Network Rumor Propagation Prediction Method Based on Users' Short-Term Emotions and Evolutionary Game

Abstract

The present invention proposes a method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games, belonging to the field of data processing. The method includes obtaining users' basic information, user behavior data, and user text data on a social network platform and performing preprocessing; calculating users' own attributes, user influence, user topic participation, message popularity, friend driving force, and message sentiment scores; calculating rumor influence and rumor refutation influence, and obtaining mutual influence through the method of evolutionary game; using a node embedding algorithm to map user nodes into a vector space, and using the CSR2Vec algorithm based on mutual influence to generate a feature topology matrix; splicing the feature topology matrix and the user adjacency matrix to obtain a feature vector matrix; inputting the feature vector matrix into a graph attention network model with an attention mechanism to output the propagation prediction results of users; the present invention can be applied to multiple fields such as rumor control and green network security.

Description

Technical Field

[0001] The present invention belongs to the field of data processing of social network platforms, and mainly relates to user influence discovery, user behavior analysis, especially a method for predicting the spread of online rumors based on short-term user emotions and evolutionary games. Background Art

[0002] Rumor is a complex social phenomenon, which has two very obvious characteristics: importance and ambiguity. Rumors are usually some unconfirmed information that is spread around by most people. In real life, people cannot verify the truth of the information by themselves, and often change their inner views with the people around them, resulting in herd mentality. Many rumors are spread on a large scale and are considered to be true by people. When rumors spread widely in our society, they often cause panic in people's hearts and create a situation of social instability. This series of impacts is harmful to the construction of a stable social environment in our country.

[0003] With the continuous update and iteration of science and technology, people's communication methods have undergone earth-shaking changes. From the initial offline communication and chatting to the current widely used online communication platforms, the development of science and technology is gradually changing people's living habits. Nowadays, hundreds of millions of Internet users will use devices such as mobile phones, tablets, and computers to post comments online. The convenience and applicability of the devices have greatly increased the time for Internet users to browse information online. Without leaving home, Internet users can know the latest hot events. The interactivity of online chat platforms allows users to get the most real experience. With the rapid development of mobile communication technology, the construction of network platforms has also reached a higher level, which can carry higher loads and provide users with faster update speeds, bringing a better sense of experience to users.

[0004] However, things always have two sides. For example, large online network communication platforms such as Weibo bring convenience to users for sharing news, but also bring some potential dangers to society. Due to the huge number of users on Weibo, it is more difficult to monitor the network platform. When a hot event breaks out, a large number of users often join the discussion of the topic. Due to the randomness of users' news dissemination, we often receive some news that does not conform to the facts, which is what we often call rumors. In the topic spread, there are often a large number of rumor messages, and rumor messages spread more quickly and wider than normal messages.

[0005] In recent years, researchers at home and abroad have conducted in-depth research on the field of rumor propagation. In terms of research models, the earliest ones were based on social network analysis and mathematical models, and the research on rumor propagation models was carried out based on user nodes. Among them, the infectious disease (SIR) model is the most basic model. On the other hand, research is carried out from the perspective of the characteristics of rumor propagation by combining machine learning and deep learning. In terms of research perspectives, some research scholars mainly conduct research on rumors from the aspects of users' own characteristics. Starting from the users themselves, considering the users' historical propagation behaviors and their own characteristics, to predict the attractiveness of rumors to users. Another part of the research scholars predict the forwarding behavior of users by rumors based on the propagation space of rumors. These researchers have achieved good results.

[0006] Although researchers have achieved certain results in the prediction of the spread of rumor topics, there are still some challenges:

[0007] 1. The emotional relevance between users and rumor messages. In the message propagation space, within a certain period of time, the emotion of rumor messages will largely affect users' behaviors. How to quantify the emotional influence between users and messages is a problem that needs to be solved.

[0008] 2. The complexity and diversity of the rumor message feature space. In the rumor propagation space, the complexity and diversity of users, messages, networks, and behaviors bring difficulties to effectively expressing rumor features, thus affecting the prediction accuracy of the model.

[0009] 3. The cooperation and opposition of rumor messages. In the rumor propagation space, rumor messages and rumor refutation messages exist simultaneously. Users' behaviors are greatly affected by these two types of messages. How to measure the influence of rumor messages and rumor refutation messages on users' behaviors is a problem that researchers need to consider. Summary of the Invention

[0010] In response to the above challenges, the present invention proposes a method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games. The present invention introduces evolutionary game theory to explore the factors affecting the spread of rumor messages at the individual level and the group level. Finally, a behavior prediction model of users' responses to rumor messages is established, and the group behaviors of participants in topic propagation are analyzed within different time periods to study the evolutionary law of rumors during their life cycles.

[0011] A method for detecting the spread of online rumors based on users' short-term emotions and evolutionary games, the method comprising:

[0012] Obtain the basic information, behavior data, and text data of users on a social network platform, and preprocess the obtained data;

[0013] Based on the preprocessed basic user information and user behavior data, calculate the user's own attributes, user influence, user topic participation, message popularity, and friend driving force respectively;

[0014] Based on the preprocessed user text data, generate word frequency features using the TF-IDF algorithm, and process the word frequency features using a sentiment classifier to generate message sentiment scores;

[0015] Based on the user's own attributes, user influence, user topic participation, message popularity, message sentiment scores, and friend driving force, calculate the rumor influence and refutation influence using the multiple regression linear algorithm;

[0016] Calculate the strategic benefits of users forwarding rumor information and forwarding refutation information respectively based on the rumor influence and refutation influence, and use the evolutionary game model to measure the mutual influence of rumor information and refutation information;

[0017] Use the node embedding algorithm to map user nodes to the vector space, and use the random walk algorithm based on mutual influence to generate the characteristic topology matrix of the topic network;

[0018] Concatenate the characteristic topology matrix of the topic network and the user adjacency matrix, and fuse them to form a characteristic vector matrix;

[0019] Input the fused characteristic vector matrix into the graph attention network model with an attention mechanism, and output the prediction results of users not participating in rumor topics, or forwarding rumor information or forwarding refutation information. The characteristic topology matrix of the topic network The characteristic topology matrix of the topic network

[0020] Advantages of the present invention:

[0021] The present invention collects records such as the user's personal information, user historical behavior records, and user text content, and creates a user interaction matrix, that is, an adjacency matrix, which can effectively express the interaction relationship between users. By designing a new method CSR2Vec algorithm (random walk algorithm based on mutual influence) for learning topics, the feature space is vectorized with low rank and density, effectively solving the problem of user data sparsity. At the same time, the present invention combines user emotional factors and evolutionary game theory, fully considering the emotional factors of users when facing rumors, and improving the accuracy of predicting whether users forward rumors. Finally, using a graph neural network with an attention mechanism, it can better process dynamic graph structures, and through the results of user behavior, it can effectively predict group behavior and further judge the development trend of rumors. Brief Description of the Drawings

[0022] Figure 1 It is the structural diagram of the network rumor propagation model based on users' short-term emotions and evolutionary games in the embodiments of the present invention;

[0023] Figure 2 is the flowchart of the network rumor propagation method based on users' short-term emotions and evolutionary game in the embodiments of the present invention;

[0024] Figure 3 is the schematic diagram of text representation in the embodiments of the present invention;

[0025] Figure 4 is the schematic diagram of the representation of the feature topology matrix structure in the embodiments of the present invention. Specific embodiments

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] To better elaborate the technical solutions of the present invention and make the advantages more concise and clear, the problems to be solved by the present invention will be specifically explained first, and then with reference to the accompanying drawings of the specification, the specific embodiments of the present invention will be further described in detail.

[0028] Figure 1 is the structural diagram of the network rumor propagation model based on users' short-term emotions and evolutionary game in the embodiments of the present invention; As Figure 1 shown, it shows that the input of the present invention is the basic information of users, user behavior data, and user text data of the social network platform. After evolutionary game processing, the mutual influence between rumor information and rumor refutation information is obtained. The Content Struct rumor Vec (CSR2Vec) algorithm based on mutual influence is used to generate the feature topology matrix of the topic network. After fusing the feature topology matrix and the user adjacency matrix, it is input into the GAT graph attention network, and finally a multi-classification result is output, that is, the prediction result that the user does not participate in the rumor topic, or forwards the rumor information or forwards the rumor refutation information is output.

[0029] Figure 2 is the flowchart of the network rumor propagation method based on users' short-term emotions and evolutionary game in the embodiments of the present invention; As Figure 2 shown, the method includes:

[0030] 101. Obtain the basic information of users, user behavior data, and user text data of the social network platform, and preprocess the obtained data;

[0031] In the embodiments of the present invention, the way to obtain the data of the social network platform can be to download it from the official website, or to obtain it using the mature social network public API.

[0032] The data obtained in the embodiments of the present invention can be the basic information of users participating in rumor-refuting topics in their life cycles from Twitter datasets and Sina Weibo datasets, as well as their historical behavior records and topic participation records. The user basic information is structured data recording the user's age, gender, region, etc. The topic participation record requires the time when the topic is retweeted and commented, and the historical behavior record is the time and content when the user likes, comments on, and retweets other users' messages within a certain period of time. These information are reorganized and divided into user basic information, user behavior data, and user text data.

[0033] It can be understood that the user text data is further extracted according to the topic participation record and user behavior. The original data is the record of the user liking other users and the corresponding content liked.

[0034] In the embodiments of the present invention, the original data obtained from the social network platform are all unstructured data, and some of the recorded data are missing and cannot be directly used for data analysis. Therefore, the embodiments of the present invention can structure the unstructured data through simple data preprocessing (clearing some invalid data and deleting some duplicate data).

[0035] In the embodiments of the present invention, according to the recorded information such as the collected user basic information, user behavior data, and user text data, a user interaction matrix, that is, a user adjacency matrix, is created. The user adjacency matrix can be constructed using a graph node network, that is, taking users as nodes and the relationships between users as edges, and using the graph network to determine the adjacent users of the users. In this way, the interaction relationship between users can be effectively expressed.

[0036] 102. According to the preprocessed user basic information and user behavior data, calculate the user's own attributes, user influence, user topic participation, message popularity, and friend driving force respectively;

[0037] In the embodiments of the present invention, in depicting the propagation process of rumor-refuting messages, in fact, relevant features need to be extracted from the two perspectives of messages and users and models are constructed using these features; therefore, these attribute information are divided into user attribute information and message attribute information. On the user side: such as the user's influence, activity, and their own interests; on the message side: the heat of the message, the emotional tendency of the message, etc. Different influencing factors are considered from two levels.

[0038] At the user level, this embodiment mainly considers the user's own attributes, user influence, and user topic participation. The following will analyze these three attributes respectively:

[0039] This embodiment is divided into a user set and a user participant network set Among them, U t represents the set of users who participate in the spread of hot topics during time period t, and R t represents the set of rumor participants during time period t, and A t represents the set of rumor refutation participants. Therefore, the set of rumor participants and the set of rumor refutation participants together constitute the user set. represents the set of edges of users U t who participate in the spread of rumor topics during time period t. The spread of rumor topics here is the rumor-refutation topic, which specifically includes spreading rumor messages or / and spreading rumor refutation messages.

[0040] The user's own attributes are represented as: User(u i ) = [age(u i ), gender(u i ), fans(u i ), fris(u i )](1)

[0041] Among them, age(u i ) represents the age of the user, gender(u i ) represents the age of the user, fans(u i ) represents the number of fans of the user, and fris(u i ) represents the number of friends of the user.

[0042] The user influence is represented as:

[0043]

[0044] Among them, userInfluence(u i ) represents the influence of user u i , user u i represents the user who participates in the topic, represents the average number of forwards of the content published by user u i , represents the average number of likes of the content published by user u i .

[0045] The user topic participation is represented as:

[0046]

[0047] Among them, participate(u i ) represents the topic participation of user u i , allNum(u i) represents user u i All the number of comments, forwards, and likes, transNum(u i ) represents user u i All the number of comments, forwards, and likes of user u regarding topic I.

[0048] At the message level, this embodiment mainly considers message popularity, friend driving force, and message sentiment score. The following will analyze these three attributes separately:

[0049] The message popularity is expressed as:

[0050]

[0051] Among them, Popularity(t) represents the popularity of the message at time t. Here, the message can include rumor messages or / and rumor refutation messages. t0 represents the moment when the rumor occurs. Num(t) and Num(t - 1) respectively represent the number of forwards and likes of the rumor information up to the current moment and the previous moment.

[0052] It can be understood that the message popularity here refers to the whole of a message. Both rumor messages and rumor refutation messages have a popularity, similar to a general term. When calculating the mutual influence, the present invention divides the messages into rumor messages and rumor refutation messages, and processes the relevant content of the two types of information respectively.

[0053] The friend driving force is expressed as:

[0054]

[0055] Among them, represents the friend driving force of user u i , n is the number of users in the rumor topic, represents the friend driving force of user u j towards user u i ; The specific formula is as follows:

[0056]

[0057] Among them, represents the average number of times user u i forwards the Weibo of user u j .

[0058] When a user forwards a hot topic on the Internet, they often attach a large amount of text with their own opinions, which can reflect the user's attitude towards the current topic. At the same time, most of the tweets on social networks carry strong emotional information; therefore, the present invention also considers the attribute of message sentiment score, which is the sentiment tendency of each of the six emotions contained in the given message text, including joy, sadness, surprise, anger, disgust, and fear. It can more clearly represent the positive and negative meanings of the entire text in each emotion category.

[0059] The message sentiment score is expressed as:

[0060] SenCat(T) = f(T) (7)

[0061] where, for the given text T i and the sentiment classifier f(·). f(T) is the category to which the given text T belongs.

[0062] However, how to select the given text is a problem to be solved by the prior art, and the present invention will describe the specific calculation process of the message sentiment score in step 104.

[0063] 103. According to the preprocessed user text data, use the TF-IDF algorithm to generate word frequency features, and use the sentiment classifier to process the word frequency features to generate a message sentiment score;

[0064] In the embodiments of the present invention, it is found that false rumors trigger negative emotions such as fear, disgust, and surprise in their responses, while true statements tend to be positive emotions such as joy, sadness, trust, and expectation. If the emotional state of a user within a certain period is close to the emotional tendency contained in the hot topic they are facing, then the user has a high probability of choosing to forward the speech related to this topic. And the emotions of users are dynamic. It is very crucial to detect the emotions of users at a finer time granularity to judge the forwarding behavior of users. On social network platforms, users are more inclined to forward other people's microblogs or comments than to post a microblog by themselves. Moreover, both microblogs and comments are short texts with short lengths, and a certain hot word in them will have a greater impact on the emotion of the entire short text. Compared with the short text research method based on word vectors, TF-IDF can better reduce the proportion of hot words in the whole sentence and document. The present invention defines the TF-IDF value of the short text as:

[0065]

[0066] Among them, TF represents the term frequency, that is, the number of times a word appears in a document, which may be positively correlated with the length of the document. Therefore, it is necessary to normalize the term frequency, usually by dividing the number of times it appears by the total number of words in the document:

[0067]

[0068] In formula (9), the numerator n i,j represents the frequency of word f i in document j, and the denominator ∑ k n k,j represents the sum of all word frequencies in document j.

[0069] Since the discrimination of the traditional TF value calculation is not obvious enough after the term frequency of a word exceeds a certain threshold, the present invention standardizes according to the maximum TF value of the document, as shown in formula (10):

[0070]

[0071] Among them, K is an adjustable value between 0.5 and 1. There is a problem with the traditional TF value calculation. If a word appears 300 times in document A and 150 times in document B, it does not mean that the relevance in A is twice that in B. That is, after the term frequency of a word exceeds a certain threshold, the discrimination of TF is not so obvious and it is not linearly correlated, and there is a problem of long and short documents. Therefore, the present invention avoids the linearization of relevance through a logarithmic function and standardizes it at the same time; on the other hand, it ensures that the discrimination of TF will not be too obvious. Since standardizing the maximum TF value each time may cause unnecessary operations, that is, it may cause the calculated TF value to deviate from the actual TF value after being too obvious. Based on this, the present invention sets that when the term frequency exceeds the threshold, formula (10) is used to calculate the TF value, and when the term frequency does not exceed the threshold, formula (9) is used for calculation.

[0072] In a preferred embodiment of the present invention, the calculation process of the threshold may include selecting several documents / sentences that are the closest to the current document / sentence as the evaluation documents / sentences, that is, the documents / sentences sent by the same user in the recent period, or the documents / sentences sent by the neighbor users of the user in the recent period; taking the TF mean value of each document / sentence in the current several documents / sentences as the initial value of the corresponding document / sentence, minimizing the cost function, so as to determine the correlation coefficient corresponding to the document / sentence; under this correlation coefficient, by changing the initial value, the calculated value under the minimized cost function is determined; the sum of the products of the correlation coefficient of each document / sentence and the calculated value under the minimized cost function of the corresponding document / sentence is used as the threshold of the current document / sentence.

[0073] Among them, the cost function can adopt any existing cost function, aiming to minimize the gap between the initial value and the TF value of the actual document / sentence.

[0074] IDF represents the inverse document frequency, which can measure whether a vocabulary can fully express the text. The smaller the IDF value, the fewer texts contain this vocabulary, and then this vocabulary can represent this text. For example, the word 'de' in a sentence has a very high frequency of occurrence in the document, but this word does not play a role in expressing the sentiment of the sentence.

[0075]

[0076] Active users often forward, like, and comment on the Weibo of other users on social network platforms. The present invention summarizes the texts associated with users, performs time slicing on the text set, so as to provide the sentiment of a user within a certain time segment.

[0077]

[0078] Among them, represents the set of Weibo texts forwarded by user u within time period t, represents the set of Weibo liked by user u within time period t, then represents the comments of user u within time period t.

[0079] Such as Figure 3 As shown, after processing the text with TF and IDF, a vector representation can finally be obtained, that is, calculate the TF-IDF value of each text in the corpus according to formula (8) TF pq =fw p1 ,fw p2 ,fw p3 ,...,fw pq ), where p represents the text number, q represents the qth word in the text, and fw pq represents the TF-IDF value of its word. Finally, the text will be represented as the frequency value of the word frequency.

[0080] On social networks, when users face a hot topic, sometimes they will spontaneously forward corresponding messages according to their own preferences. Sometimes, they will also forward rumor messages or refutation messages under the influence of other users on the network. According to the above content, the present invention summarizes the factors affecting users' forwarding behavior into six: user's own attributes, user influence, user topic participation, message popularity, message sentiment score, and friend driving force. Based on step 102 and step 103 to quantify the different influence intensities of these factors, it is possible to identify the key factors affecting users' forwarding behavior.

[0081] 104. Calculate the rumor influence and the rumor refutation influence by using the multiple regression linear algorithm based on the user's own attributes, user influence, user topic participation, message popularity, message sentiment score, and friend driving force;

[0082] In the embodiment of the present invention, considering that the rumor influence and the rumor refutation influence are mainly composed of the user factor userfactor(u i ) and the message factor mesfactors(I), the corresponding rumor influence and rumor refutation influence can be obtained through the various factors calculated in steps 102 and 103. Among them:

[0083] The composition of the user factor includes: the user's own attributes, user influence, and topic participation, that is:

[0084] userfactor(u i ) = User(u i ) * userInfluence(u i ) * participate(u i ) (13)

[0085] Among them, represents the user factor of user u i .

[0086] The composition of the message factor includes: message popularity, message sentiment score, and friend driving force, that is:

[0087]

[0088] Among them, mesfactors(u i , u j the message factor between user u i and user u j , mesfactors rumor (u i , u j ) represents the message factor of the rumor message spread between user u i and user u j , mesfactors anti-rumor (u i , u j ) represents the message factor of the rumor refutation message spread between user u i and user u j .

[0089] Combining the above two factors, use the multiple linear regression algorithm to construct the rumor and rumor refutation influence functions as:

[0090]

[0091]

[0092] Among them, are the partial regression coefficients obtained by training with the multiple linear regression algorithm respectively, represents the overall influence coefficient during the training process, represents the influence coefficient of user factors on user behavior during the training process, represents the influence coefficient of message factors on user behavior during the training process.

[0093] 105. Calculate the strategic benefits of users' forwarding rumor information and forwarding refuting rumor information respectively according to the rumor influence and the refuting rumor influence, and use the evolutionary game model to measure the mutual influence between rumor information and refuting rumor information;

[0094] In the embodiments of the present invention, due to the complexity of the social network, while rumors are spreading, there are often some refuting rumor messages attached, such as official refuting rumor announcement information. They have a relationship of mutual promotion and confrontation, and this relationship is also an important factor affecting users' forwarding behavior. Therefore, the present invention uses evolutionary game theory to quantify the rumor-refuting rumor mutual influence. First, according to the knowledge of evolutionary game theory, the present invention defines two game strategies: "forwarding rumor messages" and "forwarding refuting rumor messages". The benefit functions of the two strategies are respectively:

[0095] Ben rumor (u i ,u j ) = P1 × influence rumor (u i ,u j ) (17)

[0096] Ben anti-rumor (u i ,u j ) = P2 × influence anti-rumor (u i ,u j ) (18)

[0097] Among them, P1 and P2 are the proportions of users u i 's friends and fans who spread rumor messages and refuting rumor messages. Among them, the user nodes that do not participate in forwarding messages do not affect other user nodes, so they are ignored. Therefore, P1 + P2 = 1. Then, use evolutionary game theory to measure the rumor mutual influence:

[0098]

[0099]

[0100] Among them, MutualInf rumor (u i , u j ), MutualInf anti-rumor (u i , u j ) respectively represent the influence of the user u j spreading rumor messages and rumor-refuting messages on the user u i 's spreading behavior.

[0101] Finally, according to the evolutionary game theory, there will be competition between rumor messages and rumor-refuting messages, and finally the rumor-refuting influence characteristic matrix is obtained:

[0102]

[0103] Among them, m(u i , u j ) = MutualInf rimor (u i , u j ) - MutualInfrumor(u i , u j ). If i = j, then m(u i , u j ) = 0.

[0104] 106. The node embedding algorithm is used to map user nodes into the vector space, and the random walk algorithm based on mutual influence is used to generate the characteristic topology matrix of the topic network;

[0105] In the example of the present invention, in the topic propagation space, the user's own attributes and the relationship network between users will affect the user's forwarding behavior. Based on the advantage of Node2Vec in expressing the network topology structure, the present invention represents the rumor network and its attributes. As Figure 4 shown, after collecting the user's basic information, user behavior data, and user text data, the user relationship and user attributes can be obtained. Users with the same attributes are divided into a user group, and the forwarding relationship of users is determined using the user's own attributes and user relationships.

[0106] The topic network at time T is represented as The purpose of the algorithm is to learn the characteristic vector f(w) of the user nodes in the social network under the hot topic. The input of the algorithm is the network topology with increased weights, and the output is the vector representation of the nodes. Therefore, the objective function is defined as:

[0107] Max∑ w logPr(N(w)|f(w)) (22)

[0108] Among them, N(w) represents the neighborhood nodes of user node w, and Pr(N(w)|f(w)) represents the probability of the neighborhood nodes appearing for the user node vector f(w). It can be obtained from formula (23):

[0109]

[0110] Considering the influence factors of node attributes, the algorithm redefines the node walking strategy. Specifically, the transition probability formula of the node is defined as shown in formula (24), the starting node is defined as c0, and the i-th node in the subsequent walk is defined as c i :

[0111]

[0112] Among them,, P(c i =x|c i-1 =w) represents the probability of walking to user node x based on user node w. c i-1 represents the current user node, c i represents the next user node, α p,q (w,x) represents the weight adjustment parameter between user node w and user node x. Its calculation formula is as shown in formula (25), and β(w,x) is the similarity between user node w and user node x. Extract the relevant attributes of the user, calculate the Euclidean distance between the two users to measure the similarity between the two users. Of course, it can also be measured by other metrics, and the present invention does not limit this. γ(w,x) represents the edge weight value of the rumor topic network propagated between user node w and user node x; E represents the edge set of the topic network, and z represents the number of nodes in the topic network. Since the social relationship between potential users and hot users will affect their participation behavior in hot topics, and the rumor-refutation mutual influence of hot users on potential users will affect their behavior of participating in topic dissemination. First, use the similarity between users to assign edge weights to the network G t The definition of the edge weight (w,x) is as shown in formulas (25) and (26).

[0113]

[0114]

[0115]

[0116] Among them, I w,x represents whether user x follows user w, and text kb is user x's behavior b towards user w iThe generated text, where b = 1 represents a forwarding behavior, b = 2 represents a commenting behavior, b = 3 represents a liking behavior, t represents the time of the current topic, t k represents the time when user w generates the k-th text, and K represents the total number of texts generated by user w. W(w,x) represents the mutual influence between user node w and user node x, and interact(w,x) represents the interaction degree between user node w and user node x. Since the user interaction intensity has timeliness, a time decay function is introduced for dynamic optimization.

[0117] Finally, the CSR2Vec algorithm is used to represent the hot topic topology as a low-rank dense feature vector matrix, as shown in formula (28):

[0118]

[0119] where, E S represents the feature topology matrix of the topic network, d s represents the dimension of the feature vector, and N represents the number of user nodes in the topic space.

[0120] 107. Concatenate the feature topology matrix of the topic network and the user adjacency matrix, and fuse them to form a feature vector matrix;

[0121] In the embodiment of the present invention, it is necessary to fuse the two key features of the feature topology matrix and the user adjacency matrix. Concatenate the feature topology matrix E S of the topic network and the user adjacency matrix A to fuse them into a feature vector matrix E, as shown in formula (29) specifically:

[0122]

[0123] Use E = {E1, E2, E3,..., E n} to represent the input data, where E i is a vector of size d s +d a , and the adjacency matrix A = N×N, which is used to represent the connection information of nodes in the rumor topic propagation hybrid network.

[0124] 108. Input the fused feature vector matrix into the graph attention network model with an attention mechanism, and output the prediction results of whether the user participates in the rumor topic, or forwards the rumor information or forwards the rumor refutation information.

[0125] Considering the short-term emotional tendency of users towards rumor-refuting messages, the emotion contained in the user text is extracted as one of the main influencing features, and evolutionary game theory is introduced to depict the game process between rumor messages and refuting messages. Since social networks are typical non-Euclidean data and traditional convolutional networks cannot process them, researchers use GCN networks to process graph-structured data. However, because GCN convolution is not flexible enough when fusing graph-structured features, the graph attention network GAT with attention is selected in this invention. Finally, a network rumor propagation model based on users' short-term emotions and evolutionary games (Cooperation and Opposition influence GAT, abbreviated as CO-GAT) is proposed. The goal of the model of this invention is to predict the participation of potential user nodes in rumor topics. If they participate, it is judged whether the users forward rumor messages or refuting messages. Therefore, the prediction task can also be transformed into a three-classification task.

[0126] Input the fusion vector E = {E1, E2, E3, …, E n} into the model, and output E′ = {E′1, E′2, E′3, …, E′ n}; where E′ i is a vector of size d s +d a , specifically as shown in formula (30):

[0127]

[0128] Among them, sigmod represents a non-linear function, N i represents the domain composed of all nodes adjacent to node i, h J represents the feature vector of node j, and W is a weight matrix of size (d s +d a ) × ((d s +d a ). Among them, α ij represents the attention coefficient between the i-th node and the j-th node, and the calculation formula is as follows:

[0129]

[0130] Among them, the calculation formula of e ij is:

[0131]

[0132] Among them, LeakyReLU is a non-linear activation function, is a vector of size 2(d s +d aThe weight vector of ( ) is multiplied with it to get a real number, and finally a single number, i.e., the attention coefficient, is obtained.

[0133] Finally, through the fully connected layer for the output of the pooling layer, let the model output Z = P(r, a, d|u i ), and the specific definition is as follows:

[0134]

[0135] After the output of the model, the forwarding behavior of the user at time t + 1 can be obtained. Y = 1 indicates that the user will forward the rumor in the next stage, Y = 0 indicates that the user will no longer participate in the rumor topic in the next stage, and Y = -1 indicates that the user will forward the rumor refutation message in the next stage.

[0136] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium can include: ROM, RAM, magnetic disk, optical disc, etc.

[0137] 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 spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games, characterized in that, The method includes: Obtain the user basic information, user behavior data, and user text data of the social network platform, and preprocess the obtained data; Calculate the user's own attributes, user influence, user topic participation, message popularity, and friend driving force respectively according to the preprocessed user basic information and user behavior data; Generate word frequency features using the TF-IDF algorithm according to the preprocessed user text data, and process the word frequency features using a sentiment classifier to generate message sentiment scores; Calculate the rumor influence and the rumor refutation influence using the multiple regression linear algorithm according to the user's own attributes, user influence, user topic participation, message popularity, message sentiment scores, and friend driving force; Calculate the strategic benefits of the user forwarding rumor information and forwarding rumor refutation information respectively according to the rumor influence and the rumor refutation influence, and use an evolutionary game model to measure the mutual influence between rumor information and rumor refutation information; Map user nodes to a vector space using a node embedding algorithm, and generate a feature topology matrix of the topic network using a random walk algorithm based on mutual influence; Concatenate the feature topology matrix of the topic network and the user adjacency matrix, and fuse them to form a feature vector matrix; Input the fused feature vector matrix into a graph attention network model with an attention mechanism, and output the prediction results of whether the user participates in the rumor topic, or forwards rumor information or forwards rumor refutation information.

2. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 1, characterized in that, The step of generating word frequency features using the TF-IDF algorithm according to the preprocessed user text data and processing the word frequency features using a sentiment classifier to generate message sentiment scores includes performing time slicing processing on the preprocessed user text data, and selecting the text set that the user forwards, likes, and comments on within a certain time period; using the TF-IDF algorithm to perform word frequency processing on the text set of the user within a certain time period to generate the word frequency features of the user within a certain time period, and inputting the word frequency features into the sentiment classifier to generate the message sentiment scores of the user within a certain time period.

3. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 1 or 2, characterized in that, The rumor influence and the rumor refutation influence are respectively expressed as: Among them, influence rumor (u i ,u j ) represents the rumor influence between user u i and user u j . influence anti-rumor (u i ,u j ) represents the rumor refutation influence between user u i and user u j . are the partial regression coefficients obtained by training with the multiple linear regression algorithm respectively, represents the overall influence coefficient during the training process, represents the influence coefficient of user factors on user behavior during the training process, represents the influence coefficient of message factors on user behavior during the training process; userfactor(u i ) represents the user factor of user u i , userfactor(u i ) = User(u i ) * userInfluence(u i ) * participate(u i ), User(u i ) represents the self - attribute of user u i , userInfluence(u i ) represents the influence of user u i , participate(u i ) represents the topic participation degree of user u i ; mesfactors rumor (u i ,u j ) represents the message factor of the rumor message spread between user u i and user u j , mesfactors anti-rumor (u i ,u j ) represents the message factor of the rumor refutation message spread between user u i and user u j . Popularity(t) represents the message popularity of the rumor message or the rumor refutation message at time t, represents the message sentiment score of user u i , represents the friend driving force of user u i towards user u j .

4. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 1, characterized in that, The mutual influence between the rumor information and the rumor refutation information is expressed as: where, m(u i , u j ) = MutualInf rumor (u i , u j ) - MutualInf rumor (u i , u j ). If i = j, then m(u i , u j ) = 0. MutualInf anti-rumor (u i , u j ) represents the influence of user u j 's spreading of rumor-refuting messages on the spreading behavior of user u i . MutualInf rumor (u i , u j ) represents the influence of user u j 's spreading of rumor messages on the spreading behavior of user u i .

5. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 1, characterized in that, The step of generating a feature topology matrix of the topic network using a random walk algorithm based on mutual influence includes calculating the interaction degree between users according to the text set of the user's behavior based on forwarding, liking, and commenting, and calculating the edge weights of the rumor topic network using the mutual influence and the interaction degree between users; calculating the weight parameters of the users according to the path length between users; calculating the transition probability of user nodes according to the similarity between users, the edge weights of the rumor topic network, and the weight adjustment parameters between users; and generating a feature topology matrix of the topic network by walking according to the transition probability of user nodes.

6. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 5, characterized in that, The step of generating a feature topology matrix of the topic network by walking according to the transition probability of user nodes includes calculating the probability of the neighborhood node vector appearing in the user node vector according to the transition probability of the user node, and obtaining the representation of each user node by maximizing the probability of the neighborhood node vector appearing in all user node vectors, and forming a feature topology matrix of the topic network.

7. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 5 or 6, characterized in that, The transition probability of user nodes is expressed as: Among them, P(c i = x|c i-1 = w) represents the probability of walking to user node x based on user node w, where c i-1 represents the current user node, c i represents the next user node, α p，q (w, x) represents the weight adjustment parameter between user node w and user node x, β(w, x) is the similarity between user node w and user node x; γ(w, x) represents the edge weight value of the rumor topic network propagated between user node w and user node x, E represents the set of edges of the topic network, and z represents the number of nodes in the topic network.

8. The method for predicting the spread of online rumors based on users' short-term emotions and evolutionary games according to claim 7, characterized in that, The edge weight value of the rumor topic network spread by the user node w and the user node x is expressed as: Among them, interact(w, x) represents the interaction degree between user node w and user node x, W(w, x) represents the mutual influence between user node w and user node x, and U t represents the set of user nodes at time t.

9. A method for predicting the spread of online rumors based on users' short-term emotions and evolutionary game according to claim 8, characterized in that The interaction degree between the user node w and the user node x is expressed as: Among them, I w，x indicates whether user x follows user w, text kb is the text generated by user x for user w based on behavior b i where b = 1 represents a forwarding behavior, b = 2 represents a commenting behavior, b = 3 represents a liking behavior, t represents the time of the current topic, t k represents the time when user w generates the k-th text, and K represents the total number of texts generated by user w.

Images