A method for predicting influential users in a topic network based on depth propagation and breadth propagation

Abstract

The present invention belongs to the field of social network analysis, and particularly relates to a method for predicting topic network influential users based on deep propagation and breadth propagation. The method includes: obtaining topic network data and performing preprocessing; calculating user intimacy and user credibility according to the preprocessed topic network data; optimizing the DSU2vec algorithm according to user intimacy and user credibility; using the optimized DSU2vec algorithm to extract hidden information of the topic network to obtain a deep propagation feature vector matrix of the topic network; performing community division on the topic network to obtain the divided communities; using a graph convolutional neural network to extract features of community nodes to obtain a breadth propagation feature vector matrix of the topic network; using a multi-dimensional propagation network prediction model to process the deep propagation feature vector matrix and the breadth propagation feature vector matrix to obtain a prediction result of topic network influential users. The prediction result of the present invention has high accuracy and good application prospects.

Description

Technical Field

[0001] The present invention belongs to the field of social network analysis, and particularly relates to a method for predicting topic network influential users based on depth propagation and breadth propagation. Background Art

[0002] In recent years, with the rapid development of the Internet, social networks have been continuously expanding, and the communication between people has become increasingly frequent. A variety of social platforms have become powerful tools for people to communicate and are closely related to people's lives. Currently, the mainstream social platforms include Weibo, Twitter, and Facebook, among which the largest social platform in China is Sina Weibo. Social networks are composed of large and small topic networks. A topic network is filled with a vast amount of users and information, and these complex and dense data often have great value for analyzing user influence. Therefore, how to obtain information crucial for analyzing influence from the complex data has become the focus of topic networks.

[0003] At the same time, analyzing the influence of key users plays an important role in aspects such as public opinion control, advertising placement, and disaster control. The purpose of predicting user influence is actually to select opinion leaders to prepare for maximizing influence. Accurately predicting the most influential users is the most crucial step in the process of topic diffusion. In this context, researching the influence of topic network users is meaningful. In traditional influence research, generally only the topic propagation direction in a single dimension is considered, and the overall propagation trend of the entire topic is not considered jointly from the two propagation dimensions of depth and breadth. Therefore, it is necessary to analyze topic user behavior from multiple dimensions for better researching user influence.

[0004] The analysis of user behavior data is a key issue in studying influence. At present, a large number of scholars have studied user influence. However, in most of the current influence research models, the consideration of different dimensions of the topic space is ignored, resulting in unconvincing conclusions. For example, Han et al. (Han M, Yan M, Cai Z, et al. An exploration of broader influence maximization in timeliness networks with opportunistic selection[J]. Journal of Network and Computer Applications, 2016, 63: 39 - 49.) proposed an influence maximization model based on time delay effect and breadth diffusion. This model considered the change of topic propagation decay over time, but it only regarded breadth as an influencing factor, could not deeply mine the hidden information of users in the breadth propagation network, and did not analyze the differences between different user behaviors. It regarded the edges of different behaviors as the same type of edges, thus affecting the final prediction accuracy. Summary of the Invention

[0005] Based on the existing research on user influence in the comprehensive topic network, it is found that there are still some challenges in predicting user influence:

[0006] 1. The topic propagation space is multi - dimensional. The propagation mode of the topic network is not single. It not only has a chain - like propagation mode but also has a star - shaped diffusion propagation mode. The topic space cannot be analyzed from a single propagation dimension. It is necessary to comprehensively analyze the depth and breadth of the topic propagation to represent and analyze the nodes more comprehensively.

[0007] 2. The topic network structure is complex. The cascade length of information propagation represents the propagation depth of the topic. Different cascade lengths make the propagation space structure more complex and diverse. How to mine the potential relationships between users from the complex structure is a difficult point.

[0008] 3. The uncertainty of the topic propagation scale. The number of communities of information propagation reflects the propagation breadth of the topic. The more the number of communities, the wider the influence range of the topic. If the factor of topic breadth can be considered, the accuracy of user influence prediction can be improved.

[0009] Aiming at the deficiencies of the existing technology, the present invention proposes a method for predicting users' influence in a topic network based on depth and breadth propagation. The method includes:

[0010] S1: Obtain topic network data and pre - process the topic network data;

[0011] S2: Calculate the user intimacy and user credibility based on the preprocessed topic network data;

[0012] S3: Define the random walk strategy of the DSU2vec algorithm according to the user intimacy and user credibility to optimize the DSU2vec algorithm;

[0013] S4: Use the optimized DSU2vec algorithm to extract the hidden information of the topic network and obtain the deep propagation feature vector matrix of the topic network;

[0014] S5: Perform community division on the topic network to obtain the divided communities;

[0015] S6: Use a graph convolutional neural network to extract the features of community nodes and obtain the wide propagation feature vector matrix of the topic network;

[0016] S7: Use a multi-dimensional propagation network prediction model to process the deep propagation feature vector matrix and the wide propagation feature vector matrix to obtain the prediction result of the influential users in the topic network.

[0017] Preferably, the formula for calculating user intimacy is:

[0018]

[0019] where Int(u i , u j ) represents the intimacy between the i-th user u i and the j-th user u j , X i represents the interaction weight of the i-th interaction method, Num[Interact i (u i , u j )] represents the total number of times between the i-th user u i and the j-th user u j under the i-th interaction method, Num[Interact i u i represents the total number of times between the i-th user u i and all users in the network under the i-th interaction method, Num[Interact i u j represents the total number of times between the j-th user u j and all users in the network under the i-th interaction method.

[0020] Preferably, the formula for calculating user credibility is:

[0021] Cre(u i ) = α·Num[Interact(u i)]+β·Num[Interacted(u i )]

[0022] Among them, Cre(u i ) represents the credibility of the i-th user u i , α represents the first attenuation coefficient, Num[Interact(u i )] represents the total number of interactions of the i-th user with the messages posted by their friends, β represents the second attenuation coefficient, and Num[Interacted(u i )] represents the total number of interactions of the messages posted by the i-th user u i by their friends.

[0023] Preferably, the random walk strategy of the DSU2vec algorithm is as follows:

[0024]

[0025] w(u i ,u j ) = Int(u i ,u j ) + ε

[0026] Among them, P(u j |u i ) represents the transition probability from user node u i to user node u j , w(u i ,u j ) represents the edge weight from user node u i to user node u j , Cre(u i ) represents the credibility of the i-th user u i , z represents the scaling factor, Int(u i ,u j ) represents the intimacy between the i-th user u i and the j-th user u j , and ε represents the propagation depth coefficient.

[0027] Preferably, the process of community partitioning of the topic network includes: calculating the edge similarity of user nodes according to the topic network data; fusing the two edges with the highest similarity according to the edge similarity to form a community; calculating the partitioning density value of the community; continuously fusing the two with the highest similarity until the partitioning density value is the largest, stop fusing, and obtain the partitioned community.

[0028] Furthermore, the formula for calculating the edge similarity of user nodes is:

[0029]

[0030] Among them, represents the edge e ik and edge e jk of the edge similarity, represents the edge e ik and edge e jk whether the edge types are the same, Common represents the intersection number of the neighbor nodes of the user node u i and the user node u j Number represents the union number of the neighbor nodes of the user node u i and the user node u j and the user node u.

[0031] Preferably, the process of extracting the features of community nodes by using a graph convolutional neural network includes:

[0032] Obtain the user feature vectors of the users in each community according to the community, and obtain the user feature matrix according to the user feature vectors;

[0033] Obtain the neighbor matrix and degree matrix of the community according to the community; input the user feature matrix, neighbor matrix and degree matrix of each community into the graph convolutional neural network respectively to obtain the breadth propagation feature vector matrix of the topic network.

[0034] Preferably, the process of processing the depth propagation feature vector matrix and the breadth propagation feature vector matrix by using a multi-dimensional propagation network prediction model includes:

[0035] Concatenate the depth propagation feature vector matrix and the breadth propagation feature vector matrix to obtain a concatenated matrix;

[0036] Process the concatenated matrix by using an attention mechanism to obtain an attention distribution matrix;

[0037] Process the attention distribution matrix by using two fully connected layers to obtain the prediction result of the influential users of the topic network.

[0038] The beneficial effects of the present invention are as follows: Starting from the depth and breadth dimensions of topic propagation, the present invention analyzes the user behaviors in different dimensions, extracts the user hidden features in the two dimensions respectively, and through fusing the features in the two dimensions, obtains the prediction result of the influential users of the final topic network; compared with the prior art, the present invention considers the discovery of user hidden information in multiple dimensions, excavates the user hidden information from the depth and breadth dimensions, and considers the different impacts brought by different user behaviors in both dimensions, improving the accuracy of the prediction result. The relevant public opinion departments can use the prediction result to widely spread or suppress the spread of the topic. Description of the Drawings

[0039] Figure 1Flowchart of the topic network influence user prediction method based on depth propagation and breadth propagation in the present invention;

[0040] Figure 2 Schematic diagram of the process of obtaining the depth propagation feature vector matrix in the present invention

[0041] Figure 3 Schematic diagram of the process of obtaining the breadth propagation feature vector matrix in the present invention. Detailed implementation manners

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] The present invention proposes a topic network influence user prediction method based on depth propagation and breadth propagation. As Figure 1 shown, the method includes the following contents:

[0044] S1: Obtain topic network data and preprocess the topic network data.

[0045] Online obtain topic network data. There are mainly two ways to obtain topic network data. One is to search for a suitable data set on the websites of domestic and foreign public data sets and download and use it. The other is to obtain topic network data by using the public API interface provided by the social platform. The topic network data mainly includes the user basic information data under the topic network and the interaction behavior data during the topic life cycle; the user basic information data includes the user id, and the interaction behavior data includes the follow-up relationship, reply relationship, mention relationship, and forwarding relationship formed by the user with other users.

[0046] Preprocess the topic network data: The directly obtained original data is usually unstructured data and cannot be directly used for data analysis. It is necessary to structure the unstructured data through simple data cleaning. For example, perform operations such as deleting null values and duplicate data on the original topic network data to obtain the preprocessed topic network data.

[0047] S2: Calculate the user intimacy and user credibility according to the preprocessed topic network data.

[0048] In the topic network, it includes all user nodes participating in the topic propagation, and the relationship edges formed by users in the topic propagation, where the relationships include follow-up, mention, forwarding, and reply; extract the relevant attributes of the user according to the preprocessed topic network data, and the relevant attributes include the user intimacy and user credibility;

[0049] In a topic network, the intimacy between users can effectively measure the user relationship. User intimacy behaviors are manifested in the attention, forwarding, replying, and mentioning behaviors between users. Generally, the more frequent the interaction between users, the higher the user intimacy. The formula for calculating user intimacy is as follows:

[0050]

[0051] where Int(u i ,u j ) represents the intimacy between the i-th user u i and the j-th user u j ; X i represents the interaction weight of the i-th interaction method. Different interaction relationships have different interaction weights, and the closer the relationship, the higher the interaction weight. Preferably, in the present invention, the interaction weights of mentioning, forwarding, replying, and attention are 0.4, 0.2, 0.3, and 0.1 respectively; Num[Interact i (u i ,u j )] represents the total number of times of the i-th user u i and the j-th user u j under the i-th interaction method, Num[Interact i u i represents the total number of times of the i-th user u i and all users in the network under the i-th interaction method, and Num[Interact i u j represents the total number of times of the j-th user u j and all users in the network under the i-th interaction method.

[0052] In a topic network, there are malicious users such as zombie fans and water army, so it is necessary to identify users. At the same time, the more credible a user is, the greater the role in topic dissemination and the more likely to become an influential user; the formula for calculating user credibility is:

[0053] Cre(u i ) = α·Num[Interact(u i )] + β·Num[Interacted(u i )]

[0054] where Cre(u i ) represents the credibility of the i-th user u i , Num[Interact(u i)] represents the total number of interactions of the i-th user with the messages posted by his friends; α represents the first attenuation coefficient, and β represents the second attenuation coefficient. According to experience, α, β∈(0,1), preferably, α and β are 0.6 and 0.8 respectively; the smaller the values ​​of α and β, the more likely the user u i The smaller the credibility is, the smaller the role is in the topic dissemination process; on the contrary, the greater the role is; Num[Interacted(u i )] represents the i-th user u i The total number of interactions that friends have received on the published messages. Interactions include being followed, replied to, mentioned, forwarded, and other behavioral interactions.

[0055] S3: Based on user intimacy and user credibility, a random walk strategy of the DSU2vec algorithm is defined to optimize the DSU2vec algorithm.

[0056] For in-depth topic dissemination, the interaction between users tends to be more vertical, forming a chain of dissemination. This network is usually generated by the interaction between relatives and friends. Figure 2 As shown, the present invention proposes a DSU2vec (deep space user vector representation) method for node relationships in deep propagation networks. With user intimacy and propagation depth as the driving force, the nodes are vectorized and embedded into a low-dimensional dense vector space to mine the hidden relationships between nodes.

[0057] Calculate the edge weights between two adjacent user nodes using the following formula:

[0058] w(u i ,u j )=Int(u i ,u j )+ε

[0059]

[0060] Where ε represents the propagation depth coefficient, d i Represents user node u i Continue to propagate the topic to the depth reached. The objective function optimized by the DSU2vec method is:

[0061]

[0062] in, is the mapping function of the vertex u embedding vector, N s (u) is the set of neighboring vertices sampled by vertex u through the formulated walking strategy S. Represents the probability of the neighboring nodes of vertex u appearing.

[0063] To achieve the above goals, the present invention redefines the random walk strategy of the DSU2vec algorithm to optimize the DSU2vec algorithm. The random walk strategy of the DSU2vec algorithm is as follows:

[0064]

[0065] where P(u j |u i ) represents the transition probability from user node u i to user node u j , w(u i ,u j ) represents the edge weight from user node u i to user node u j , z represents the scaling factor, taking the maximum value of w(u i ,u j )Cre(u i ), and the transition probability is scaled to the interval (0, 1).

[0066] S4: Use the optimized DSU2vec algorithm to extract the hidden information of the topic network, and obtain the deep propagation feature vector matrix of the topic network.

[0067] Use the DSU2vec method to extract the hidden information of the topic network from the dimension of propagation depth, vectorize the nodes, and embed them into a low-dimensional dense vector space. The output of the DSU2vec algorithm is expressed as:

[0068] N = [n 1 ,n 2 ,n 3 ,...,n K

[0069] where N represents the deep propagation feature vector matrix of the topic network, and n k represents the vector representation of the k-th user node.

[0070] Since the depth of topic propagation does not necessarily mean the wide spread of information, next, starting from the breadth dimension of topic propagation, the influence of users within the breadth range is studied. As Figure 3 shown, first, the topic network is divided into different community structures to obtain basic attributes such as the structure, quantity, and size of the community where the nodes are located. Then, according to the node attributes and the community network structure, the graph convolutional neural network GCN is used to perform feature representation on the user nodes in different communities.

[0071] S5: Perform community division on the topic network to obtain the divided communities.

[0072] ​Since the breadth and dissemination scope of a topic are closely related to the community information where the user is located, the present invention divides communities for the topic network and believes that multiple communities existing in the social network overlap and are related to each other. There may be some special nodes that are closely connected to multiple communities, and these nodes should belong to multiple communities. The MB-Link partitioning algorithm is proposed, and the specific process is as follows:

[0073] The MB-Link algorithm is an overlapping community detection algorithm that divides communities by aggregating edges. The basic idea of this algorithm is: calculate the similarity between each adjacent edge of the overlapping nodes, sort them according to the similarity size, and merge the edges with high similarity to obtain a community structure based on edge partitioning.

[0074] The present invention defines the similarity of edges as being associated with different interaction behaviors of users, and assigns different weights to different behavioral relationships. For example, the user mention relationship obviously indicates a closer connection between the two than the user follow relationship, so the weight is higher; calculate the edge similarity of user nodes according to the topic network data, and the formula for calculating the edge similarity of user nodes is:

[0075]

[0076] Wherein, represents the edge similarity between the edge e ik and the edge e jk ; represents whether the edge types of the edge e ik and the edge e jk are the same. If they are different, then is 0. If they are the same, then is determined by the weights of four different edge types: follow, reply, forward, and mention. Preferably, the weights of follow, reply, forward, and mention are taken as 0.1, 0.3, 0.2, and 0.4 respectively; Common represents the number of intersection of neighbor nodes of the user node u i and the user node u j , and is defined as:

[0077] Common = n + (i) ∩ n + (j)

[0078] Wherein, n + (i) represents all neighbor nodes of the user node u i , n + (j) represents all neighbor nodes of the user node u j , and the user node u i and the user node u j are two nodes that are not shared with the edge e ik and the edge e jk .

[0079] Number represents the user node u i and the union number of the neighbor nodes of the user node u j is defined as:

[0080] Number = n + (i) ∪ n + (j)

[0081] Fuse the two edges with the highest similarity according to the edge similarity to form a community; use the partition density as the criterion for evaluating the partition quality. When the partition density is the largest, the partition quality is the best. Assume that the number of edges in the topic network is M and the number of nodes is K. Define C = {C 1 , C 2 ,..., C e} as a community partition structure of the entire network. The formula for calculating the partition density is:

[0082]

[0083] where m c represents the number of edges in the c-th community C c , and k c represents the number of user nodes in the c-th community C c .

[0084] Continuously fuse the two with the highest similarity until the partition density value is the largest, then stop fusing to obtain the partitioned community.

[0085] S6: Use a graph convolutional neural network to extract the features of community nodes to obtain the breadth propagation feature vector matrix of the topic network;

[0086] The diffusion range of the breadth propagation network is closely related to the community where the user is located. Obtain the user feature vector of each user in the community according to the community. The present invention uses the number of communities where the node is located, the size of the community where the node is located, and the number of connections between nodes as the user's own attributes. The single user feature vector is expressed as:

[0087] S a = {Num community , Num size , Num degree}

[0088] Obtain the user feature matrix according to the user feature vector, which is expressed as: X = k c × S a .

[0089] The present invention uses a two-layer GCN model to represent the features of community nodes, fully considering the interaction between the node's own attributes and the network topology structure where it is located, and realizes the mining of hidden information of nodes with different community structures. Obtain the neighbor matrix and degree matrix of the community according to the community; input the user feature matrix X, neighbor matrix A, and degree matrix D of each community into the GCN respectively to obtain the breadth propagation feature vector matrix of the topic network. The specific process includes the following content:

[0090] The output of each layer of the GCN model is expressed as:

[0091]

[0092] Among them, x takes 0 or 1, H(0) = X, that is, the user feature matrix, and σ represents a non-linear activation function; W (x) represents the weight matrix of the (x + 1)-th layer, and the output of the previous layer is the input of the next layer; preferably, when σ is the ReLu activation function in the present invention, the output result of the first layer is:

[0093]

[0094]

[0095] Among them, represents the averaging process of the degree matrix, is the adjacency matrix after normalization, which balances the influence degree between nodes, I is a diagonal matrix.

[0096] The above processing is performed for each community. For the vector representation of nodes that are simultaneously in multiple communities, the vector representation of these nodes is obtained by taking the average value. The final output of the GCN model, that is, the vector representation of the nodes, is combined to obtain the breadth propagation feature vector matrix W = [w 1 , w 2 , w 3 ,..., w K .

[0097] S7: Use the multi-dimensional propagation network prediction model to process the depth propagation feature vector matrix and the breadth propagation feature vector matrix to obtain the prediction result of the influential users in the topic network.

[0098] The present invention uses the attention mechanism to fuse the node vector representations of the two propagation direction networks, and proposes a multi-dimensional propagation network prediction model. The final user influence prediction is defined as a binary classification problem, that is, whether the user has influence.

[0099] The process of using the multi-dimensional propagation network prediction model to process the depth propagation feature vector matrix and the breadth propagation feature vector matrix includes:

[0100] Concatenate the deep propagation feature vector matrix and the breadth propagation feature vector matrix to obtain a concatenated matrix;

[0101] First, concatenate the node features of deep propagation, i.e., the deep propagation feature vector matrix N = [n 1 , n 2 , n 3 ,..., n K and the node features of breadth propagation, i.e., the breadth propagation feature vector matrix W = [w 1 , w 2 , w 3 ,..., w K , that is

[0102] Considering that the attention mechanism can automatically weigh the importance of each input and reduce the impact of a single propagation path on user prediction, the attention mechanism is used to fuse the node features of multiple propagation dimensions; the attention mechanism is used to process the concatenated matrix P obtained after concatenation to obtain an attention distribution matrix R = [r 1 , r 2 , r 3 ,..., r K , where the calculation formula of r i , i ∈ [1, K] is:

[0103]

[0104]

[0105] Among them, q represents the query vector, which is determined by the input information, Q represents the input dimension, and K represents the maximum index of the input information.

[0106] Use two fully connected layers to process the attention distribution matrix to obtain the prediction result of the topic network influence on users; specifically:

[0107] Use two fully connected layers to perform dimensionality reduction on the result of the attention mechanism, and further synthesize vector information, and use the sigmoid function to obtain the final binary classification result. The output corresponding to each user node in the second layer is a one-dimensional vector, and the one-dimensional vector has two values, 0 and 1. 0 indicates no influence, and 1 indicates influence. The output result can be expressed as:

[0108] Y = d K

[0109] The result of Y has two categories. Y = 1 indicates that the user is an influential user, and Y = 0 indicates that the user is a non-influential user.

[0110] The probability value mapped by the sigmoid function for user influence, that is, the probability that a user is an influential user, is as follows:

[0111]

[0112] Among them, B represents the user embedding output by the fully connected layer.

[0113] Through the prediction results of user influence, users with great influence on topic dissemination can be screened out. In the public opinion part, the overall dissemination trend of the topic can be controlled by encouraging or restricting such users, so as to achieve the accelerated dissemination of the target topic or the rapid suppression of the target topic.

[0114] The above-mentioned embodiments have further elaborated on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above-mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting influential users in a topic network based on depth propagation and breadth propagation, characterized in that, it includes: S1: Obtain topic network data and preprocess the topic network data; S2: Calculate user intimacy and user credibility according to the preprocessed topic network data; S3: Define a random walk strategy for the DSU2vec algorithm according to user intimacy and user credibility to optimize the DSU2vec algorithm; The random walk strategy of the DSU2vec algorithm is: w(u i ,u j ) = Int(u i ,u j ) + ε Among them, P(u j |u i ) represents the transition probability from user node u i to user node u j , w(u i , u j ) represents the edge weight from user node u i to user node u j , Cre(u i ) represents the credibility of the i-th user u i , z represents the scaling factor, Int(u i , u j ) represents the intimacy between the i-th user u i and the j-th user u j , and ε represents the propagation depth coefficient; S4: Use the optimized DSU2vec algorithm to extract the hidden information of the topic network to obtain the depth propagation feature vector matrix of the topic network; S5: Perform community division on the topic network to obtain the divided communities; The process of performing community division on the topic network includes: calculating the edge similarity of user nodes according to the topic network data; fusing the two edges with the highest similarity according to the edge similarity to form a community; calculating the division density value of the community; continuously fusing the two with the highest similarity until the division density value is the largest, stop fusing, and obtain the divided communities; S6: Use a graph convolutional neural network to extract the features of community nodes to obtain the breadth propagation feature vector matrix of the topic network; The process of using a graph convolutional neural network to extract the features of community nodes includes: Obtain the user feature vectors of users in each community according to the community, and obtain the user feature matrix according to the user feature vectors; Obtain the neighbor matrix and degree matrix of the community according to the community; Input the user feature matrix, neighbor matrix, and degree matrix of each community into the graph convolutional neural network respectively to obtain the breadth propagation feature vector matrix of the topic network; S7: Use a multi-dimensional propagation network prediction model to process the depth propagation feature vector matrix and the breadth propagation feature vector matrix to obtain the prediction result of influential users in the topic network; The process of using a multi-dimensional propagation network prediction model to process the depth propagation feature vector matrix and the breadth propagation feature vector matrix includes: Concatenate the depth propagation feature vector matrix and the breadth propagation feature vector matrix to obtain a concatenated matrix; Use an attention mechanism to process the concatenated matrix to obtain an attention distribution matrix; Use two fully connected layers to process the attention distribution matrix to obtain the prediction result of influential users in the topic network.

2. The method for predicting influential users in a topic network based on depth propagation and breadth propagation according to claim 1, characterized in that, The formula for calculating user intimacy is: Among them, Int(u i , u j ) represents the intimacy between the i-th user u i and the j-th user u j . X i represents the interaction weight of the i-th interaction method. Num[Interact i (u i , u j )] represents the total number of times the i-th user u i and the j-th user u j have interacted in the i-th interaction method. Num[Interact i u i represents the total number of times the i-th user u i has interacted with all users in the i-th interaction method. Num[Interact i u j represents the total number of times the j-th user u j has interacted with all users in the i-th interaction method.

3. The method for predicting influential users in a topic network based on depth propagation and breadth propagation according to claim 1, characterized in that, The formula for calculating user credibility is: Cre(u i ) = α·Num[Interact(u i )] + β·Num[Interacted(u i )] Among them, Cre(u i ) represents the credibility of the i-th user u i , α represents the first attenuation coefficient, Num[Interact(u i )] represents the total number of interactions of the i-th user with the messages posted by their friends, β represents the second attenuation coefficient, Num[Interacted(u i )] represents the total number of interactions of the messages posted by the i-th user u i by their friends.

4. The method for predicting influential users in a topic network based on depth propagation and breadth propagation according to claim 1, characterized in that, The formula for calculating the edge similarity of user nodes is: Among them, represents the edge similarity between edge e ik and edge e jk ; represents whether the edge types of edge e ik and edge e jk are the same. Common represents the number of intersection nodes between user node u i and user node u j . Number represents the number of union nodes between user node u i and user node u j .

Images