Hot topic propagation prediction method based on implicit relationship and cascade length

Abstract

The invention belongs to the technical field of Internet application, and particularly relates to a hot topic propagation prediction method based on a recessive relationship and cascade length. Comprising the steps of obtaining topic data and preprocessing the topic data; extracting user related attributes from the pre-processed topic data, wherein the user related attributes comprise user own attributes, user similarity and user activeness; according to the related attributes of the users, mining potential relationships among the users by adopting an association rule algorithm, and constructing a user implicit relationship network; constructing a user attention relation network according to the attributes of the user, and combining the user implicit relation network and the user attention relation network to obtain a user full relation network; learning structural features of the users according to the user full-relation network to obtain a structural feature matrix of the topic network; inputting the structural feature matrix of the topic network and the user activeness into a trained topic propagation prediction model for processing to obtain a topic propagation prediction result; the accuracy of topic propagation prediction can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet applications, and in particular relates to a hot topic propagation prediction method based on implicit relations and cascade lengths. Background Art

[0002] With the rapid development of wireless communications and the Internet, as well as the miniaturization and popularity of mobile devices, online social networks have become an important channel for people to make friends and share information. Hot news and information will attract many users to participate in discussions, thus forming hot topics. Hot topics have the characteristics of rapid spread, short-term, and great influence. If hot topics are not supervised and controlled, they may continue to spread, leading to group division and confrontation, and affecting social stability. Therefore, it is particularly important to predict user behavior on social networks and supervise and control hot topics. Understanding the spread of hot topics is crucial to correctly guide online topics.

[0003] At the same time, accurate user behavior can enable the government to more quickly control the direction of topics. Information dissemination prediction is generally divided into three categories from the perspective of research methods. (1) Feature-based methods: extract some features within the observation time window, such as content features, user features, etc. Through future popularity, supervised information is directly used for learning. However, feature extraction has strong human subjective factors and is not universal. (2) Generative methods: mainly describe social phenomena in real networks as a series of events in a continuous time domain, and regard the spread and accumulation of information as the arrival process of a forwarding behavior. It mainly focuses on independently modeling the rate function of each message forwarding process. (3) Deep learning-based methods: inspired by the excellent performance of deep learning models in many fields, aim to automatically and deeply represent the message propagation structure, participating users, message content, and time sequence information of the propagation event in an end-to-end manner, and perform single factor or multi-factor fusion to predict the final scale of the information cascade. In particular, when modeling structural factors, the first step is to pass the propagation cascade graph through multiple random walks to obtain a set of propagation cascade paths. Each cascade path not only carries the information of the messenger, but also describes the flow path of the information flow.

[0004] However, most existing methods only consider the implicit relationship between users, and each propagation path modeled by random walk is independent of each other, ignoring the time series factor. In reality, due to factors such as user group belonging and identity, common interests, trust and influence transmission, there are many implicit relationships between users, and the decline of topic popularity exacerbates the uncertainty of information dissemination strategies. Therefore, how to improve the user relationship network and better integrate the time series factor into the neural network information dissemination prediction method has become the main research goal of scholars in recent years.

[0005] Based on the existing research on social network behavior prediction, we found that there are still some challenges in user behavior prediction:

[0006] 1. The hidden nature of the key user relationship network. In social networks, even if two users do not have a direct following relationship, due to factors such as user group belonging and identity, common interests, trust and influence transmission, the hidden relationship between users will affect the depth of topic dissemination.

[0007] 2. The factors that influence the leadership of opinion leaders are complex. Users often make decisions based on the opinions of opinion leaders, but the length of the communication cascade and the differences among users lead to different effects of leadership transmission. This complexity makes users face more choices and confusion when discussing topics and making decisions, increasing the uncertainty of decision-making.

[0008] 3. The decline in topic popularity has exacerbated the uncertainty of information dissemination strategies. The decline in topic popularity reflects the shift in public interest and the fragmentation of attention. This phenomenon makes information dissemination increasingly tend to be carried out within specific interest groups, thereby shortening the length of the dissemination cascade. Summary of the invention

[0009] In view of the shortcomings of the prior art, the present invention proposes a hot topic propagation prediction method based on implicit relationship and cascade length, which includes:

[0010] S1: Obtain topic data and preprocess it; extract user-related attributes from the preprocessed topic data, where the user-related attributes include user attributes, user similarity, and user activity;

[0011] S2: Based on user-related attributes, the association rule algorithm is used to mine the potential relationships between users and build a user implicit relationship network;

[0012] S3: Build a user attention relationship network based on the user's own attributes, merge the user's implicit relationship network and the user's attention relationship network to obtain the user's full relationship network;

[0013] S4: Learn the structural features of users based on the user full relationship network and obtain the structural feature matrix of the topic network;

[0014] S5: Input the structural feature matrix of the topic network and the user activity into the trained topic propagation prediction model for processing to obtain the topic propagation prediction result.

[0015] Preferably, the process of extracting user-related attributes from the preprocessed topic data includes:

[0016] Extracting user attributes is represented as:

[0017] Att(u i )={{history(u i )},{hab(u i )},{focus(u i )},{fans(u i )}}

[0018] Among them, {history(u i )} represents user u i Historical behavior; {hab(u i )} represents user u i interest tags, {focus(u i )} represents user u i 's attention information, {fans(u i )} represents user u i Fan information of

[0019] The extracted user similarity is expressed as:

[0020]

[0021] Among them, Similarity(u i ,u j ) represents user u i with u j similarity, α represents the first adjustment parameter, β represents the second adjustment parameter, Mutal(u i ,u j ) represents user u i with u j The degree of interaction, hab(u i ) represents user u i Collection of hobbies, hab(u j ) represents user u j is a set of hobbies, m represents the number of user behavior types, W k The interaction weight of the k-th user behavior, Num[interact k t(u i ,u j )] represents user u i with u j The total number of times the kth user behavior is performed, participate k (u i ) represents user u i The total number of times the kth user behavior is performed, participate k (u j ) represents u j The total number of times all users have the kth user behavior;

[0022] Extracting user activity is expressed as:

[0023]

[0024] Among them, UserActivity(u i ) k Represents user u i The activity of the k-th user behavior, UserActivity(u i ) represents user u i activity, n represents the total number of users, follow(u i ) k Represents user u i The total number of k-th behaviors is obtained from the published posts, and μ represents an adjustable hyperparameter.

[0025] Preferably, the process of using the association rule algorithm to mine potential relationships between users and construct a user implicit relationship network includes:

[0026] Construct transaction data sets under multiple user behaviors based on user-related attributes;

[0027] The association rule algorithm is used to mine the potential relationship between users and obtain the implicit relationship user item set;

[0028] Based on the transaction data set under multiple user behaviors, multiple initial supports of each user item set in the implicit relationship user item set are calculated; the multiple initial supports of each user item set are averaged to obtain the final support of each user item set;

[0029] Determine whether the final support of the user item set is greater than the preset support threshold. If so, calculate the confidence of the user item set. Otherwise, it means that there is no implicit relationship between the users in the user item set.

[0030] The confidence of the user item set and the similarity between users in the user item set are weighted summed to obtain the final confidence of the user item set;

[0031] Determine whether the final confidence of the user item set exceeds the preset confidence threshold. If so, it means that there is an implicit relationship between the users in the user item set. Otherwise, it means that there is no implicit relationship between the users in the user item set. Construct a user implicit relationship network based on the implicit relationship between users.

[0032] Furthermore, the formula for calculating the initial support of the user item set is:

[0033]

[0034] Among them, Support(P j ) kRepresents the user item set P under the kth user behavior j The initial support of Contain(P j ) k The transaction data set representing the kth user behavior contains the user item set P j The number of transactions, D k A transaction dataset representing the k-th user behavior.

[0035] Furthermore, the formula for calculating the confidence of the user item set is:

[0036]

[0037] Among them, Confidence(u i →u j ) represents the confidence of the user item set, Contain(u i ∪u j ) indicates that the transaction dataset contains user u i with u j The number of transactions, Contain(u i ) indicates that the transaction dataset contains user u i The number of transactions.

[0038] Preferably, the process of obtaining the structural feature matrix of the topic network includes:

[0039] The edge weights between users are calculated according to user similarity and propagation cascade length, and the walking strategy is defined according to the edge weights between users. According to the walking strategy, the CI2vec algorithm is used to learn the structural characteristics of users and obtain the structural characteristic matrix of the topic network.

[0040] Furthermore, the formula for calculating the edge weight between users is:

[0041] w(u i ,u i+1 )=Similarity(u i ,u i+1 )+ρ*legth

[0042] Among them, w(u i ,u i+1 ) represents user u i with u i+1 The edge weight between i ,u i+1 )User u i with u i+1 , ρ represents the propagation cascade length coefficient, and legth represents the propagation cascade length.

[0043] Preferably, the propagation prediction model processes the user's structural feature matrix and user activity, including: calculating the weight coefficient between users based on the user activity and the edge weight between users; inputting the weight coefficient between users and the user's structural feature matrix into a two-layer graph convolutional neural network including a Dropout layer for processing to obtain a topic propagation prediction result.

[0044] Furthermore, the formula for calculating the weight coefficient between users is:

[0045]

[0046] Among them, w i,j Represents user u i with u j The weight coefficient between i ,u j ) represents user u i with u i+1 The edge weight between t and t, σ represents the time decay factor; i , t j is the timestamp of the input time and output time of the propagation prediction model; τ represents the activity decay factor, UserActivity(u j ) represents user u j activity.

[0047] The beneficial effects of the present invention are as follows: This paper proposes a new user behavior prediction method to solve the problem of predicting user behavior in social networks after a period of time, and provide effective decision support for applications such as rumor evaluation and website recommendation. It mainly includes: 1) Constructing a user implicit relationship network through user historical behavior to improve the user network structure. 2) Representing user feature vectors through a deep walk algorithm combined with user-related attributes. 3) Considering that the topic heat decay increases the uncertainty of information dissemination strategies, a time and activity decay function is designed and combined with a graph neural network to predict user behavior. The present invention enhances the understanding and prediction capabilities of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a structural block diagram of the hot topic propagation prediction method based on implicit relationship and cascade length in the present invention;

[0049] Figure 2 A schematic diagram of constructing a user implicit relationship matrix based on the association rule mining method in the present invention;

[0050] Figure 3 It is a schematic diagram of user feature vector representation based on cascade length in the present invention;

[0051] Figure 4 Schematic diagram of the UTAD-GCN model of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The present invention proposes a hot topic propagation prediction method based on implicit relationship and cascade length, such as Figure 1 As shown, the method includes:

[0054] S1: Obtain topic data and preprocess it; extract user-related attributes from the preprocessed topic data, where the user-related attributes include user attributes, user similarity, and user activity.

[0055] To obtain topic data, you can directly collect data from the social network platform. What needs to be obtained here is the basic information of the topic in the life cycle of the topic activity, the subscription relationship between users, the historical behavior of users in various periods (such as likes, comments, forwarding), the user's own attribute set (age, hobbies, etc.); pre-process the topic data, including deleting duplicate data, cleaning invalid information, standardizing data format, etc. This improves the quality of the data, making it more suitable for subsequent data analysis and modeling work, and provides a reliable data foundation for subsequent network initialization and model building.

[0056] In social networks, users' participation in topics is affected by many factors, such as user interests, neighboring users' participation in activities, and various interactive information between users. Based on this, the main goal of the present invention is to predict user group behavior in the hot topic space based on given user-related attributes and user networks; therefore, the user-related attributes are extracted from the pre-processed topic data, specifically:

[0057] User attributes:

[0058] The basic attributes of users, such as user interests, user attention information, and fan information, have a certain impact on whether users participate in topic dissemination. The present invention defines the basic attributes of users as follows:

[0059] Att(u i )={{history(u i )},{hab(u i )},{focus(u i )},{fans(u i)}}

[0060] Among them, {history(u i )} represents user u i Historical behavior; {hab(u i )} represents user u i interest tags, {focus(u i )} represents user u i 's attention information, {fans(u i )} represents user u i Fan information of .

[0061] User similarity:

[0062] User similarity is a measure of the similarity between two users in terms of interests, behaviors, preferences, and attributes. User similarity plays an important role in discovering implicit relationships between users. Users with higher similarity are more likely to have implicit relationships. User similarity is defined as:

[0063]

[0064] Among them, Similarity(u i ,u j ) represents user u i with u j similarity, α represents the first adjustment parameter, β represents the second adjustment parameter, Mutal(u i ,u j ) represents user u i with u j The degree of interaction, hab(u i ) represents user u i Collection of hobbies, hab(u j ) represents user u j Collection of hobbies.

[0065] In social networks, the degree of interaction between users can effectively measure user relationships. The interactive behaviors between users are reflected in the user's likes, comments, forwarding and other behaviors. The more behaviors between users, the higher the user interaction. At the same time, different weight coefficients W are given to different behaviors. k ; Therefore, user interaction is defined as:

[0066]

[0067] Wherein, m represents the number of user behavior types, preferably, m=3 (like, comment, forward); W k The interaction weight of the k-th user behavior, Num[interact k t(ui ,u j )] represents user u i with u j The total number of times the kth user behavior is performed, participate k (u i ) represents user u i The total number of times the kth user behavior is performed, participate k (u j ) represents u j The total number of times all users have the kth user behavior.

[0068] User activity:

[0069] In a topic network, user activity can effectively measure the user's interest and participation in a specific topic. User activity affects the user's forwarding behavior to a certain extent. The more active a user is at a certain moment, the more likely the user is to participate in the topic at the next moment. User activity is defined as:

[0070]

[0071] Among them, UserActivity(u i ) k Represents user u i The activity of the k-th user behavior, UserActivity(u i ) represents user u i activity, n represents the total number of users, follow(u i ) k Represents user u i The total number of k-th behaviors is obtained from the published posts, and μ represents an adjustable hyperparameter.

[0072] S2: Based on user-related attributes, the association rule algorithm is used to mine the potential relationships between users and construct a user implicit relationship network.

[0073] Build transaction data sets under multiple user behaviors based on user-related attributes, specifically:

[0074] like Figure 2As shown in the figure, in view of the hidden nature of the user relationship network, association rule mining is used to discover the potential relationships between users from their historical behaviors. Association rule mining is a method of exploring potential associations and connections between users by analyzing user behavior data. This technology can not only reveal hidden connections between users, but also help understand the behavior patterns and interaction rules of user groups. In order to construct a transaction set in association rule mining. Through the user's historical behavior, the likes, comments, and reposts of each user for one month are obtained as transactions in the association rule. The transaction data set of association rule mining is represented as D; D k =

[0075] {T1,T2,......,T n}, where D k is the transaction data set under the kth user behavior, T n For transaction, transaction T n Represents user u n For example, user u1 likes users u7, u8, and u 12 ; then T1={u7,u8,u 12}.

[0076] The association rule algorithm is used to mine the potential relationship between users and obtain the implicit relationship user item set {P1,P2,...P j}, P j is the user set {u p ,u q}, each D k Under these conditions, we can get a set of implicit relation user items.

[0077] Based on the transaction data set under multiple user behaviors, multiple initial supports of each user item set in the implicit relationship user item set are calculated. The calculation formula is:

[0078]

[0079] Among them, Support(P j ) k Represents the user item set P under the kth user behavior j The initial support of Contain(P j ) k The transaction data set representing the kth user behavior contains the user item set P j The number of transactions, D k A transaction dataset representing the k-th user behavior.

[0080] The average of multiple initial supports of each user item set is calculated to obtain the final support of each user item set Support (P j). An item set with high support means that it is more common in the data set. It is determined whether the final support of the user item set is greater than the preset support threshold. If it is greater, the confidence of the user item set is calculated. Otherwise, it means that there is no implicit relationship between the users in the user item set.

[0081] Use confidence to filter out reliable item sets. Confidence indicates the probability that a transaction containing a certain item set also contains another item set. It is used to measure the reliability of the rule; the formula for calculating the confidence of the user item set is:

[0082]

[0083] Among them, Confidence(u i →u j ) represents the confidence of the user item set, Contain(u i ∪u j ) indicates that the transaction dataset contains user u i with u j The number of transactions, Contain(u i ) indicates that the transaction dataset contains user u i The number of transactions.

[0084] The confidence of the user item set and the similarity between users in the user item set are weighted summed to obtain the final confidence of the user item set.

[0085] Determine whether the final confidence of the user item set exceeds the preset confidence threshold. If so, it means that there is an implicit relationship between the users in the user item set. Otherwise, it means that there is no implicit relationship between the users in the user item set. Construct a user implicit relationship network based on the implicit relationship between users to obtain the user implicit relationship matrix.

[0086] S3: Construct a user-following relationship network based on the user's own attributes, merge the user's implicit relationship network with the user's follow-up relationship network, and obtain the user's full relationship network.

[0087] S4: Learn the structural characteristics of users based on the user's full relationship network and obtain the structural characteristic matrix of the topic network.

[0088] Considering that the influencing factors of user influence vary with the cascade length of the transmission and the differences among users, the inaccurate quantification of user influence will cause deviations in group behavior prediction. The present invention uses the CI2vec algorithm to extract the structural feature matrix. CI2vec is a multi-dimensional user node feature representation method based on Node2vec. This method realizes user feature representation by simulating the user's experience of hot topic transmission, and more richly vectorizes the low-rank feature space of hot topic. Figure 3 As shown, specifically:

[0089] Calculate the edge weights between users based on user similarity and propagation cascade length:

[0090] w(u i ,u i+1 )=Similarity(u i ,u i+1 )+ρ*legth

[0091] Among them, w(u i ,u i+1 ) represents user u i with u i+1 The edge weight between i ,u i+1 )User u i with u i+1 similarity, ρ represents the propagation cascade length coefficient, legth represents the propagation cascade length obtained by the breadth-first algorithm, which can be obtained by the breadth-first algorithm.

[0092] Define the walking strategy based on the edge weights between users:

[0093] Assume that in the random walk path, the maximum walk length is set to L, the root node user u0 is used as the starting point, and the sequence of randomly visited user vertices is u0,u1,u2...u i-1 ,u i ,u i+1 . Current user node u i The wandering transition probability can be expressed as:

[0094]

[0095] Among them, P(u i+1 |u i ) represents user node u i To user node u i+1 The transition probability, w(u i ,u i+1 ) represents user node u i With node u i+1 The edge weight between p,q (u i-1 ,u i+1 ) is the walk bias term, and z is the normalization coefficient.

[0096]

[0097] Among them, p is the probability of returning to the previous node; q reflects the walking strategy; Represents user node u i-1 and node u i+1The shortest distance between two nodes is {0,1,2}.

[0098] According to the walking strategy, the CI2vec algorithm is used to learn the structural features of users and obtain the structural feature matrix of the topic network. The optimization objective function of the CI2vec method is:

[0099]

[0100] Where Ns(u) is the set of neighboring nodes of vertex u, f(u) represents the mapping function of the embedding representation vector of user node u, and P(Ns(u)|f(u)) represents the probability of the appearance of the neighboring nodes of vertex u.

[0101] Through the node feature mapping function f(u) at different times, the structural feature matrix M of different learning periods t can be obtained t The output of the CI2vec algorithm is expressed as:

[0102] M t =[m1,m2,m3,...,m k ]

[0103] Among them, M t The eigenvector matrix representing the topic network at time t; m k Represents the final feature vector representation of the k-th user node.

[0104] S5: Input the structural feature matrix of the topic network and the user activity into the trained topic propagation prediction model for processing to obtain the topic propagation prediction result.

[0105] The participation behavior of users in topic propagation activities is not only related to their neighboring users, but also to internal factors of users and topics. The decline of topic popularity exacerbates the uncertainty of information dissemination strategies. The decline of topic popularity reflects the migration of public interests and the fragmentation of attention. This phenomenon makes information dissemination increasingly tend to be carried out within specific interest groups, thereby reducing the length of the dissemination cascade. In order to solve this problem, the present invention proposes a topic propagation prediction model called UATD-GCN by designing time and user activity decay functions and fusing them. It uses graph convolutional neural networks as deep learning models, which can capture the relationships and structural characteristics between nodes in graph data.

[0106] Preferably, Figure 4 As shown in Figure 1, the topic propagation prediction model is a two-layer graph convolutional neural network with a dropout layer. The prediction of user group behavior is defined as a binary classification problem, that is, through the network-wide feature vector of the hot topic in time period t, predict whether the potential group users in time period t+1 will participate in the spread of the hot topic.

[0107] Feature matrix Mt =N×F, where N represents the number of user nodes in the hot topic propagation network, and F represents the input feature dimension of each node. The adjacency matrix G of the user at time t t = {U t ,E t}, represents the user full relationship network under the hot topic in time period t, where U t represents the set of user nodes in the propagation space at time t, E t Represents the set of edges between nodes in the propagation space.

[0108] Randomly initialize the weights W and bias items. In the first layer of GCN, first convert the feature matrix M t The edge weights are adjusted by the time decay function and the activity decay function, that is, the weight coefficients between users are calculated based on the user activity and the edge weights between users. The calculation formula is:

[0109]

[0110] Among them, w i,j Represents user u i with u j The weight coefficient between i ,u j ) represents user u i with u i+1 The edge weight between t and t, σ represents the time decay factor; i , t j is the timestamp of the input time and output time of the propagation prediction model; τ represents the activity decay factor, UserActivity(u j ) represents user u j activity.

[0111] The weight coefficients between users and the structural feature matrix of users are input into a two-layer graph convolutional neural network containing a Dropout layer for processing to obtain the topic propagation prediction result, which is expressed as:

[0112]

[0113] Among them, W l is the weight matrix corresponding to the lth layer network in the graph convolutional network, Add self-loop I to the adjacency matrix G. The Sigmod function maps the probability value of user behavior:

[0114]

[0115] Since the Sigmod function can compress the output value into the range of [0,1], if the output value Z is greater than 0.5, it is considered that the user participates in the topic propagation in the current period, and if it is less than 0.5, it is considered that the user does not participate.

[0116] If the experimental data set size is N, Y i The true sample label of the i-th user annotated in the dataset. is the output result of the user behavior prediction model. During training, the loss function L of the model is defined as:

[0117]

[0118] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hot topic propagation prediction method based on implicit relationship and cascade length, characterized in that: include: S1: Obtain topic data and preprocess it; extract user-related attributes from the preprocessed topic data, where the user-related attributes include user attributes, user similarity, and user activity; S2: Based on user-related attributes, the association rule algorithm is used to mine the potential relationships between users and build a user implicit relationship network; S3: Build a user attention relationship network based on the user's own attributes, merge the user's implicit relationship network and the user's attention relationship network to obtain the user's full relationship network; S4: Learn the structural features of users based on the user full relationship network and obtain the structural feature matrix of the topic network; S5: Input the structural feature matrix of the topic network and the user activity into the trained topic propagation prediction model for processing to obtain the topic propagation prediction result.

2. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 1, characterized in that: The process of extracting user-related attributes from preprocessed topic data includes: Extracting user attributes is represented as: Att(u i )={{history(u i )},,{hab(u i )},{focus(u i )},{fans(u i )}} Among them, {history(u i )} represents user u i Historical behavior; {hab(u i )} represents user u i interest tags, {focus(u i )} represents user u i 's attention information, {fans(u i )} represents user u i Fan information of The extracted user similarity is expressed as: Among them, Similarity(u i ,u j ) represents user u i with u j similarity, α represents the first adjustment parameter, β represents the second adjustment parameter, Mutal(u i ,u j ) represents user u i with u j The degree of interaction, hab(u i ) represents user u i Collection of hobbies, hab(u j ) represents user u j is a set of hobbies, m represents the number of user behavior types, W k The interaction weight of the k-th user behavior, Num[interact k t(u i ,u j )] represents user u i with u j The total number of times the kth user behavior is performed, participate k (u i ) represents user u i The total number of times the kth user behavior is performed, participate k (u j ) represents u j The total number of times all users have the kth user behavior; Extracting user activity is expressed as: Among them, UserActivity(u i ) k Represents user u i The activity of the k-th user behavior, UserActivity(u i ) represents user u i activity, n represents the total number of users, follow(u i ) k Represents user u i The total number of k-th behaviors is obtained from the published posts, and μ represents an adjustable hyperparameter.

3. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 1, characterized in that: The process of using the association rule algorithm to mine the potential relationships between users and construct the user implicit relationship network includes: Construct transaction data sets under multiple user behaviors based on user-related attributes; The association rule algorithm is used to mine the potential relationship between users and obtain the implicit relationship user item set; Based on the transaction data set under multiple user behaviors, multiple initial supports of each user item set in the implicit relationship user item set are calculated; the multiple initial supports of each user item set are averaged to obtain the final support of each user item set; Determine whether the final support of the user item set is greater than the preset support threshold. If so, calculate the confidence of the user item set. Otherwise, it means that there is no implicit relationship between the users in the user item set. The confidence of the user item set and the similarity between users in the user item set are weighted summed to obtain the final confidence of the user item set; Determine whether the final confidence of the user item set exceeds the preset confidence threshold. If so, it means that there is an implicit relationship between the users in the user item set. Otherwise, it means that there is no implicit relationship between the users in the user item set. Construct a user implicit relationship network based on the implicit relationship between users.

4. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 3 is characterized in that: The formula for calculating the initial support of the user item set is: Among them, Support(P j ) k Represents the user item set P under the kth user behavior j The initial support of Contain(P j ) k The transaction data set representing the kth user behavior contains the user item set P j The number of transactions, D k A transaction dataset representing the k-th user behavior.

5. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 3, characterized in that: The formula for calculating the confidence of a user item set is: Among them, Confidence(u i →u j ) represents the confidence of the user item set, Contain(u i ∪u j ) indicates that the transaction dataset contains user u i with u j The number of transactions, Contain(u i ) indicates that the transaction dataset contains user u i The number of transactions.

6. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 1, characterized in that: The process of obtaining the structural feature matrix of the topic network includes: The edge weights between users are calculated according to user similarity and propagation cascade length, and the walking strategy is defined according to the edge weights between users. According to the walking strategy, the CI2vec algorithm is used to learn the structural characteristics of users and obtain the structural characteristic matrix of the topic network.

7. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 6, characterized in that: The formula for calculating the edge weight between users is: w(u i ,u i+1 )=Similarity(u i ,u i+1 )+ρ*legth Among them, w(u i ,u i+1 ) represents user u i with u i+1 The edge weight between i ,u i+1 )User u i with u i+1 , ρ represents the propagation cascade length coefficient, and legth represents the propagation cascade length.

8. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 1, characterized in that: The propagation prediction model processes the user's structural feature matrix and user activity, including: calculating the weight coefficient between users based on user activity and the edge weight between users; inputting the weight coefficient between users and the user's structural feature matrix into a two-layer graph convolutional neural network containing a Dropout layer for processing to obtain the topic propagation prediction result.

9. The method for predicting the spread of hot topics based on implicit relationships and cascade length according to claim 8, characterized in that: The formula for calculating the weight coefficient between users is: Among them, w i,j Represents user u i with u j The weight coefficient between i ,u j ) represents user u i with u i+1 The edge weight between t and t, σ represents the time decay factor; i , t j Timestamps for input and output times of the propagation prediction model; τ represents the activity decay factor, UserActivity(u j ) represents user u j activity.

Images