Malicious-anti-malicious information propagation prediction method based on high-order propagation network

Abstract

The application belongs to the field of information propagation analysis and processing, and particularly relates to a malicious-anti malicious information propagation prediction method based on a high-order propagation network. The method comprises the following steps: obtaining user behavior records, information participation records and information propagation conditions under a propagation topic of a social network platform, and performing preprocessing; extracting user self attributes, user group attributes and user influence; calculating multi-dimensional information influence of the user; calculating propagation driving forces of malicious information and anti malicious information according to a dynamic game theory; dividing the user into four types of states, and establishing a propagation model; and according to the propagation driving forces of the malicious information and the anti malicious information, solving the probability of participation of any user in malicious information and anti malicious information propagation at any time by using a dynamic equation of the propagation model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of information propagation analysis and processing, and relates to user information propagation law analysis, in particular, common propagation analysis of malicious and anti-malicious information in a directed social network, and specifically relates to a malicious-anti-malicious information propagation prediction method based on a high-order propagation network. BACKGROUND

[0002] Malicious information in an online social network platform refers to fraudulent, hype, rumor and other false information published and propagated by users; compared with traditional interpersonal relationship propagation, malicious information propagation in an online social network platform has the characteristics of faster information propagation speed, wider influence, stronger power and stronger concealment. Since the development of new social media platforms such as Facebook, Twitter and microblog, people can receive and publish any information they want at any time and any place. In addition, since network propagation is mostly anonymous, people publish information with or without intention and irresponsibly, which leads to a doubling of the number of malicious information propagated on the network; the publication and propagation of such malicious information disrupts the good network social environment; therefore, the research on the trend of malicious information propagation is of great significance for the guidance of social public opinion and the control of emergencies. In recent years, the industry and academia have conducted a lot of research on malicious information propagation, and the current research on malicious information propagation is mainly divided into two directions: one is to establish a traditional information propagation model to predict the propagation trend, and the classic ones are the Susceptible-Infected-Recovered model SIR and the Susceptible-Infected-Susceptible model SIS. The second is based on deep learning algorithm, according to some user features and network features in social propagation network to make propagation prediction.

[0003] Although the above methods can solve the problem of malicious information propagation prediction in a social network platform to some extent, the deep learning method is difficult to understand and explain the simulation results, and the infectious disease model has high requirements for the artificial selection and extraction method of features, and there are still some challenges:

[0004] 1. The behavior of individuals in the propagation process is affected by the emotions implied in malicious information. Malicious information contains positive and negative emotions, and the propagation of individuals will be promoted or inhibited by such emotions, and how to quantify such inhibition and promotion becomes a challenge.

[0005] 2. There is a difference in propagation efficiency between different users. There is a hierarchical phenomenon in the user group in the propagation network, and the propagation efficiency of malicious information between layers has differences, and how to analyze the influence of such differences on information propagation becomes a problem.

[0006] 3. Symbiosis and antagonism of malicious information propagation. Malicious information and anti-malicious information exist simultaneously in the propagation space, and user propagation behavior is greatly influenced by the symbiosis and antagonism of the two. How to quantify this influence is a problem that needs to be considered in the process of propagation trend prediction. SUMMARY

[0007] In view of the problems existing in the prior art, the present application considers that in real social network propagation, the propagation behavior of a user is often influenced by multiple aspects, both the influence of the outside world on the user and the influence of the user's own preferences. Some users tend to propagate information containing positive emotions to establish their own social image, and some users tend to propagate information containing negative emotions to vent their own pressure through social platforms. At the same time, users will have different reactions when receiving information from different groups. For example, if the source of an information is a field expert with a large number of fans, the user will be more likely to believe it and spread and propagate the information; on the contrary, if a user who is unknown in the field suddenly publishes an information, it may not be very convincing. On the basis of this phenomenon, the present application proposes a malicious-anti-malicious information propagation prediction method based on a high-order propagation network. The method starts from the emotional needs of user propagation behavior, combines topic, user characteristics, and other quantitative multiple influences. At the same time, it comprehensively considers the differences in propagation efficiency between different user groups and the game relationship between malicious and anti-malicious information, combines an infectious disease model to construct an information propagation dynamics model, analyzes the information propagation trend under a certain topic, and verifies it through a data set. The results show that the model can more accurately depict the trend of malicious information propagation on real social networks.

[0008] A malicious-anti-malicious information propagation prediction method based on a high-order propagation network, the method comprising:

[0009] Obtaining user behavior records, information participation records, and information propagation conditions under a propagation topic of a social network platform, and preprocessing the obtained data;

[0010] From the preprocessed data, extracting user's own attributes, user group attributes, and user influence;

[0011] According to the user's own attributes, user group attributes, and user influence, calculating the multi-dimensional information influence of the user;

[0012] Based on the multi-dimensional influence of the user, calculating the propagation driving force of malicious information and anti-malicious information according to dynamic game theory;

[0013] The user is divided into a user group which has not contacted malicious-anti-malicious information, a user group which spreads anti-malicious information, a user group which selects to spread malicious information, and a user group which has contacted malicious-anti-malicious information and selected to ignore, and a spreading model is established;

[0014] According to the spreading driving force of malicious information and anti-malicious information, the probability of participation of any user in the spreading of malicious information and anti-malicious information at any time is solved by using the dynamic equation of the spreading model.

[0015] The present application has the following beneficial effects:

[0016] 1. The users are clustered into different groups by KNN clustering of the similarity of the spreading behaviors between different users, the user group attribute is calculated according to the group to which the user belongs, the user influence is calculated from the group level, and the driving effect of the group dimension on the user spreading decision is reflected.

[0017] 2. The score value of the user is calculated by the Pagerank algorithm combined with the relevance of the user and the topic, the core node set is established according to the score value, the high-order node influence and the low-order node influence are calculated according to the flow direction of information from the core node and the first-order neighbor node and the second-order neighbor node, so that the model is more three-dimensional, the influence of different users is more accurately described and conforms to the reality.

[0018] 3. The efficiency factor is described by the user score value and the scaling factor, the low-order node influence and the high-order node influence are measured by the efficiency factor, the different spreading situations of the social network are refined, and the fitting of the influence calculation and the actual decision is increased.

[0019] 4. The analysis and fitting of the existing malicious information spreading data are helpful to find the influencing factors of malicious information spreading, master the key links of information spreading, facilitate the relevant departments to take a series of effective measures to curb the malicious information spreading, so as to purify the network environment. In addition, the existing model can also be combined to predict the scale of future malicious information spreading, so as to take precautions in advance and better control the network public opinion. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a malicious-anti-malicious spreading dynamics model based on a high-order spreading network of an embodiment of the present application;

[0021] Figure 2 is an information spreading model architecture diagram of an embodiment of the present application;

[0022] Figure 3 is a multi-dimensional influence constitution diagram of an embodiment of the present application;

[0023] Figure 4 is a user spreading prediction schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0025] Figure 1 is a framework diagram of an information propagation prediction method according to an embodiment of the present application, as shown in Figure 1 which shows that the input of the present application is user behavior records, information participation records and information propagation conditions under a certain propagation topic, and after processing, user self attributes, user group attributes and user low-order node influence and user high-order node influence are extracted, and the multi-dimensional information influence of the user is calculated through these information; the multi-dimensional information influence of the user is calculated according to the dynamic game theory to calculate the propagation driving force of malicious information and anti-malicious information, and the malicious information driving force and the anti-malicious information driving force are input into the SAIR propagation model to output the propagation result of the user, the information propagation trend and the information game coefficient. In this way, the probability of any user propagating malicious information or anti-malicious information at any time can be effectively obtained, the propagation trend is formed, and it is more conducive for relevant institutions or departments to control information propagation according to needs.

[0026] Figure 2 is a framework diagram of an information propagation model according to an embodiment of the present application, as shown in Figure 2 which shows that the present application first abstracts a complex social network into a core node layer, a strong node layer and a weak node layer, calculates the corresponding low-order node influence and high-order node influence according to different levels, and calculates the propagation behavior similarity between different users according to the number of times of the same propagation behavior for a certain information; the propagation behavior similarity is KNN clustered, and the users are clustered into different groups; on the other hand, the individual influence is described by using emotional information, and the multi-dimensional information influence is calculated by using the low-order node influence and high-order node influence set group attributes; the real social network is abstracted by using the multi-dimensional influence, and finally the propagation model is established; the propagation of malicious information and anti-malicious information is predicted by using the dynamic game method, and the probability of any user participating in the propagation of malicious information and anti-malicious information at any time is obtained.

[0027] In the embodiments of the present application, the malicious-anti-malicious information propagation prediction method based on a high-order propagation network of the present application can refer to the following steps:

[0028] 101, Obtain user behavior records, information participation records and information propagation conditions under a spreading topic of a social network platform, and pre-process the obtained data;

[0029] In the embodiment of the application, the spreading topic data of the social network platform can be obtained online, and the data can be obtained from a public data website or by using a mature social network public API. What is needed here is the behavior records and information participation records of all participants in a certain topic during the life cycle of malicious-anti-malicious information and the propagation conditions of the information. The information participation records need to obtain the time when the information is propagated, the basic information of the participating users and the friend relationship information of the participating users; the behavior records of the information propagation participants include the malicious information records historically propagated by the user.

[0030] In the embodiment of the application, when the above original data is obtained, the obtained original data is mostly chaotic and cannot be directly used for data analysis and processing. Data cleaning can structure most of the chaotic data, delete abnormal data, and facilitate subsequent calculation. Further, the data after simple data cleaning needs to be stored by using a database, the data is further standardized by using a table structure, and the database can greatly improve the data retrieval efficiency and the mapping of the inter-table relationship.

[0031] 102, Extract user attributes, user group attributes and user influence from the pre-processed data;

[0032] In the embodiment of the application, related attributes need to be extracted from the pre-processed data. The three key influencing factors of emotional similarity, content matching degree and individual activity degree are mined from user labels and historical behaviors. At the same time, the connectivity of the user and the friend nodes and the propagation behavior similarity are extracted. Finally, the message influence is quantified by using a polynomial equation from the two angles of the user friend relationship and the information itself.

[0033] For the extraction of related attributes in the above step 102, the following three steps are mainly divided:

[0034] Exploring the malicious information propagation process based on a high-order propagation network is essentially extracting related features from the user himself, the user's friends and the user's group to which the user belongs and modeling analysis. The application extracts related attributes that affect information propagation from the user's own attributes, the influence of the same-order friends on the user, the influence of the high-order friends on the user and the influence of the group attributes on the user, and analyzes the dynamic causes driving information propagation.

[0035] S21: User's own attributes.

[0036] S211: Emotional similar Semo(v i )

[0037] Emotional similar is used to describe the similarity between the emotional polarity of the user's recent information dissemination and the emotional polarity of the malicious information. The higher the similarity, the greater the probability of user dissemination. In the embodiment of the present application, the Bert model is constructed through keras-bert library to analyze the emotional polarity of malicious information and user historical information, and output the emotional polarity label. The emotional similar is defined as follows:

[0038]

[0039] Wherein, when the user v i has the same emotional label as the emotional polarity of the malicious-anti-malicious information, the emotional similar Semo(v i ) is 1, and when the user has different emotional labels, the emotional similar Semo(v i ) is 0.

[0040] S212: Content matching degree Info match (v i )

[0041] Info match (v i ) is used to describe the matching of malicious-anti-malicious information label and user favorite label. The higher the matching degree, the greater the probability of user dissemination. The definition of content matching degree in the present application is as follows:

[0042]

[0043] Wherein, Tag(mess) is the malicious-anti-malicious information content label, and Tag(user) is the user favorite label.

[0044] S213: Individual initiative Initiative(v i

[0045] Individual initiative is defined as the degree of user dissemination of information without relying on other users, wherein n represents that there are n users in the social network, and d(v i ,v j ) represents the shortest distance from user v i to all other users, then the user initiative can be defined as:

[0046]

[0047] S22: User influence ​

[0048] S221: Friend contact co(v i , v j )

[0049] contact co(v i , v j ) is used to describe the closeness of the relationship between users and friends. The more common friends and the more similar information preferences, the stronger the contact between them, and the more likely the user is to spread the information. The definition of the calculation formula is as follows:

[0050] co(v i , v j ) = similartag(v i , v j ) + similarfriends(v i , v j ) (4)

[0051] Wherein, similartag(v i , v j ) represents the similarity of tags between two users, and similarfriends(v i , v j ) represents the number of common friends.

[0052] S222: Friend drive Frideve(v i )

[0053] Friend drive(v i ) is defined as the degree of information spread by the user's friends being viewed, commented on, and spread by others. The greater this value indicates that the user's friends have stronger driving force, and the user is more likely to be affected and thus spread. Based on the user's historical data, the formula is defined as follows:

[0054]

[0055] Wherein, Nums[comt(v j )], Nums[ret(v j )], Nums[read(v j )] represent the average number of comments, spreads, and readings of each information of the friends, respectively, and μ is a weakening factor.

[0056] S223: Node influence

[0057] The user influence is actually the friend influence, and the user influence is divided into high-order node influence and low-order node influence in the application; in order to calculate the high-order node influence and the low-order node influence, a high-order propagation network needs to be constructed, and the high-order propagation network can divide the users into different levels and is represented as {N c},{N h},{N l}.

[0058] The application calculates each connection between user nodes as a vote by the Pagerank algorithm, combines the importance of the voting nodes, and finally obtains a weighted sum as a score value; the user with a score pr(v i ) exceeding a preset threshold is regarded as a core node, and a core node set is established. Then, the node connected with the core node by a direct edge is defined as a strong node, and the node connected with the strong node by a direct edge is defined as a weak node, and finally, the strong node set, the weak node set and the core node set are obtained; that is, the first-order neighbor node of the core node is regarded as a strong node; the second-order neighbor node of the core node is regarded as a weak node; the high-order node influence is calculated by flowing information from the core node to the strong node or from the strong node to the weak node; and the low-order node influence is calculated by flowing information from the weak node to the strong node or from the strong node to the core node.

[0059] S2231: high-order node influence

[0060] When the information flows from the core node to the strong node or from the strong node to the weak node, the propagation has stronger infectivity (the infectivity difference is quantified by an efficiency factor r, and r is positively correlated with the improved Pagerank algorithm comprehensive score mpr(v i )), and the definition of hnodeinf(v i ,g h ) is as follows:

[0061] hnodeinf(v i ,g h ) = co(v i , v j ) * Fridrive(v i ) * (1 + r) (6)

[0062] S2232: low-order node influence

[0063] Similarly, the low-order node influence represents the information flowing from the weak node to the strong node or from the strong node to the core node, and has a similar influence calculation method as the high-order node influence, and the difference is that the information propagation sources of the two are different, and therefore the propagation efficiency is reduced.

[0064] lnodeinf(vi ,g l ) = co(v i , v j ) * Fridrive(v i ) * (1 - r) (7)

[0065] In the present application, the efficiency factor is calculated as follows:

[0066] r = τ * mpr(v i ) (8)

[0067] mpr(v i ) represents the comprehensive score of user v i , which is obtained by linear combination of the score pr(v i ) of the Pagerank algorithm and the user fit degree Info match (v i ). That is, the importance of the user node in the social network topology is scored to obtain the pr(v i ) value. Secondly, the comprehensive score mpr(v i ) is obtained by linear combination of the fit degree between the user and the information. τ is a scaling factor. In this way, the efficiency factor r is used to quantify the difference in infectivity, and the score of the user is used to depict the efficiency factor, so that the influence of high-order nodes and low-order nodes is more accurate.

[0068] S23: User group attribute

[0069] S231: Divide groups {G1}, {G2}, {G3}...

[0070] In order to divide the groups, the propagation behavior similarity prb(v i ,v j ) needs to be calculated first. The propagation behavior similarity prb(v i ,v j ) is defined as the number of times that user v i and user v j make the same propagation behavior on a certain information. In the case of the same total amount of information, the more the number of information propagated by the user, the higher the intimacy between the two, and the greater the propagation promotion.

[0071] S232: The propagation behavior similarity formula is defined as follows:

[0072] prb(v i ,v j ) = | (a, v ij ) ∈ (Dv i ∩Dvj )| (9)

[0073] wherein, Dv i represents the historical behavior data set of user v i , Dv j represents the historical behavior data set of user v j , and (a, v ij ) represents the same reaction of user v j and user v j to the same information.

[0074] Then, the propagation behavior similarity is taken as the input of the KNN algorithm, and the users are clustered into different groups {G1}, {G2}, {G3}...

[0075] In the preferred embodiment of the present application, other clustering algorithms can also be used to cluster the propagation behavior similarity by using the clustering algorithm to divide the users into different groups.

[0076] S233: Group activity Gact(g i )

[0077] In the group attribute, the group activity Groupactive(g i ) refers to the historical information propagation of the group to which the user belongs, and the more frequent the propagation of the group to the information, the more active the group to which the user belongs. The user with higher group activity is more likely to participate in information propagation.

[0078] The definition of the group activity of the present application is as follows:

[0079] Gact(g i ) = σ*Nbs[origmes(v i )] + γ*Nbs[(partmes(v i )] (10)

[0080] wherein, Gact(g i ) represents the amount of original and propagated information of the group g i in a period of time before the malicious information appears, Nbs[origmes(v i )] represents the total amount of original information of the group g i in a period of time, and Nbs[(partmes(v i )] represents the total amount of propagated information of the group g i in a period of time. σ and γ are attenuation factors.

[0081] S234: Group preference Gpre(g i )

[0082] Group preference Gpre(g i ) is used to describe the similarity between the group propagation preference and the malicious information tag. The present application collects and collates the group g i The information content propagated one month before the outbreak of malicious information is extracted by the Textrank algorithm, and then normalized calculation is performed by using the Jaccard coefficient. The higher the Jaccard coefficient is, the greater the degree of fit between the group propagation preference and the current malicious-anti-malicious information, and the more likely the user in the group is to make a propagation behavior. The group preference is defined as follows:

[0083]

[0084] Wherein, Gtag(g i ) is the group propagation keyword, representing the group's preference for information propagation, and Tag(mess) is the malicious-anti-malicious information keyword, representing the individual's preference for information propagation.

[0085] S235: User group attribute Groupatt(v i ):

[0086] Groupatt(v i )=Gact(g i )*Gpre(g i ) (12)

[0087] In the embodiment of the present application, a high-order propagation network with group attributes is established according to the extracted related attributes. Secondly, on the established propagation network, the propagation driving force of malicious information and anti-malicious information is calculated according to dynamic game theory. Finally, a malicious information propagation driving force model is constructed based on the infectious disease model. The model shows that whether a user participates in the malicious-anti-malicious information propagation is affected by the user's friend relationship, the user's propagation preference, and the cognitive game result.

[0088] The model mainly includes three stages: quantifying the multi-dimensional information influence of different groups on the propagation network, quantifying the malicious and anti-malicious information driving force, measuring the probability of user information propagation, and designing the model algorithm. In the first stage, the multi-dimensional information influence is quantified by multiple linear regression by comprehensively considering the user's own attributes, low-order node influence, high-order node influence, and group attributes. In the second stage, the propagation driving force is calculated by using game theory by comprehensively considering the different benefits of the user in the propagation process when propagating malicious information and propagating anti-malicious information. In the third stage, an information propagation dynamics prediction model based on high-order propagation network and infectious disease model is established to predict and analyze the propagation situation of malicious and anti-malicious information. The specific implementation of each stage will be described below in combination with steps 103-106.

[0089] 103. Calculate the multi-dimensional information influence of the user according to the user's own attributes, user group attributes and user influence;

[0090] In a social network, whether a piece of information can be successfully spread is affected by many aspects and dimensions. From the individual level, the information propagation influence is affected by the user's own attributes, such as emotional similarity, individual initiative and content fit; from the group level, the influence is affected by the group attributes, such as group similarity and group activity. From the low dimension, the propagation behavior of friends will enhance the influence of information; from the high dimension, the information propagated by high-order nodes is more likely to be accepted by people and has higher propagation efficiency. As shown in the following formula, the definition of multi-dimensional information influence is as follows: Figure 3

[0091] Multeff(v i )=γ0+γ1*Useratt(v i )+γ2*lnodeinf(v i ,g l )+γ3*hnodeinf(v i ,g h )+γ4*Groupatt(v i ) (13)

[0092] Wherein, γ0, γ1, γ2, γ3, γ4 represent different partial regression coefficients, which are fitted by a multiple linear regression model, γ0 represents the balance coefficient of multi-dimensional information, γ1, γ2, γ3, γ4 reflect the weight coefficients of user's own attributes, low-order node influence, high-order node influence and group attributes in information influence in turn; Useratt(v i ) represents the own attributes of user v i , lnodeinf(v i ,g l ) represents the low-order node influence of user v i , hnodeinf(v i ,g h ) represents the high-order node influence of user v i , Groupatt(v i ) represents the group attributes of user v i .

[0093] 104. Based on the multi-dimensional influence of the user, calculate the propagation driving force of malicious information and anti-malicious information according to the dynamic game theory;

[0094] ​In this embodiment of the invention, considering that social networks are complex network environments where malicious and anti-malicious information often coexist, when users face a choice between these two types of information, their own profit-seeking and herd mentality will drive them to choose information with wider dissemination and higher returns. Therefore, evolutionary game theory is introduced to obtain user v i Drif1 (v) is a driving force behind the spread of malicious information. i ):

[0095]

[0096] Here, earn1 and earn2 represent the profit functions for spreading malicious information and countering malicious information, respectively. Assume user v... i If the proportion of friends spreading malicious information is P1, then the proportion of friends spreading anti-malicious information is 1-P1 (there may be friends who do not forward either type of information, but they are not necessarily interested in user v). i The forwarding strategy has no impact. Therefore, the definitions of earn1 and earn2 are as follows:

[0097] earn1(v i =P1*Multeff(v i (15)

[0098] earn2(v i )=(1-P1)*Multeff(v i (16)

[0099] Similarly, we can obtain the user's anti-malicious information propagation driving force Drif2(v i ):

[0100]

[0101] 105. Divide users into groups that have not yet been exposed to malicious-anti-malicious information, groups that spread anti-malicious information, groups that choose to spread malicious information, and groups that have been exposed to malicious-anti-malicious information but have chosen to ignore it, and establish a propagation model.

[0102] In this embodiment of the invention, considering the inhibitory effect of anti-malicious information on the spread of malicious information in real social networks, the present invention introduces an anti-malicious information state (Anti-infection: refers to the user group that spreads anti-malicious information) in the traditional SIR model; while state S refers to Susceptible: the user group that has not yet been exposed to malicious-anti-malicious information; state I refers to the malicious information spread state Infection: the user group that has chosen to spread malicious information; state R refers to the removed state Removed: the user has been exposed to both types of messages and has chosen to ignore them.

[0103] In order to better realize the propagation prediction, the rumor forwarding rules are defined in the embodiment of the present application as follows:

[0104] The propagation model defined in the present application is based on the following three assumptions:

[0105] Since the propagation of malicious information has the characteristics of outbreak and short duration, the present application considers that the total number of users participating in information propagation remains basically unchanged during the propagation period, S+A+I+R=1.

[0106] Meanwhile, the malicious information is contact transmission, and as long as the new user contacts the malicious information, it will have a certain infection rate.

[0107] The malicious information does not exist permanently, and its life cycle is set to 12 hours from the time the user contacts it.

[0108] On the basis of the above assumptions, the present application defines the malicious information propagation rules of social network as follows:

[0109] After the susceptible node contacts the malicious information node and the anti-malicious information node, it is converted into a malicious anti-malicious information node with a probability of α and β, respectively. Since part of the nodes do not participate in information propagation from beginning to end, α+β<1.

[0110] Since the information propagation has a certain life cycle, the malicious-anti-malicious information node will change into a removed node after a period of time. The present application assumes that the probabilities are ε and γ, respectively. And α, β, γ, ε∈(0, 1).

[0111] According to the above rules, the following kinetic equations are obtained:

[0112]

[0113] The information propagation has unidirectionality, and the state transition of the user is also the same. The user changes from the susceptible state to the malicious information propagation state or the anti-malicious information propagation state, and finally becomes a removed state with the passage of time. Assuming that a user v i has m neighbors, then the probability that k neighbors participate in malicious information propagation satisfies the binomial distribution:

[0114]

[0115] Therefore, the probability that any user participates in malicious information propagation at time t is:

[0116]

[0117] Similarly, the probability that any user participates in anti-malicious information propagation at time t is:

[0118]

[0119] Combining the mean field equation, we have:

[0120]

[0121] 106、According to the propagation driving force of malicious information and anti-malicious information, the probability of any user participating in the propagation of malicious information and anti-malicious information at any time is solved by using the kinetic equation of the propagation model.

[0122] Figure 4 is the schematic diagram of information propagation prediction of the user in the embodiment of the application, as shown in Figure 4 It can be seen that the state of each type of user at the next time can be calculated by the above-mentioned SAIR model.

[0123] Through the malicious information dynamic propagation model constructed by the application, the system can predict the malicious information propagation trend of a certain topic on the current social network, output the user state ratio at each time and the propagation trend diagram. On the basis of the output data, the relevant departments can understand the propagation situation and scale of the malicious information, and effectively manage and control the network public opinion as soon as possible.

[0124] Although the embodiments of the application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and modifications can be made to these embodiments without departing from the principles and spirits of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the spread of malicious-anti-malicious information based on a high-order propagation network, characterized in that, The method comprises: Obtaining user behavior records, information participation records and information dissemination conditions under a spread topic of a social network platform, and preprocessing the obtained data; From the preprocessed data, user self attributes, user group attributes and user influence are extracted; The user self attributes comprise: Useratt(v i ) = Semo(v i ) * Info match (v i ) * Initiative(v i ) Among them, Semo(v i ) represents user v i Emotional similarity, Info match (v i ) represents user v i Content relevance, Initiative(v i ) represents user v i Individual initiative; The user group attributes comprise: Groupatt(v i ) = Gact(g i ) * Gpre(g i ) Among them, Gact(g i ) represents user v i Group activity, Gpre(g i ) represents user v i Group preferences; The user influence comprises high-order node influence and low-order node influence; wherein the calculation formula of node influence comprises combining the PageRank algorithm with the user-topic relevance degree to obtain the score of each user, taking the user whose score exceeds a preset threshold as a core node; taking the first-order neighbor node of the core node as a strong node; taking the second-order neighbor node of the core node as a weak node; calculating the high-order node influence by flowing information from the core node to the strong node or from the strong node to the weak node; calculating the low-order node influence by flowing information from the weak node to the strong node or from the strong node to the core node; According to the user self attributes, user group attributes and user influence, the multi-dimensional information influence of the user is calculated; The multi-dimensional information influence comprises: Multeff(v i ) = γ0+ γ1*Useratt(v i )+ γ2*lnodeinf(v i ,g l )+ γ3 hnodeinf(v i ,g h )+γ4*Groupatt(v i ) Wherein, γ0, γ1, γ2, γ3, γ4 represent different partial regression coefficients, fitted by a multiple linear regression model, γ0 represents a balance coefficient of multi-dimensional information, γ1, γ2, γ3, γ4 reflect the weight coefficients of user's own attributes, low-order node influence, high-order node influence, and group attributes in information influence in turn; Useratt(v i ) represents the own attributes of user v i , lnodeinf(v i , g l ) represents the low-order node influence of user v i , hnodeinf(v i , g h ) represents the high-order node influence of user v i , and Groupatt(v i ) represents the group attributes of user v i ; Based on the multi-dimensional influence of the user, the propagation driving force of malicious information and anti-malicious information is calculated according to the dynamic game theory; User v i Malicious information propagation driving force Drif1(v i ): User's anti-malicious information propagation driving force Drif2(v i ): Wherein, earn1 and earn2 represent the income functions of propagating malicious information and anti-malicious information respectively; The users are divided into a user group that has not contacted malicious-anti-malicious information, a user group that propagates anti-malicious information, a user group that propagates malicious information, and a user group that has contacted malicious-anti-malicious information and selected to ignore, and a propagation model is established; According to the propagation driving force of malicious information and anti-malicious information, the dynamics equation of the propagation model is used to solve the probability of any user participating in the propagation of malicious information and anti-malicious information at any time; Dynamics equations of the propagation model: 2.The method of claim 1, wherein, The division method of the user group comprises calculating the propagation behavior similarity between different users according to the number of times that different users make the same propagation behavior on a piece of information; and performing KNN clustering on the propagation behavior similarity to cluster the users into different groups. 3.The method of claim 1, wherein, The user v i The calculation of the group activity of the user v includes: Gact(g i ) = σ * Nbs[origmes(v i )] + γ * Nbs[(partmes(v i )] where Nbs [origmes(v i )] represents the total amount of original information in the population g i at a time; Nbs [(partmes(v i )] represents the total amount of propagated information in the population g i at a time; and σ and γ are different decay factors. 4.The method of claim 1, wherein, User v i The calculation method of the group preference of the user v includes collecting the group g i The information content propagated in a period before the malicious information outbreak, the keywords of the malicious-anti malicious information content are extracted by Textrank algorithm, and the Jaccard coefficient is used for normalization calculation, which is expressed as: where Gtag(g i ) is the group g i propagation keyword, Tag(mess) is the malicious-anti malicious information keyword.

5. The method of claim 1, wherein the method is characterized by, The user influence comprises: hnodeinf(v i , g h ) = co(v i , v j ) * Fridrive(v i ) * (1 + r) lnodeinf(v i , g l ) = co(v i , v j )*Fridrive(v i )*(1-r) wherein lnodeinf(v i ,g l ) represents the low-order node influence of user v i ,hnodeinf(v i ,g h ) represents the high-order node influence of user v i ,co(v i ,v j ) represents the friend connection of user v i and user v j ,Fridrive(v i ) represents the friend drive of user v i ,and r represents the efficiency factor. wherein lnodeinf(v i ,g l ) represents the low-order node influence of user v i ,hnodeinf(v i ,g h ) represents the high-order node influence of user v i ,co(v i ,v j ) represents the friend connection of user v i and user v j ,Fridrive(v i ) represents the friend drive of user v i ,and r represents the efficiency factor.

6. The method of claim 5, wherein the method is characterized by, The calculation method of the efficiency factor comprises: r = τ * mpr(v i ) where τ is a scaling factor, mpr(v i ) represents the overall score of user v i .

Images