A topic propagation prediction method based on topic relevance

Abstract

The application belongs to the field of network public opinion analysis, and particularly relates to a topic propagation prediction method based on topic association; comprising: obtaining topic information, extracting internal attributes and external attributes of the topic information; selecting user interest feature keywords and user cognitive feature keywords from a user topic content set by using a DTR2vec algorithm, performing vector representation on the selected keywords, and obtaining a user historical behavior feature vector; quantifying the influence of a topic on a user by using evolutionary game theory according to all internal attributes and part of the external attributes, and obtaining a topic influence adjacency matrix; extracting network structure features according to the topic influence adjacency matrix, and obtaining a user network structure feature vector; inputting the user historical behavior feature vector and the user network structure feature vector into a DT-GCN model, and obtaining a propagation prediction result of the user on the topic; and the application can help the public opinion department to quickly take targeted measures, and has a good application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of network public opinion analysis, and particularly relates to a topic propagation prediction method based on topic association. BACKGROUND

[0002] Generally, a topic refers to a hot issue that is most concerned by the public within a certain time and a certain range. Various diversified topics are widely propagated in social networks, carrying a large number of speeches and information behaviors of network users. These topic data reflect user interests, behaviors and social relationships, and the use of these data for research can effectively perform information recommendation. However, at the same time, in the real world, the information contained in the topic is mixed with true and false, and when the topic is widely propagated in the social group, it will cause a series of influences on people's cognition and the stability of the society.

[0003] With the development of the Internet, the mode of topic propagation has undergone tremendous changes. On the one hand, the booming social platforms such as microblog, WeChat and forum provide information exchange channels for the propagation of topics across time and space, covering all people. On the other hand, the sinking of Internet users makes the composition of network groups more diversified, the flow and propagation of information more flat, and the development and derivation of topics more complex. This means that compared with the traditional mode of information propagation, the topic propagation nowadays is faster, has a wider influence and is more complex. Therefore, the research on the propagation situation of topics can better understand the characteristics of information propagation, and is of great significance for the prevention of emergencies and the control of public opinion.

[0004] In recent years, scholars have made a series of researches on the propagation situation of topics in social networks from multiple dimensions and achieved remarkable results; with the continuous growth of the data volume of network social platforms and the maturity of deep learning technology, neural network and deep learning model based topic propagation prediction is favored. However, there are still many challenges for topic propagation, such as: 1. The relevance and complexity of the feature space of derivative topics. Compared with a single topic, derivative topics and original topics are intertwined with each other, and constantly exchange information dynamically, which is a challenge for how to effectively extract topic features; 2. The complex relationship of users in the original-derivative topic propagation process. In the process of topic propagation, the original topic and the derivative topic interact with each other, how to quantify the user influence of the original topic and the derivative topic and dig out the hidden relationship between users is a problem to be solved urgently; 3. The stage and timeliness of the dynamic evolution of derivative topics. The evolution trend of derivative topics changes dynamically with time and is affected by the original topic, and the topic state alternately evolves, how to dynamically analyze the propagation situation of derivative topics is a difficulty in current research.

[0005] Therefore, the application proposes a topic association-based information propagation prediction method, which introduces derivative topics, can not only effectively predict the propagation trend of topics, but also more truly reflect the correlation and game relationship between original topics and derivative topics in the propagation process. SUMMARY

[0006] In view of the deficiencies in the prior art, the application proposes a topic association-based topic propagation prediction method, which comprises the following steps: obtaining topic information and preprocessing the topic information; inputting the preprocessed topic information into a topic association-based topic propagation prediction model to predict the propagation trend of the topic by the user; and controlling the topic propagation trend according to the propagation trend of the topic by the user.

[0007] The process of processing the preprocessed topic information by the topic association-based topic propagation prediction model comprises the following steps:

[0008] S1: extracting internal attributes and external attributes of the topic information; the internal attributes include user activity, user topic perception rate and user historical forwarding rate, and the external attributes include user friend driving force, topic heat and user topic content set;

[0009] S2: selecting user interest feature keywords and user cognitive feature keywords from the user topic content set, and performing vector representation on the selected keywords to obtain a user historical behavior feature vector;

[0010] S3: quantifying the influence of the topic on the user by using evolutionary game theory according to all internal attributes and the user friend driving force and topic heat in the external attributes, to obtain a topic influence adjacency matrix;

[0011] S4: extracting network structure features according to the topic influence adjacency matrix to obtain a user network structure feature vector;

[0012] S5: inputting the user historical behavior feature vector and the user network structure feature vector into a DT-GCN model to obtain a user topic propagation prediction result, including whether the user participates in the topic propagation and the type of topic participated by the user.

[0013] Preferably, the process of selecting the user interest feature keywords and the user cognitive feature keywords from the user topic content set by using the DTR2vec algorithm comprises the following steps:

[0014] S21: the user topic content set comprises an original topic content set, a derivative topic content set and a user social content set; the three content sets are input into an LDA topic identification model respectively to obtain an original topic keyword set , a derivative topic keyword set and a user content keyword set ;

[0015] S22: Calculate the relevance between the original topic keyword set and the derived topic theme keyword set; calculate the similarity between the original topic content set and the derived topic content set, and put the two keywords with the relevance greater than the similarity into the topic association feature keyword set

[0016] S23: Calculate the interest weight and the cognitive weight of the user content keyword set according to the original topic keyword set, the derived topic theme keyword set, and the topic association feature keyword set

[0017] S24: Select Top-k keywords from the user content keyword set as the user interest feature keyword and the user cognitive feature keyword according to the interest weight and the cognitive weight respectively.

[0018] Further, the similarity between the original topic content set and the derived topic content set is calculated by using the BM25 algorithm, and the formula is:

[0019]

[0020] wherein, sim(Q, d) represents the similarity score of the original topic content set Q and the derived topic content set d, Q represents the original topic content set, d represents the derived topic content set, tf(Q, i) represents the word weight, Qi represents the i-th word in the set Q, |Q| represents the total number of words in the original topic content set Q, sim(Qi, d) represents the relevance between the i-th word in the set Q and the set d.

[0021] Further, the formula for calculating the interest weight and the cognitive weight of the user content keyword set is:

[0022]

[0023]

[0024] wherein, w(Qi) represents the interest weight of the i-th word in the user content keyword set, w(Qi) represents the cognitive weight of the i-th word in the user content keyword set, sim(Qi, T) represents the relevance between the i-th candidate keyword and the derived topic theme keyword set T, ​​​​​​​​​​​Similarity between them This represents the set of feature words related to the topic for the i-th candidate keyword. Similarity between them This represents the total number of words in the user content keyword set. This represents the set of candidate keywords and original topic keywords for the i-th keyword. Similarity between them Indicates the current moment. Indicates the moment when the initial topic was generated. This represents the regularization factor.

[0025] Preferably, the process of quantifying the influence of a topic on users using evolutionary game theory includes:

[0026] S31: Calculate internal influence based on internal attributes, and calculate external influence based on user friend influence and topic popularity;

[0027] S32: Based on internal and external influence, a multiple linear regression algorithm is used to calculate the influence of the original topic and the influence of the derived topic;

[0028] S33: Define two game strategies. Calculate the first payoff based on the first game strategy and the influence of the original topic, and calculate the second payoff based on the second game strategy and the influence of the derived topic.

[0029] S34: Calculate the influence of users' topic dissemination behavior in the original topic and the influence of users' topic dissemination behavior in the derivative topic based on the first and second benefits;

[0030] S35: Calculate the topic influence adjacency matrix based on the influence of users' topic propagation behavior in the original topic and the influence of users' topic propagation behavior in the derived topic.

[0031] Furthermore, the formulas for calculating internal and external influence are as follows:

[0032]

[0033]

[0034] in, Indicates user Its internal influence Indicates user For users External influence user activity level Indicates user Historical forwarding rate Indicates user Topic awareness rate representing the user to the user 's friend driving force, representing the topic heat of the current time t.

[0035] Further, the formula for calculating the topic propagation behavior influence of the user in the original topic and the topic propagation behavior influence of the user in the derivative topic is:

[0036]

[0037]

[0038] wherein, representing the influence of the user to the user , representing the influence of the user to the user , representing the first income, representing the second income.

[0039] Further, the topic influence adjacency matrix is represented as:

[0040]

[0041] wherein, representing the topic influence adjacency matrix, , representing the topic propagation behavior influence between users, representing the influence of the user to the user , representing the influence of the user to the user .

[0042] Preferably, the processing process of the DT-GCN model on the user historical behavior feature vector and the network structure feature vector of the user includes: obtaining an adjacency matrix according to the connection information between nodes in the original topic and derivative topic propagation mixed network, inputting the user historical behavior feature vector and the network structure feature vector of the user into a CNN network for convolution to obtain a feature matrix; inputting the feature matrix and the adjacency matrix into a GCN network, adding a drop layer in the GCN network, and finally adopting a softmax function for processing to obtain a propagation prediction result of the user on the topic.

[0043] Preferably, the expression of the DT-GCN model is:

[0044]

[0045] wherein, represents the category probability of user output, represents a preprocessed adjacency matrix, represents an adjacency matrix, represents an input layer, represents that a drop layer is added to randomly sample parameters according to a certain probability, represents a CNN network, is a weight matrix of the i-th layer network.

[0046] The present application has the beneficial effects that: the present application quantifies user interest and cognitive process by using the hidden relationship between original and derivative topics, focuses on the influence of the propagation process of associated topics on user behavior, and constructs a topic propagation prediction model based on topic association by combining topic features, user features and powerful neural networks; the present application introduces associated topics, which can not only effectively predict the propagation trend of topics, but also more truly reflect the correlation and game relationship between original and derivative topics in the propagation process, so that the prediction effect is more in line with the actual situation and has high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a structure schematic diagram of the topic propagation prediction method based on topic association in the present application;

[0048] Figure 2 is a process schematic diagram of extracting the network structure feature vector of the user in the present application;

[0049] Figure 3 is a structure schematic diagram of the DT-GCN model in the present application. DETAILED DESCRIPTION

[0050] 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. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The present application provides a topic propagation prediction method based on topic association, as shown in Figure 1 The method comprises the following contents: obtaining topic information and preprocessing the topic information; inputting the preprocessed topic information into a topic propagation prediction model based on topic association to predict the propagation trend of the user to the topic; and controlling the topic propagation trend according to the propagation trend of the user to the topic.

[0052] Topic information can be obtained from publicly available data websites or through mature social network public APIs. This includes the historical behavior information of all participating users during the lifecycle of the original and derivative topics, topic participation information, and user basic attribute information. The user's historical behavior information includes information such as the user's historical forwarding and commenting. The topic participation information includes information such as the time when the original and derivative topics were forwarded and commented on. The user's basic attribute information includes the friend relationship information of the participating users.

[0053] The topic information is preprocessed. Specifically, the topic information is cleaned to make most of the unstructured data structured, so that outliers or null values ​​no longer appear, reducing the inconvenience to subsequent calculations.

[0054] The process of the topic propagation prediction model based on topic association in processing preprocessed topic information includes the following:

[0055] S1: Extract the internal and external attributes of topic information; internal attributes include user activity, user topic awareness rate, and user historical forwarding rate, while external attributes include user friend influence, topic popularity, and user topic content set.

[0056] User activity :

[0057] User activity level is the degree to which a user's forwarding behavior is more active compared to other users; users who forward more content within a certain period are more actively forwarding. This invention defines a user... The activity level is:

[0058]

[0059] in, This refers to the period before the emergence of derivative topics. Internal users The number of original posts This refers to the period before the emergence of derivative topics. Internal users The number of reposts and posts; since the prediction is the probability of users reposting a topic, it is defined as... This represents the weakening coefficient.

[0060] User topic awareness :

[0061] User topic awareness rate reflects the probability of a user encountering a new topic, and is represented by the number of topics a user follows. To a certain extent, it reflects the probability of a user participating when a new topic emerges. This invention defines user topic awareness rate as:

[0062]

[0063] wherein, represents the number of attention of the user , represents the average number of attention of all users in the social network.

[0064] User historical forwarding rate :

[0065] User historical forwarding rate reflects the user's tendency to forward to some extent. The main source of users to obtain the topic is the information of friends. Therefore, the user historical forwarding rate is defined as:

[0066]

[0067] wherein, represents the number of historical forwarding microblog of the user, represents the total number of microblog forwarded by the user from the friends, the friends are the users followed by the user.

[0068] User friend driving force :

[0069] In the topic network, users are usually affected by the propagation behavior of the attention user to participate in a topic, the higher the interaction frequency between users, the higher the driving force between users, and the higher the probability of mutual forwarding topics. Different friends of the user have different driving forces, which are constructed as:

[0070]

[0071] wherein, represents the driving force of the user to the friend of the user , represents the average number of original microblog forwarded by the user to the friend user, represents the number of friends of the user, if or the user is not the friend of the user , then .

[0072] Topic heat :

[0073] The topic heat is embodied in the topic forwarding, commenting, liking and so on in the social network. The topic heat will rise rapidly in a short time, but when its heat reaches the top, it will fall rapidly. Considering that this process is similar to the half-life of elements, the half-life function is introduced, and the topic heat is defined as:

[0074]

[0075] wherein, represents the topic heat of the current time t, and respectively represent the forwarding quantity of the topic as of the current time and the previous time, represents the time of initial topic generation, represents a regularization factor.

[0076] User topic content set :

[0077] During the propagation of the topic, due to the uniqueness of people, the ideas and comments of each user are different when facing the same topic, so the topic comments can reflect the topic attributes and characteristics, and the topic characteristics can also change with the propagation of the topic, and the topic content set is represented as:

[0078]

[0079] wherein, info(t) represents the comments published by users in the topic propagation space in a time period The original topic content set is represented as , and the derived topic content set is represented as .

[0080] S2: The DTR2vec algorithm (derived topic representation learning algorithm based on topic association) is adopted to select user interest feature keywords and user cognitive feature keywords from the user topic content set, and vector representation is performed on the selected keywords to obtain a user historical behavior feature vector.

[0081] The DTR2vec algorithm designed by the application firstly utilizes an LDA (latent Dirichlet allocation) topic identification model to construct original- derived topic association features and user features, then extracts user cognitive accumulation and interest degree according to the state transition of the user facing the original- derived topic, and finally utilizes representation learning to vectorize the low-dimensional vectors.

[0082] S21: The user topic content set includes an original topic content set, a derived topic content set and a user social content set; the three content sets are respectively input into an LDA topic identification model to obtain an original topic keyword set , a derived topic topic keyword set and a user content keyword set .

[0083] ​The LDA model is used to extract the correlation features of the original topic and the derivative topic. Specifically, the microblog content published and forwarded by the user is regarded as a paragraph, thereby forming content in the form of an article. The content is divided into an original topic content set, a derivative topic content set and a user social content set. The original topic content set, the derivative topic content set and the user social content set are regarded as an article respectively, and the LDA model is used to process the three content sets respectively. The optimal number of themes is obtained by using different theme numbers for multiple clustering experiments.

[0084] The original topic keyword set , the derivative topic theme keyword set and the user content keyword set are obtained through the LDA model processing.

[0085] S22: Calculate the correlation degree of the original topic keyword set and the derivative topic theme keyword set . Calculate the similarity of the original topic content set and the derivative topic content set. Put the two keywords with a correlation degree greater than the similarity into the topic correlation feature word set .

[0086] The derivative topic develops from the original topic, so there is a certain correlation between the original topic and the derivative topic. The BM25 (Best Match) algorithm is used to obtain the similarity score of the original topic content set and the derivative topic content set, which is used as the correlation degree threshold of the original-derivative topic. The calculation formula is:

[0087]

[0088] wherein, represents the similarity score of the original topic content set and the derivative topic content set, Q represents the original topic content set, d represents the derivative topic content set, represents the word weight, represents the i-th word in the set Q, represents the total number of words in the original topic content set Q, represents the correlation degree of the i-th word in the set Q and the set d.

[0089] The correlation degree between the original topic theme keyword set and the derivative topic theme keyword set is calculated by using the cosine calculation formula, and the calculation formula is:

[0090]

[0091] wherein, , , respectively represent the vectors of keywords key1, key2, and m represents the dimension of the keyword vector.

[0092] Two keywords with a correlation degree greater than a similarity degree are put into a topic correlation characteristic keyword set .

[0093] S23: According to the correlation degrees of the original topic keyword set , the derived topic theme keyword set , and the topic correlation characteristic keyword set , the interest weight and the cognitive weight of the user content keyword set are calculated.

[0094] Whether a user forwards a topic is closely related to the interest and the cognition of the user to the topic. Considering that the user will accumulate certain cognition to the original topic after a period of time after the appearance of the original topic and the interest in the common characteristics of the original- derived topic is weakened, the present application uses a cosine distance as a weight to select keywords, and the formula for calculating the interest weight and the cognitive weight is as follows:

[0095]

[0096]

[0097] wherein, represents the interest weight of the i-th word in the user content keyword set, represents the cognitive weight of the i-th word in the user content keyword set, represents the similarity between the i-th candidate keyword and the derived topic theme keyword set , represents the total number of words in the user content keyword set, represents the similarity between the i-th candidate keyword and the topic correlation characteristic keyword set , represents the similarity between the i-th candidate keyword and the original topic keyword set , represents the current time, represents the time when the initial topic is generated, represents a regularization factor.

[0098] S24: According to the interest weight and the cognitive weight, respectively, the top-k keywords are selected from the user content keyword set as the user interest feature keywords and the user cognitive feature keywords.

[0099] The selected keywords (user interest feature keywords and user cognitive feature keywords) are represented by vectors based on a Doc2vec algorithm, and a user historical behavior feature vector is output , is expressed as:

[0100]

[0101] Wherein K represents the number of users under the derived topic, represents the user interest feature vector and the user cognitive feature vector.

[0102] S3: According to all internal attributes and external attributes, the influence of the topic on the user is quantified by using evolutionary game theory, and a topic influence adjacency matrix is obtained.

[0103] The original topic and the derived topic interact with each other in the process of propagation, so when predicting the propagation behavior of the user, the influence of the interaction between the topics should be considered. The original- derived topic influence on the user is quantified by introducing evolutionary game theory, and the process is as follows:

[0104] S31: According to the internal attribute, the internal influence is calculated, and the external influence is calculated according to the user friend driving force and the topic heat.

[0105] As shown in Figure 2 , the internal influence is composed of internal attributes, and the external influence is composed of external attributes, and the calculation formula is as follows:

[0106]

[0107]

[0108] Wherein, represents the internal influence of the user , represents the external influence of the user on the user , the activity of the user , represents the historical forwarding rate of the user , represents the topic perception rate of the user .

[0109] S32: According to the internal influence and the external influence, the original topic influence and the derived topic influence are calculated by using a multiple linear regression algorithm.

[0110] Comprehensive internal influence factors and external factors, using multiple linear regression algorithm, construct the topic influence function of the original topic and the derived topic:

[0111]

[0112]

[0113] wherein, denotes the native topic influence of a user on a user , denotes the derivative topic influence of a user on a user , denotes the external influence of a user on the native topic of a user , denotes the external influence of a user on the derivative topic of a user , , , are the first, second and third partial regression coefficients respectively trained using a multiple linear regression algorithm.

[0114] S33: define two game strategies, calculate the first revenue according to the first game strategy and the native topic influence, and calculate the second revenue according to the second game strategy and the derivative topic influence.

[0115] According to the principle of game theory, the present application defines two game strategies: strategy 1: "forwarding native topic", strategy 2: "forwarding derivative topic". Let , respectively represent the proportion of forwarding native topics and derivative topics among the adjacent users of the target user, and the revenue functions of the two strategies are:

[0116]

[0117]

[0118] wherein, denotes the first revenue, i.e. the revenue obtained by the user forwarding the native topic, denotes the second revenue, i.e. the revenue obtained by the user forwarding the derivative topic.

[0119] S34: calculate the topic propagation behavior influence of the user in the native topic and the topic propagation behavior influence of the user in the derivative topic according to the first revenue and the second revenue.

[0120]

[0121]

[0122] wherein, , respectively represent the original and derived topic users after evolution game the topic propagation behavior influence of the user .

[0123] S35: Calculate the topic influence matrix according to the topic propagation behavior influence of the user in the original topic and the topic propagation behavior influence of the user in the derived topic.

[0124] Considering the competitiveness of the original and derived topics, the final original- derived topic influence adjacency matrix is obtained

[0125]

[0126] wherein, represents the topic propagation behavior influence between users, if , then .

[0127] S4: Extract the network structure features according to the topic influence adjacency matrix, and obtain the network structure feature vector of the user.

[0128] Considering the complexity of the social network structure, the SDNE (graph network embedding algorithm) method is used for network representation learning in the present application, and the original- derived topic influence adjacency matrix is directly used as the input, and the output is the network structure feature vector of the user, which is represented as:

[0129]

[0130] wherein, K is the number of network nodes, i.e. the number of users under the derived topic, is the social structure feature vector corresponding to the user.

[0131] S5: Input the user historical behavior feature vector and the network structure feature vector of the user into the DT-GCN model, and obtain the user topic propagation prediction result, including whether the user participates in the topic propagation and the type of topic participated by the user.

[0132] The prediction task target of the present application is to predict whether the potential user node will participate in the forwarding of the associated topic, to judge whether the user forwards or does not forward, if the user forwards, whether the user forwards the original topic or the derived topic, and then the three classification tasks can be converted; considering that directly splicing two types of features will cause the input of the model to be too long, such as Figure 3 ​As shown, the application designs a DT-GCN (graph convolutional neural network based on related topics) model, which adds a CNN (convolutional neural network) layer in front of a GCN (graph convolutional neural network) model; the processing process of the DT-GCN model on a user historical behavior feature vector and a user network structure feature vector includes: constructing a native topic and derivative topic propagation mixed network according to topic information, obtaining an adjacency matrix according to the connection information between nodes in the native topic and derivative topic propagation mixed network, inputting the user historical behavior feature vector and the user network structure feature vector into the CNN network for convolution to obtain a feature matrix, inputting the feature matrix and the adjacency matrix into the GCN network, adding a drop layer in the GCN network, and finally processing by using a softmax function, wherein the softmax function converts the graph convolution output into probability values of different nodes and different classifications to obtain a propagation prediction result of the user on the topic.

[0133] In the GCN network, the adjacency matrix is preprocessed According to the adjacency matrix, the following is calculated:

[0134]

[0135] Among them, The original adjacency matrix is represented as A, which is obtained according to the connection information between nodes in the native topic and derivative topic propagation mixed network, The degree matrix of the adjacency matrix is represented as D.

[0136] The expression of the DT-GCN model is as follows:

[0137]

[0138] Among them, The category probability output by the user is represented as y; The activation function is represented as f; The drop layer added to the model randomly samples the parameters according to a certain probability p, that is, the i-th layer and the j-th neuron in the network layer are sampled, and some values are discarded to prevent the model from overfitting; The activation function is represented as f, The preprocessed adjacency matrix is represented as A, The adjacency matrix is represented as A, The node feature attribute of each layer is represented as x, and for the input layer That is, the user historical behavior feature vector and the user network structure feature vector, The activation function is represented as f, The CNN network is represented as CNN, The weight matrix of the i-th layer network is represented as W.

[0139] The application discusses a three-classification prediction problem, so the model output is set to y. , and are defined as follows:

[0140]

[0141] wherein, represents the probability of a user forwarding the original topic, represents the probability of a user not participating in the topic, represents the probability of a user forwarding the derivative topic, if the corresponding , it is determined that the potential user will forward the original topic in the next time period; if , it is determined that the potential user will forward the derivative topic in the next time period; otherwise, the potential user will not participate in the hot topic forwarding in the next time period.

[0142] The topic propagation prediction method based on topic association designed by the application first starts from the association relationship between the characteristics of the original-derivative topics in different evolution stages, proposes a DTR2vec algorithm, represents the cognitive accumulation and interest degree of the user, so as to make the topic feature space low vectorization. Then, considering the antagonism and symbiosis of the original-derivative topics in the propagation process, the evolutionary game theory is introduced to form the network structure feature representation of the influence of the original-derivative topics between users; finally, the association and game relationship of the original-derivative topics in the propagation process are analyzed, and the topic propagation trend is predicted and analyzed by comprehensively considering the above information.

[0143] The application can predict the propagation trend of the topic by predicting the propagation trend of the topic, so as to predict the propagation trend of the topic, can be applied to user personalized recommendation and marketing, and is helpful to mine the forwarding preference of the user, accurately put the advertisement and formulate the propaganda plan. At the same time, the forwarding and propagation trend of the hot topic can be understood and monitored in advance, targeted measures can be taken to help the public opinion department to quickly take targeted measures to purify the network environment, and the topic of carrying forward positive energy can be guided according to circumstances, and the correct value orientation is established in the society.

[0144] The above examples further illustrate the purpose, technical solutions and advantages of the application. It should be understood that the above examples are only preferred embodiments of the application and are not intended to limit the application. Any modification, equivalent replacement, improvement, etc. made to the application within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A topic propagation prediction method based on topic association, characterized in that, The method comprises the following steps: acquiring topic information and preprocessing the topic information; inputting the preprocessed topic information into a topic propagation prediction model based on topic correlation to predict the propagation trend of the user on the topic; controlling the topic propagation trend according to the propagation trend of the user on the topic; the process of processing the preprocessed topic information based on the topic correlation includes: S1: extracting internal attributes and external attributes of the topic information; the internal attributes include user activity, user topic perception rate and user historical forwarding rate, and the external attributes include user friend driving force, topic heat and user topic content set; S2: selecting user interest feature keywords and user cognitive feature keywords from the user topic content set by using the DTR2vec algorithm, and performing vector representation on the selected keywords to obtain a user historical behavior feature vector; the process of selecting the user interest feature keywords and the user cognitive feature keywords from the user topic content set by using the DTR2vec algorithm includes: S21: the user topic content set includes a native topic content set, a derivative topic content set and a user social content set; the three content sets are respectively input into an LDA topic identification model to obtain a native topic keyword set , a derivative topic keyword set and a user content keyword set ; S22: Calculate the set of original topic keywords and related topic keyword collection The relevance is calculated; the similarity between the original topic content set and the derived topic content set is calculated, and two keywords with a relevance greater than a similarity are added to the topic-related feature word set. ; S23: calculating the interest weight and the cognition weight of the user content keyword set according to the original topic keyword set , the derived topic theme keyword set , and the topic association characteristic keyword set ;​ S24: selecting Top-k keywords as the user interest feature keywords and the user cognition feature keywords from the user content keyword set according to the interest weight and the cognition weight respectively S24: selecting Top-k keywords as the user interest feature keywords and the user cognition feature keywords from the user content keyword set according to the interest weight and the cognition weight respectively S3: quantifying the influence of the topic on the user according to all internal attributes and the user friend driving force and the topic heat in the external attributes by using evolutionary game theory to obtain a topic influence adjacency matrix; S4: extracting network structure features according to the topic influence adjacency matrix to obtain a network structure feature vector of the user; S5: inputting the user historical behavior feature vector and the network structure feature vector of the user into a DT-GCN model to obtain a user topic propagation prediction result, including whether the user participates in the topic propagation and the type of the topic participated by the user.

2. The topic propagation prediction method based on topic correlation according to claim 1, wherein the similarity between the original topic content set and the derived topic content set is calculated by using the BM25 algorithm, and the formula is: the process of quantifying the influence of the topic on the user by using the evolutionary game theory includes: ； wherein, represents a similarity score between the native topic content set and the derived topic content set, Q represents the native topic content set, d represents the derived topic content set, represents a word weight, represents the i-th word in the set Q, represents the total number of words in the native topic content set Q, represents the i-th word in the set Q and the set d. 3.The topic propagation prediction method based on topic relevance according to claim 1, characterized in that, Computing a set of user content keywords The formula for the interest weight and the awareness weight of the user content keyword set is: ； ； in, This represents the interest weight of the i-th word in the user content keyword set. This represents the cognitive weight of the i-th word in the user content keyword set. This represents the set of candidate keywords and related topic keywords for the i-th candidate keyword. Similarity between them This represents the set of feature words related to the topic for the i-th candidate keyword. Similarity between them This represents the total number of words in the user content keyword set. This represents the set of candidate keywords and original topic keywords for the i-th keyword. Similarity between them Indicates the current moment. Indicates the moment when the initial topic was generated. This represents the regularization factor.

4. The topic propagation prediction method based on topic relevance according to claim 1, characterized in that, S31: calculating internal influence according to the internal attributes, and calculating external influence according to the user friend driving force and the topic heat; S32: calculating original topic influence and derived topic influence by using a multiple linear regression algorithm according to the internal influence and the external influence; S33: defining two game strategies, calculating first income according to the first game strategy and the original topic influence, and calculating second income according to the second game strategy and the derived topic influence; S34: calculating topic propagation behavior influence of the user in the original topic and topic propagation behavior influence of the user in the derived topic according to the first income and the second income; S35: calculating the topic influence adjacency matrix according to the topic propagation behavior influence of the user in the original topic and the topic propagation behavior influence of the user in the derived topic. The formula for calculating the internal influence and the external influence is:

5. The topic propagation prediction method based on topic relevance according to claim 4, characterized in that, the formula for calculating the topic propagation behavior influence of the user in the original topic and the topic propagation behavior influence of the user in the derived topic is: ； ； wherein, represents the internal influence of a user , represents the external influence of a user on a user , represents the activity of a user , represents the historical retweet rate of a user , represents the topic awareness rate of a user , represents the friend driving force of a user on a user , represents the topic heat at the current time t.

6. The topic propagation prediction method based on topic relevance according to claim 4, characterized in that, the topic influence adjacency matrix is represented as: ； ； wherein, represents the influence of users in the original topic on users, represents the influence of users in the derived topic on users, represents the first revenue, represents the second revenue.​​ 7. The topic propagation prediction method based on topic relevance according to claim 4, characterized in that, ​ ； wherein, represents a topic influence adjacency matrix, , represents a topic propagation behavior influence between users, represents a user influence in a native topic, on a user, , represents a user influence in a derived topic, on a user, . 8.The topic propagation prediction method based on topic relevance according to claim 1, characterized in that, The processing procedure of the DT-GCN model on the user historical behavior feature vector and the network structure feature vector of the user includes: obtaining an adjacency matrix according to the connection information between nodes in the original topic and the derived topic propagation mixed network, inputting the user historical behavior feature vector and the network structure feature vector of the user into a CNN network for convolution to obtain a feature matrix; inputting the feature matrix and the adjacency matrix into a GCN network, adding a drop layer in the GCN network, and finally adopting a softmax function for processing to obtain a propagation prediction result of the user on the topic.

9. The topic propagation prediction method based on topic relevance according to claim 1, characterized in that, The expression of the DT-GCN model is: ； wherein, represents the category probability output by the user, represents a pre-processed adjacency matrix, represents an input layer, represents that a drop layer is added to randomly sample parameters according to a certain probability, represents a CNN network, is a weight matrix of the i-th layer network.

Images