mechanism based on complex topics is constructed, which can better predict the topic dissemination trend. Summary of the Invention

[0008] To solve the above problems, the present invention takes as the starting points the content of hot topics and the complex association relationships among users, and mines the hidden information in two aspects: users' pre-emotions and user relationships. The evolutionary game theory is introduced to quantify the driving force of user behavior, and the pre-emotions of users are synthesized. Finally, according to the dynamics of topic dissemination, a time decay function is introduced to further optimize the prediction of user behavior.

[0009] The present invention proposes an information dissemination prediction method based on the game relationship between user information and topics, including the following steps:

[0010] S1. Download data sources related to the target native topic from existing Web-based research recommendation systems or obtain them using the public APIs of mature social platforms;

[0011] S2. According to the obtained data sources, extract the user attributes of the participating users. The user attributes include basic user attributes and user emotion indices, and establish a user information matrix based on the user attributes;

[0012] S3. According to the obtained data sources, construct a full user relationship network of participating users and potential users, calculate the neighbor message propagation influence based on the full user relationship network to obtain a user relationship matrix; fuse the user information matrix and the user relationship matrix to obtain a fusion matrix;

[0013] S4. Use dynamic game strategies to calculate the total benefits brought by forwarding the native topic and the derivative topic respectively, combine the fusion matrix to calculate the driving force of the native topic and the driving force of the derivative topic, and integrate the two driving forces into the full user relationship network to obtain the latest full user relationship network matrix;

[0014] S5. Use a two-layer graph convolutional neural network with an intermediate Dropout layer to construct a topic dissemination prediction model, input the latest full user relationship network matrix into the topic dissemination prediction model, and predict the dissemination trend of the target native topic at the next moment.

[0015] Furthermore, the data sources obtained include topic participation records, message forwarding status and behavior records of participating users within the life cycle of the target native topic; topic participation records include basic information, follow-up information and followed information of participating users, as well as topic forwarding time; behavior records of participating users include historical forwarding information and historical comment information of participating users.

[0016] Furthermore, user attributes include user basic attributes and user sentiment index.

[0017] The basic attributes of users are expressed as:

[0018] Attr(u i )=act(u i )×trans(u i )×cog(u i )

[0019] Among them, Attr(u i ) represents the participating user u i Basic attributes of users, act(u i ) represents the participating user u i Activity of trans(u i ) represents the participating user u i The historical forwarding rate, cog(u i ) represents the participating user u i awareness;

[0020] The user sentiment index is expressed as:

[0021]

[0022] Among them, Emo t (u i ) represents the participating user u i User sentiment index as of time t, Num[interact(u i ,t)] represents the participating user u i As of time t, the number of interactions with the target native topic; Num[interact(u i ,t-1)] represents the participating user u i As of time t-1, the number of interactions with the target native topic, Num[bebavior(u i ,t)] represents the participating user u i The total number of behaviors up to time t, Num[bebavior(u i ,t-1)] represents the participating user u i The total number of behaviors up to time t-1, α represents the impact factor, α∈[0,1].

[0023] Furthermore, participating user u i The activity is defined as:

[0024]

[0025] tm(u i ) represents the participating user u i Number of Internet access in the past month, tm(u p ) represents the maximum value of the number of online visits of all participating users in the past three months, tm(u q ) represents the minimum value among all participating users’ Internet access times within three months;

[0026] Participating user u i The historical forwarding rate is defined as:

[0027]

[0028] tr(u i ) represents the participating user u i The number of microblogs published and forwarded in the past three months, tr(u m ) represents the maximum value in the set of microblogs published and forwarded by all participating users in the past three months, tr(u n ) represents the minimum value in the set of the number of microblogs published and forwarded by all participating users in the past three months;

[0029] Participating user u i Awareness is defined as:

[0030]

[0031] Wr represents the participating user u i Keywords related to rumors in all Weibo content, Wh represents the number of users involved i The most frequently appearing words in Weibo content in the past three months.

[0032] Furthermore, a full user relationship network of participating users and potential users of the target native topic at time t is constructed Expressed as:

[0033]

[0034] Among them, U pa represents the set of users participating in the target native topic at time t, U po represents the potential user set of the target native topic at time t, Represents the set of social relationships between participating users and potential users;

[0035] Calculate the influence of neighbor messages generated by neighbor users on each participating user u in the entire user relationship network, expressed as: i

[0036]

[0037] Num(U opt ) represents the number of optimistic neighbor users of the participating user u i , Num(U pes ) represents the number of pessimistic neighbor users of the participating user u i , and β represents the attenuation coefficient.

[0038] Furthermore, the process of users selecting the best game strategy is characterized as the process of the spread of the original topic and the derivative topic in the social network. A dynamic game strategy is designed with the goal of high returns after participating users forward the topic. The dynamic game strategy consists of two parts

[0039] The first part is the game between the participating user and its neighbor users, expressed as:

[0040]

[0041]

[0042] pro1(u i ) represents the profit function obtained by the participating user u i selecting Strategy 1, that is, forwarding the target original topic, after the game in the first part; pro2(u i ) represents the profit function obtained by the participating user u i selecting Strategy 2, that is, forwarding the derivative topic related to the target original topic, after the game in the first part; M xy represents the specific profit when the participating user u i executes Strategy x = 1, 2 for the target original topic, and at the same time, the neighbor users of the participating user u i execute Strategy y = 1, 2; K represents the number of neighbor users of the participating user u i , represents the number of neighbor users who forward the target original topic of the participating user u i at time t - 1; represents the number of neighbor users who forward the derivative topic related to the target original topic of the participating user u i at time t - 1, and

[0043] The second part is the game between the participating user and itself, expressed as:

[0044] ​

[0045]

[0046] pro′1(u i ) represents the payoff function obtained by user u participating in the game in the second part and choosing Strategy 1, that is, forwarding the target original topic; pro′2(u i ) represents the payoff function obtained by user u participating in the game in the second part and choosing Strategy 2, that is, forwarding the derivative topic related to the target original topic; m i ) represents user u participating in the game i and choosing Strategy 2, that is, the payoff function obtained by forwarding the derivative topic related to the target original topic; m xy represents user u participating in the game i and executing Strategy x = 1, 2 in the previous topic participated in for the target original topic and executing Strategy y = 1, 2 for the target original topic; N represents the number of Weibo posts of user u i in the past three months, represents that at time t - 1, user u i forwards the number of target original topics, represents that at time t - 1, user u i forwards the number of derivative topics related to the target original topic.

[0047] Further, the total payoffs of each participating user choosing Strategy 1 and Strategy 2 are calculated according to the dynamic game strategy, expressed as:

[0048] pro nat (u i ) = pro1(u i ) × pro′1(u i )

[0049] pro der (u i ) = pro2(u i ) × pro′2(u i )

[0050] Among them, pro nat (u i ) represents the total payoff function obtained by user u i choosing Strategy 1, that is, forwarding the original topic; pro der (u i ) represents the total payoff function obtained by user u i choosing Strategy 2, that is, forwarding the derivative topic.

[0051] Further, the topic propagation prediction model is expressed as:

[0052]

[0053] Among them, X represents the user information matrix, G represents the user relationship matrix, and W (0) represents the weight matrix of the first-layer graph convolutional neural network, and W (1) represents the weight matrix of the second-layer graph convolutional neural network, represents the degree matrix transformation of G, that is, I is the identity matrix, and H (1) represents the output of the first-layer graph convolutional neural network, and Z represents the output result of the topic propagation prediction model.

[0054] Advantages of the present invention:

[0055] The present invention proposes an information propagation prediction method based on user information and topic game relationships, introducing sentiment analysis and game benefits between interaction behaviors. It can not only perform more accurate social network public opinion analysis but also distinguish what types of topics users want to forward. It mainly constructs a user attribute matrix and a user relationship matrix based on prepositional sentiment, combines them with game strategies, and uses two driving forces to analyze the revenue influence of public opinion propagation. Most previous scholars studied public opinion propagation through epidemic models, without analyzing users' prepositional sentiment and not delving deeply into the topics in public opinion. Usually, topics in our social networks are not only spread among those participating in the topic, but the topics will be processed during this process, and derivative topics will appear. The derivative topics may not be consistent with the original topics and may even be opposite, such as rumors, where there may be two situations: rumor refutation and rumor promotion. Which of these two situations has the optimal benefit for users requires the use of the optimal game strategy. In summary, we introduce the prepositional sentiment driving force and the topic game mechanism, and use a deep learning model to study the potential mechanism in the process of public opinion analysis, making up for the deficiencies of traditional epidemic models. Brief Description of the Drawings

[0056] Figure 1 is the process framework diagram of the embodiment of the present invention;

[0057] Figure 2 is the topic propagation schematic diagram of the embodiment of the present invention. Detailed Embodiment

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] Hot topics are the hottest issues that the public is most concerned about within a certain period of time and within a certain range, such as education, social security, medical care, the real estate market, the stock market, employment issues, etc. For example, Figure 2 As shown, hot topics will attract wide attention from users after rapid dissemination through social platforms. However, under people's dissemination and discussion, information will gradually deviate from the original version, spreading explosive topics to attract the attention of a large number of people. The impact of topics is immeasurable. Especially the authenticity of some topics is questionable, and a series of hot topics may bring troubles to the people involved in the topics.

[0060] The present invention provides an information dissemination prediction method based on user information and topic game relationship, including three major parts: online data acquisition, extraction of relevant attributes, and model establishment. Among them:

[0061] 1. Online data acquisition: The data that the present invention needs to acquire are topic participation records, message forwarding situations, and behavior records of participating users during the life cycle of hot topics; topic participation records include basic information, attention information, and followed information of participating users, as well as topic forwarding time; behavior records of participating users include historical forwarding information and historical comment information of participating users; use the acquired data to construct a user relationship matrix of participating users of the topic; the message forwarding situation refers to the forwarding data of the original topic and the derivative topic, as well as the competitive and cooperative relationship between the two types of topics.

[0062] 2. Extraction of relevant attributes: According to the data acquired online, extract attributes such as basic information, activity, historical forwarding rate, and sentiment index of participating users, construct a user information matrix of participating users according to the relevant attributes, and fuse the user relationship matrix and the user information matrix.

[0063] 3. Model establishment: First, design a dynamic game strategy to quantify the user behavior driving force of the original topic - derivative topic, and at the same time consider the user's pre - emotion driving force, and discretize continuous time; whether a user participates in topic discussion is related to the user's pre - emotion, friend interaction information, and the game between the original topic and the derivative topic, and construct a topic dissemination prediction model for public opinion analysis.

[0064] In an embodiment, as Figure 1 shown, an information dissemination prediction method based on user information and topic game relationship includes the following steps:

[0065] S1. Download data from an existing Web - based research - type recommendation system or obtain data sources using the public API of a mature social platform;

[0066] Specifically, generally speaking, the acquired original data is unstructured and cannot be directly used for data analysis. Therefore, simple data cleaning operations are adopted to convert the acquired data source into structured data, so that outliers or null values no longer appear, reducing the inconvenience of subsequent calculations.

[0067] Meanwhile, the data source after data cleaning is stored in a database, and the cleaned data source is further standardized through the table structure. Moreover, the database can greatly improve the data retrieval efficiency and the mapping of inter-table relationships.

[0068] S2. According to the acquired data source, extract the user attributes of the participating users. The user attributes include user basic attributes and user sentiment index, and establish a user information matrix based on the user attributes;

[0069] S3. According to the acquired data source, construct the full user relationship network of the participating users and potential users, and obtain the user relationship matrix based on calculating the influence of neighbor message propagation in the full user relationship network; fuse the user information matrix and the user relationship matrix to obtain the fusion matrix;

[0070] S4. Adopt the dynamic game strategy to calculate the total benefits brought by forwarding the original topic and the derivative topic respectively, calculate the driving force of the original topic and the driving force of the derivative topic in combination with the fusion matrix, and integrate the two driving forces into the full user relationship network to obtain the latest full user relationship network;

[0071] S5. Construct a topic propagation prediction model using a two-layer graph convolutional neural network with an intermediate Dropout layer, input the latest full user relationship network matrix into the topic propagation prediction model, and predict the propagation trend of the target original topic at the next moment.

[0072] In one embodiment, the object of the present invention is to predict the user's forwarding behavior and the propagation trend of the topic at the next moment according to the user's basic information, past behaviors, and the game relationship of hot topics. There are many driving forces affecting user behavior, and the present invention analyzes from three aspects: user attributes, user relationships, and topic influence.

[0073] Specifically, the user attributes include user basic attributes and user sentiment index,

[0074] The user basic attributes are expressed as:

[0075] Attr(u i )=act(u i )×trans(u i )×cog(u i )

[0076] Where Attr(u i ) represents the participating user u iThe basic attributes of user u, act(u i ) represents the activity of user u participating i in it, trans(u i ) represents the historical forwarding rate of user u participating i in it, and cog(u i ) represents the awareness of user u participating i in it;

[0077] In this embodiment, the Internet access frequency of the user is used to represent the user activity. Logically speaking, the higher the Internet access frequency of a user, the greater the possibility that the user participates in the spread of rumors. The activity of user u participating i is defined as:

[0078]

[0079] tm(u i ) represents the number of Internet accesses of user u participating i in the past month, tm(u p ) represents the maximum value in the set of the number of Internet accesses of all participating users in the past three months, and tm(u q ) represents the minimum value in the set of the number of Internet accesses of all participating users in three months;

[0080] In this embodiment, the frequency of the user publishing and forwarding Weibo is used to represent the historical forwarding rate. Similarly, the more times a user publishes or forwards Weibo, the greater the possibility that the user participates in the spread of rumors. The historical forwarding rate of user u participating i is defined as:

[0081]

[0082] tr(u i ) represents the number of Weibo posts and forwards of user u participating i in the past three months, tr(u m ) represents the maximum value in the set of the number of Weibo posts and forwards of all participating users in the past three months, and tr(u n ) represents the minimum value in the set of the number of Weibo posts and forwards of all participating users in the past three months;

[0083] In this embodiment, the user's awareness of the topic is measured by whether the user has retrieved similar topics before and the number of retrievals. The awareness of user u participating i is defined as:

[0084]

[0085] Wr represents the keywords related to rumors in all the Weibo content of user u participating i in it, and Wh represents the Weibo content of user u participatingi The words with the highest frequency of occurrence in the Weibo content within the past three months; in this implementation, the TextRank algorithm is used to extract the keywords related to the topic from the Weibo posts and forwards of participating users, and then the Jaccard coefficient is used to measure the cognitive degree of each user.

[0086] At time t1, the user is affected by the relevant topic and contacts the topic at time t2 with an emotional tendency. This emotional tendency will affect the user's forwarding behavior, and the group emotional fluctuation affects the user's cognition.

[0087] Therefore, in this embodiment, the user emotion index is defined as:

[0088]

[0089] where Emo t (u i ) represents the user emotion index of participating user u i as of time t, Num[interact(u i ,t)] represents the number of participation interactions of participating user u i as of time t with the target original topic; Num[interact(u i ,t-1)] represents the number of participation interactions of participating user u i as of time t-1 with the target original topic, Num[bebavior(u i ,t)] represents the total number of behaviors of participating user u i as of time t, Num[bebavior(u i ,t-1)] represents the total number of behaviors of participating user u i as of time t-1. The total number of behaviors refers to the number of participation interactions of participating user u i with multiple topics including the target original topic. The participation interaction mainly refers to the forwarding situation of the participating user to the topic. α represents the influence factor, α∈[0,1]. Theoretically, the larger α is, the higher the current user's emotion is and the higher the user's tendency to participate in the interaction is.

[0090] Specifically, after extracting the relevant attributes from the data, a user information matrix is constructed. Among them, the information of each user node info(u i ) is expressed as:

[0091] info(u i )=Attr(u i )×Emo t (u i )

[0092] In one embodiment, according to the obtained data source, a full user relationship network of participating users and potential users is constructed. At time t, the target native topic spreads in the full user relationship network, and the full user relationship network at this time is expressed as:

[0093]

[0094] where U pa represents the set of participating users of the target native topic at time t, and U po represents the set of potential users of the target native topic at time t. represents the set of social relationships between participating users and potential users;

[0095] In the full user relationship network, calculate the influence of the neighbor users of each participating user u i on the propagation of neighbor messages generated by it, which is expressed as:

[0096]

[0097] Num(U opt ) represents the number of optimistic neighbor users of the participating user u i , and Num(U pes ) represents the number of pessimistic neighbor users of the participating user u i . In this embodiment, it is considered that optimistic neighbor users have a greater impact on participating users. Therefore, a decay coefficient β is given to pessimistic neighbor users. At the same time, in this embodiment, the user optimism index Opt(u i ) of the user u i is set. If , then the user u i is an optimistic user, otherwise it is a pessimistic user.

[0098] In order to make user information more complete, the user information matrix and the user relationship matrix are fused to obtain a fusion matrix. The node information in the fusion matrix is expressed as:

[0099] fus(u i ) = info(u i ) + rela(u i )

[0100] where rela(u i ) represents the node information of the participating user u i in the user relationship matrix.

[0101] In one embodiment, whether a user will participate in topic forwarding depends on whether the user can obtain high returns. This embodiment generalizes the process in which a user selects the best game strategy that can obtain the maximum return to the process of the spread of the original topic and the derivative topic in the social network, so as to measure the influence between topics.

[0102] This embodiment defines a dynamic game strategy. Assuming that a user only participates in the forwarding of one topic and only forwards it once, the behavior of the user forwarding the topic can be regarded as the best decision after the game among the current user node itself, all upstream nodes (neighbor nodes of the current user node), and the current user node. The game refers to the user's historical behavior data. Specifically, assuming that C represents the target node, B represents the upstream node of the target node, that is, the neighbor node, x and y represent the forwarding strategies, and x = 1, 2, y = 1, 2, 1 and 2 respectively represent forwarding the original topic and forwarding the derivative topic related to the original topic. BC xy represents the return when the target node selects the x forwarding strategy and its neighbor node selects the y forwarding strategy; CC xy represents the return when the target node selects x in the previous participating topic of the original topic and y in the original topic.

[0103] In this embodiment, the dynamic game strategy includes two parts. The first part is the game between the participating user and its neighbor user, as shown in Table 1,

[0104] Table 1 Payoff matrix of the dynamic game with neighbor users

[0105]

[0106] Combined with the payoff matrix, the payoff of the user choosing each strategy can be quantified. Taking a user with degree K in the online social network as an example, the payoff functions obtained by choosing Strategy 1 (forwarding the original topic) and Strategy 2 (forwarding the derivative topic) can be expressed by the following formulas respectively:

[0107]

[0108]

[0109] pro1(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 1, that is, forwarding the target original topic, in the first part of the game; pro2(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 2, that is, forwarding the derivative topic related to the target original topic, in the first part of the game; M xy represents for the target original topic, the participating user u i executes the strategy x = 1, 2, and at the same time the participating user ui The specific payoffs when the neighboring users of 11 represent the participating user u i executes Strategy 1, and at the same time, the neighboring users of the participating user u i execute Strategy 1. M 12 represent the participating user u i executes Strategy 1, and at the same time, the neighboring users of the participating user u i execute Strategy 2. M 21 represent the participating user u i executes Strategy 2, and at the same time, the neighboring users of the participating user u i execute Strategy 1. M 22 represent the participating user u i executes Strategy 2, and at the same time, the neighboring users of the participating user u i execute Strategy 2. K represents the number of neighboring users of the participating user u i , represents the number of neighboring users of the participating user u i who forward the native topic of the forwarding target at time t - 1; represents the number of neighboring users of the participating user u i who forward the derivative topic related to the target native topic at time t - 1, and

[0110] The second part is the game between the participating user and himself, as shown in Table 2.

[0111] Table 2 Payoff matrix of the dynamic game with oneself

[0112]

[0113] In addition to playing games with upstream nodes, usually, users also have game behaviors with themselves. To quantify their own game behaviors, taking the total number N of all microblogs of the user in the past three months as an example, the payoff functions obtained by choosing Strategy 1 and Strategy 2 can be expressed by the following formulas respectively:

[0114]

[0115] [[ID=--]]

[0116] pro′1(u i ) represents the payoff function obtained by the participating user u i through choosing Strategy 1, that is, forwarding the target native topic, after the game in the second part; pro′2(u i ) represents the payoff function obtained by the participating user u iAfter the game in the second part, select strategy 2, that is, the profit function obtained by forwarding the derivative topic related to the target native topic; m xy Denote the participating user u i The specific profit when executing strategy x = 1, 2 in the previous participated topic of the target native topic and executing strategy y = 1, 2 for the target native topic; specifically, m 11 Denote the participating user u i The specific profit when executing strategy 1 in the previous participated topic of the target native topic and executing strategy 1 for the target native topic, m 12 Denote the participating user u i The specific profit when executing strategy 1 in the previous participated topic of the target native topic and executing strategy 2 for the target native topic, m 21 Denote the participating user u i The specific profit when executing strategy 2 in the previous participated topic of the target native topic and executing strategy 1 for the target native topic, m 22 Denote the participating user u i The specific profit when executing strategy 2 in the previous participated topic of the target native topic and executing strategy 2 for the target native topic, N denotes the participating user u i The number of Weibo posts within the past three months, Denote that at time t - 1, the participating user u i The number of times of forwarding the target native topic, Denote that at time t - 1, the participating user u i The number of times of forwarding the derivative topic related to the target native topic.

[0117] Calculate the total profit of each participating user when choosing strategy 1 and strategy 2 respectively according to the dynamic game strategy, denoted as:

[0118] pro nat (u i ) = pro1(u i ) × pro′1(u=43]] i )

[0119] pro der (u i ) = pro2(u i ) × pro′2(u i )

[0120] Among them, pro nat (u i ) denotes the total profit function obtained by the participating user u i when choosing strategy 1, that is, forwarding the target native topic, pro der (u i ) denotes the participating user u iSelect Strategy 2, that is, the total revenue function obtained by forwarding derivative topics related to the target native topic. Calculate the two driving forces of each participating user according to the above formula, and the formula is:

[0121] D0(u i )=pro nat (u i )×fus(u i )

[0122] D1(u i )=pro der (u i )×fus(u i )

[0123] Among them, D0(u i ) represents the native topic driving force of participating user u i , D1(u i ) represents the derivative topic driving force of participating user u i . The driving force can be understood as the probability of a user participating in a certain topic.

[0124] Generally speaking, the destination of the birth of each topic is destruction, that is: the popularity of the topic will decrease over time. Therefore, the present invention uses a time weight coefficient based on topic popularity to quantify the decay process of the topic. The time decay function f topic (t) is defined as:

[0125] f topic (t)=D×e -(1+θt)

[0126] Among them, D represents the initial popularity of the target topic, which is represented by the number of topic user attentions in the first three months, and θ is the decay parameter. Incorporate the time decay function into the two driving forces.

[0127] In one embodiment, the two driving forces integrated with the time decay function are respectively incorporated into the edge weights of the full user relationship network. The edge weights are proportional to the driving forces of the user nodes. If the behavioral driving force between the current user node and the upstream user node is greater, the corresponding edge weight value is greater. A topic propagation prediction model is constructed by using a two-layer graph convolutional neural network with an intermediate Dropout layer added, and the topic propagation prediction model is used to predict the propagation trend of the current original topic at the next moment. Among them, the first layer of the graph convolutional neural network is expressed as:

[0128]

[0129] H (1) represents the output of the first layer of the graph convolutional neural network, G represents the user relationship matrix, represents the degree matrix transformation of G, X represents the user information matrix, W(0) represents the weight matrix of the first-layer graph convolutional neural network, and ReLu represents the Relu activation function;

[0130] The second-layer graph convolutional neural network is represented as:

[0131]

[0132] Combining the two-layer representations, it can be expressed as:

[0133]

[0134] This paper discusses a three-classification problem, so the model output is represented by Z = P(o, d, n|u i ). The larger the Z value of different classes indicates that the user is more likely to forward this label. o represents the original topic label of the forwarding target, d represents the derivative topic label of the forwarding, and n is the non-forwarding label. Specifically as follows:

[0135]

[0136] Among them, when Y = 0, the user forwards the original topic of the target, when Y = 1, the user forwards the derivative topic, and when Y = 2, the user does not forward.

[0137] The present invention can effectively manage topics, release rumor-refuting information, prevent the spread of rumors, and contribute to public opinion management by predicting the propagation trend of topics at the next moment.

[0138] In the present invention, unless otherwise clearly specified and limited, terms such as "installation", "setting", "connection", "fixation", "rotation", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components. Unless otherwise clearly limited, for those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0139] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirits of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An information dissemination prediction method based on user information and topic game relationship, characterized in that It includes the following steps: S1. Download data sources from existing Web-based research recommendation systems or obtain them from the public APIs of social platforms; S2. According to the obtained data sources, extract the user attributes of participating users. The user attributes include user basic attributes and user sentiment indices, and establish a user information matrix based on the user attributes; S3. According to the obtained data sources, construct a full-user relationship network of participating users and potential users, and calculate the neighbor message propagation influence based on the full-user relationship network to obtain a user relationship matrix; Fuse the user information matrix and the user relationship matrix to obtain a fusion matrix; S4. Adopt a dynamic game strategy to calculate the total benefits brought by forwarding the original topic and the derivative topic respectively, combine the fusion matrix to calculate the driving force of the original topic and the driving force of the derivative topic, and integrate the two driving forces into the full-user relationship network to obtain the latest full-user relationship network matrix; Characterize the process of users choosing the best game strategy as the process of the original topic and the derivative topic spreading in the social network, and design a dynamic game strategy with the goal of participating users obtaining high benefits after forwarding the topic. The dynamic game strategy includes two parts. The first part is the game between participating users and their neighbor users, which is expressed as: pro1(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 1, i.e., forwarding the target native topic, in the game of the first part; pro2(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 2, i.e., forwarding the derivative topic related to the target native topic, in the game of the first part; M xy represents the specific payoff when the participating user u i executes Strategy x = 1, 2, and at the same time, the neighboring users of the participating user u i execute Strategy y = 1, 2 for the target native topic; K represents the number of neighboring users of the participating user u i ; represents the number of neighboring users who forward the target native topic of the participating user u i at the moment of t - 1; represents the number of neighboring users who forward the derivative topic of the participating user u i at the moment of t - 1, and The second part is the game between participating users and themselves, which is expressed as: pro′1(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 1 in the second - part game, that is, forwarding the target original topic; pro′2(u i ) represents the payoff function obtained by the participating user u i after choosing Strategy 2 in the second - part game, that is, forwarding the derivative topic related to the target original topic; m xy represents the specific payoff when the participating user u i executes Strategy x = 1, 2 in the previous participated topic of the target original topic and executes Strategy y = 1, 2 for the target original topic; N represents the number of Weibo posts of the participating user u i in the past three months, represents the number of times the participating user u i forwarded the target original topic at time t - 1, represents the number of times the participating user u i forwarded the derivative topic at time t - 1; Calculate the total benefits of each participating user choosing Strategy 1 and Strategy 2 respectively according to the dynamic game strategy, which is expressed as: pro nat (u i ) = pro1(u i ) × pro′1(u i ) pro der (u i ) = pro2(u i ) × pro′2(u i ) Among them, pro nat (u i ) represents the total revenue function obtained by participating user u i selecting Strategy 1, that is, forwarding the target native topic; pro der (u i ) represents the total revenue function obtained by participating user u i selecting Strategy 2, that is, forwarding the derivative topic related to the target native topic; S5. Construct a topic propagation prediction model using a two-layer graph convolutional neural network with an intermediate Dropout layer, input the latest full-user relationship network matrix into the topic propagation prediction model, and predict the propagation trend of the target original topic at the next moment.

2. The information dissemination prediction method based on the user information and topic game relationship according to claim 1, characterized in that, The obtained data sources include the topic participation records of the target original topic, the forwarding situation of messages, and the behavior records of participating users; the topic participation records include the basic information, attention information, and being-followed information of participating users, as well as the topic forwarding time; The behavior records of participating users include the historical forwarding information and historical comment information of participating users.

3. A method for predicting information dissemination based on user information and topic game relationship according to claim 1, characterized in that, User attributes include user basic attributes and user sentiment indices. The user basic attributes are expressed as: Attr(u i ) = act(u i ) × trans(u i ) × cog(u i ) Among them, Attr(u i ) represents the basic user attributes of user u i , act(u i ) represents the activity of user u i , trans(u i ) represents the historical forwarding rate of user u i , and cog(u i ) represents the awareness of user u i ; The user sentiment indices are expressed as: Among them, Emo t (u i ) represents the user emotion index of participating user u i up to time t, Num[interact(u i ,t)] represents the number of participation interactions of participating user u i up to time t with the target native topic; Num[interact(u i ,t-1)] represents the number of participation interactions of participating user u i up to time t-1 with the target native topic, Num[bebavior(u i ,t)] represents the total number of behaviors of participating user u i up to time t, Num[bebavior(u i ,t-1)] represents the total number of behaviors of participating user u i up to time t-1, α represents the influence factor, α ∈ [0, 1].

4. The information dissemination prediction method based on the user information and topic game relationship according to claim 3, wherein Participating user u i The activity level is defined as: tm(u i ) represents the number of times user u i has accessed the Internet in the past month. tm(u p ) represents the maximum value in the set of the number of times all participating users have accessed the Internet in the past three months. tm(u q ) represents the minimum value in the set of the number of times all participating users have accessed the Internet in three months; Participating user u i The historical forwarding rate is defined as: tr(u i ) represents the number of microblogs posted and reposted by user u i within the past three months. tr(u m ) represents the maximum value in the set of the number of microblogs posted and reposted by all participating users within the past three months. tr(u n ) represents the minimum value in the set of the number of microblogs posted and reposted by all participating users within the past three months; Participating user u i is defined as follows: Wr represents the keywords related to rumors in all the Weibo content participated by user u i and Wh represents the word with the highest frequency of occurrence in the Weibo content of user u i in the past three months.

5. The information dissemination prediction method based on the user information and topic game relationship according to claim 1, characterized in that Construct the full-user relationship network of participating users and potential users of the target native topic at time t It is expressed as: Among them, U pa represents the set of participating users of the target native topic at time t, and U po represents the set of potential users of the target native topic at time t, represents the set of social relationships between participating users and potential users; Calculate, for each participating user u in the full user relationship network i the influence of neighbor messages generated by its neighbor users, expressed as: Num(U opt ) represents the number of optimistic neighbor users participating in user u i , Num(U pes ) represents the number of pessimistic neighbor users participating in user u i , and β represents the attenuation coefficient.

6. The information dissemination prediction method based on the user information and topic game relationship according to claim 1, characterized in that The topic propagation prediction model is expressed as: Among them, X represents the user information matrix, G represents the user relationship matrix, and W (0) represents the weight matrix of the first-layer graph convolutional neural network, and W (1) represents the weight matrix of the second-layer graph convolutional neural network, represents the degree matrix transformation of G, that is H (1) represents the output of the first-layer graph convolutional neural network, and Z represents the output result of the topic propagation prediction model.

Images