Key influence user discovery method for time-varying network

Abstract

The invention belongs to the field of social network derived topic analysis, and particularly relates to a time-varying network-oriented key influence user discovery method, which comprises the following steps of: preprocessing original topic network data to obtain topic network data; slicing the topic network data along the time dimension to obtain topic network data at different time points; calculating dynamic importance of the user nodes at different time points, and obtaining a transfer sequence of each user node according to the dynamic importance; constructing a feature vector of each user node by adopting an attention mechanism; obtaining an accumulated interaction sequence and a recent interaction sequence of each user node, and obtaining an interest vector of each user node through a dual Transform structure; splicing the feature vector and the interest vector of each user node to obtain a comprehensive feature; inputting all the comprehensive features into a full connection layer and a softmax function layer to obtain a key influence user identification result; according to the method, the key influence elements in the topic space can be effectively found.

Description

Technical Field

[0001] The present invention belongs to the field of social network derived topic analysis, and in particular relates to a method for discovering key influential users for time-varying networks. Background Art

[0002] With the rapid development of social networks, people widely participate in the discussion and dissemination of various topics. Derivative topics spread rapidly in a short period of time, forming public opinion hotspots with great social impact. In this process, some key users use platform mechanisms and their own influence to promote the dissemination of topics, trying to guide the direction of public opinion, influence public thinking, and even achieve specific goals. Therefore, timely discovering the key influential users in the early stage of these topic outbreaks is of great significance for understanding the mechanism of topic dissemination, guiding the direction of public opinion, and effectively implementing public opinion management. In recent years, scholars have found that identifying and tracking core users with guiding capabilities has become an important topic in the fields of influence evaluation, public opinion governance, and target group positioning. Identifying potential core users in the guiding topic dissemination network can not only help relevant institutions grasp the public opinion dynamics more accurately, but also effectively prevent and respond to the spread of potential negative information. Therefore, scholars believe that discovering and managing core influential users in guiding topics has far-reaching significance for promoting a healthy network ecological environment.

[0003] At present, the research on the discovery of key influential users in the early stage of the outbreak of derivative topics in social networks mainly focuses on three aspects: social network topic content analysis, network structure analysis and node influence quantification. Research on topic content analysis mainly includes the mining of sentiment tendencies and content characteristics. Research on network structure analysis mainly focuses on the group structure and interaction patterns in social networks, revealing the information propagation path and core propagation groups in the network. In terms of node influence quantification, researchers evaluate the influence of nodes by quantifying indicators such as node centrality and predict their key role in communication.

[0004] In general, although the academic community has produced a lot of research results on hot topics, it still faces the following challenges in discovering key influential users in the early stages of the outbreak of derivative topics:

[0005] 1. Dynamic complexity of topic space network structure. The interaction relationship between users and topic structure in social networks changes over time. How to effectively mine and quantify these multi-dimensional features and represent them as node vectors is a challenge.

[0006] 2. The fluidity of topic content. Topic content changes over time, and it is difficult to directly analyze and model it. How to effectively capture the complex connections and similarities between topics and combine the time dimension to conduct in-depth analysis of user behavior is a difficult point. Summary of the invention

[0007] To solve the above problems, the present invention provides a method for discovering key influential users in a time-varying network, comprising the following steps:

[0008] S1. Obtaining original topic network data from the social platform and preprocessing it to obtain topic network data;

[0009] S2. Slice the topic network data along the time dimension to obtain the topic network data at different time points;

[0010] S3. Calculate the dynamic importance of the user node for the topic network data at each time point, calculate the transition probability of the random walk based on the dynamic importance, and obtain the transition sequence of each user node;

[0011] S4. Based on the transfer sequence at each time point, the attention mechanism is used to construct the feature vector of each user node;

[0012] S5. Obtain the cumulative interaction sequence and the most recent interaction sequence of each user node according to the topic network data;

[0013] S6. Based on the accumulated interaction sequence and the most recent interaction sequence, the interest vector of each user node is obtained through a dual Transformer structure;

[0014] S7. For each user node, concatenate the feature vector and interest vector to obtain a comprehensive feature; input the comprehensive features of all user nodes into the fully connected layer and the softmax function layer to obtain the key influential user identification results.

[0015] Beneficial effects of the present invention:

[0016] Starting from the dynamic characteristics of the topic network, the present invention analyzes the user behavior under the time dimension, combines the hidden features of users under different time scales, and obtains the final discrimination result of the topic network influential users by integrating with the comprehensive interest features of users. Compared with the prior art, the present invention takes into account the time sensitivity and historical influence factors, and combines the comprehensive interests of users to consider the comprehensive impact of user behavior from the time dimension level, thereby improving the discrimination accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention;

[0018] Figure 2 A schematic diagram of obtaining a user node feature vector according to the present invention. DETAILED DESCRIPTION

[0019] 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.

[0020] The present invention provides a method for discovering key influential users in time-varying networks. Figure 1 , Figure 2 As shown, the following steps are included:

[0021] S1. Obtain original topic network data from the social platform and preprocess it to obtain topic network data.

[0022] Specifically, we obtain relevant original topic network data through social media APIs and public data set platforms, including user basic information data, user historical behavior data, and relevant data during the life cycle of topics. Among them, user basic information data includes personal information such as user ID and gender, and user historical behavior data includes interactive behavior data between users and other users or topics, including likes, forwarding, etc.

[0023] The original topic network data is preprocessed, including removing some invalid data and duplicate data, and simple cleaning to transform the unstructured raw data into structured data.

[0024] S2. Slice the topic network data along the time dimension to obtain the topic network data at different time points.

[0025] S3. For the topic network data at each time point, calculate the dynamic importance of the user node, calculate the transition probability of the random walk based on the dynamic importance, and obtain the transition sequence of each user node.

[0026] Specifically, the present invention takes a user as a user node. The user relationship network topology constructed according to the topic network data at different time points is different, and the importance of the user node changes with time. The present invention defines user node i as n The dynamic importance of W i (t n ), expressed as

[0027] W i (t n )=βS i (T)+(1-β)H i (t n )

[0028] Among them, β∈[0,1] represents the balance coefficient, which is used to balance the weight of short-term time sensitivity and cumulative historical influence in the dynamic importance calculation. T represents a fixed-length time period, which is used to count the interaction behavior of user nodes within this time range; specifically, the user node i is calculated at the nth time point t n The dynamic importance of W i (t n ), statistics in the time period [t n ,t n +T] the interaction behavior of user node i. i (T) is the time sensitivity, which indicates the relative connection frequency of user node i in the time window T. i (t n ) is the cumulative historical influence, indicating the influence of user node i at the current time point t n The long-term influence of is combined with the interaction behaviors at the past m time points;

[0029]

[0030] I i (T) represents the interaction frequency between user node i and other user nodes in the time window T, Represents the total interaction frequency of all users in the time window T. α is the attenuation factor, which is used to control the impact of historical behavior on the importance of current dynamics. The larger the time interval, the smaller the weight. i (t k ) represents user node i at time point t k The frequency of interaction with other user nodes.

[0031] By adjusting β and α, the importance of user nodes in the dynamic network can be better reflected.

[0032] Specifically, the transition probability of random walk is calculated based on dynamic importance and is expressed as

[0033]

[0034] Among them, P(j,i) represents the transition probability from user node i to user node j, ω(i,j) represents the edge weight between user node i and user node j, indicating the degree of correlation between the two nodes. E represents the set of relationship edges between users participating in the topic. W j represents the dynamic importance of user node j. Z is the normalization coefficient. b p,q (i,j) represents the walk bias term, which is based on the shortest distance d from user node i to user node j. i,j The possible values ​​are [0, 1, 2], which respectively represent returning to the previous node, the direct neighbor, and the far node. That is, p and q are used to control the walking direction, including:

[0035]

[0036] S4. Based on the transfer sequence at each time point, the attention mechanism is used to construct the feature vector of each user node.

[0037] Specifically, step S4 uses the attention mechanism to construct a vector representation of each user node based on the transfer sequence at each time point, such as Figure 2 As shown, including:

[0038] S41. Perform embedding processing on the transfer sequence of each user node at each time point to obtain corresponding structural features;

[0039] S42. Perform position encoding on each time point to obtain a time position embedding vector;

[0040] S43. According to the structural features and time position embedding vectors, the input vectors of each user node at different time points are obtained, which are expressed as

[0041] x i (t n )=concat(h i (t n ),PE(t n ))

[0042] Among them, concat(·) represents the vector concatenation operation, x i (t n ) represents user node i at time point t n The input vector, h i (t n ) represents user node i at time point t n The structural characteristics of PE(t n ) represents the time point t n The temporal position embedding vector of

[0043] S44. Calculate the weight vector of each user node at different time points according to the input vector, expressed as

[0044]

[0045]

[0046] in, Represents user node i at time point t n The input vector of user node i at time point t k The attention weight of the input vector, Q i (t n ) represents the time point t according to the user node in The query vector K is calculated from the input vector i (t k ) represents the time point t according to the user node i k The key vector calculated from the input vector, d k represents the dimension of the key vector; V i (t k ) represents the time point t according to the user node i k The value vector calculated from the input vector, N represents the total number of time points, z i (t n ) represents user node i at time point t n The weight vector of ;

[0047] S45. Input the weighted vector of each user node at all time points into the feedforward neural network to obtain the feature vector of the user node.

[0048] S5. Obtain the cumulative interaction sequence and the most recent interaction sequence of each user node according to the topic network data.

[0049] Specifically, the cumulative interaction sequence and the recent interaction sequence are the interaction behavior data between the user node and other user nodes in different time periods, such as likes, forwarding, comments, etc.

[0050] S6. Based on the accumulated interaction sequence and the most recent interaction sequence, the interest vector of each user node is obtained through a dual Transformer structure.

[0051] Specifically, the dual Transformer structure includes a shallow Transformer and a deep Transformer.

[0052] Specifically, the shallow Transformer is used to capture the main points of interest in the user's recent interactions. The shallow network has high computational efficiency and can quickly extract the main features of short-term behaviors. 2 to 4-layer encoders can better balance the model complexity and the needs of short-term interest capture. The shallow Transformer in the embodiment of the present invention uses a 2-layer encoder. The deep Transformer is used to perform in-depth feature extraction on the user's historical accumulated interactions; more layers enable the model to capture complex long-term behavior patterns and global semantic features; for this reason, the deep Transformer in the embodiment of the present invention uses a 6-layer encoder. At the same time, in order to enable the shallow Transformer and the deep Transformer to work together, the two share some parameters.

[0053] The process of obtaining the interest vector of any user node through the dual Transformer structure in step S6 includes:

[0054] S61. Input the accumulated interaction sequence into the deep Transformer and output the user's long-term preference, which can be expressed as

[0055]

[0056] in, represents the long-term preference of user node i, represents the cumulative interaction sequence of user node i, Θ deep represents the deep Transformer parameters, T deep Represents a deep Transformer.

[0057] S62. Input the recent interaction sequence into the shallow Transformer and output the user's short-term preference, which can be expressed as

[0058]

[0059] in, represents the short-term preference of user node i, represents the most recent interaction sequence of user node i, Θ shallow Represents the shallow Transformer parameters, T shallow Represents a shallow Transformer.

[0060] S63. The user's long-term preference and the user's short-term preference are combined to obtain the interest vector, which can be expressed as

[0061]

[0062] Among them, pi represents the interest vector of user node i, and γ represents a weight parameter for balancing long-term and short-term interests. This combination balances the weights of short-term interests and long-term preferences, ensuring that the user's interest vector can dynamically reflect recent changes and reflect long-term stability.

[0063] S7. For each user node, concatenate the feature vector and interest vector to obtain a comprehensive feature; input the comprehensive features of all user nodes into the fully connected layer and the softmax function layer to obtain the key influential user identification results.

[0064] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0065] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for discovering key influential users in time-varying networks, characterized in that: The following steps are involved: S1. Obtaining original topic network data from the social platform and preprocessing it to obtain topic network data; S2. Slice the topic network data along the time dimension to obtain the topic network data at different time points; S3. Calculate the dynamic importance of the user node for the topic network data at each time point, calculate the transition probability of the random walk based on the dynamic importance, and obtain the transition sequence of each user node; S4. Based on the transfer sequence at each time point, the attention mechanism is used to construct the feature vector of each user node; S5. Obtain the cumulative interaction sequence and the most recent interaction sequence of each user node according to the topic network data; S6. Based on the accumulated interaction sequence and the most recent interaction sequence, the interest vector of each user node is obtained through the dual Transformer structure; S7. For each user node, concatenate the feature vector and interest vector to obtain a comprehensive feature; input the comprehensive features of all user nodes into the fully connected layer and the softmax function layer to obtain the key influential user identification results.

2. According to claim 1, a method for discovering key influential users in time-varying networks is characterized in that: User node i at time point n n The dynamic importance of W i (t n ), expressed as W i (t n )=βS i (T)+(1-β)H i (t n ) Among them, β∈[0,1] represents the balance coefficient; T represents a fixed-length time period; S i (T) represents the relative connection frequency of user node i in time window T; H i (t n ) represents the user node i at the current time point t n The long-term impact of I i (T) represents the interaction frequency between user node i and other user nodes in the time window T, α is the decay factor; I i (t k ) represents user node i at time point t k The frequency of interaction with other user nodes.

3. According to claim 1, a method for discovering key influential users in time-varying networks is characterized in that: The transition probability of random walk is calculated based on dynamic importance and is expressed as Where P(j,i) represents the transition probability from user node i to user node j, ω(i,j) represents the edge weight between user node i and user node j, and E represents the set of relationship edges between users participating in the topic; W j represents the dynamic importance of user node j, Z is the normalization coefficient; b p,q (i,j) represents the walk bias term expressed as: Among them, d i,j It represents the shortest distance from user node i to user node j, and p and q are the factors controlling the wandering bias.

4. According to claim 1, a method for discovering key influential users in time-varying networks is characterized in that: Step S4 uses an attention mechanism to construct a vector representation of each user node based on the transfer sequence at each time point, including: S41. Perform embedding processing on the transfer sequence of each user node at each time point to obtain corresponding structural features; S42. Perform position encoding on each time point to obtain a time position embedding vector; S43. According to the structural features and time position embedding vectors, the input vectors of each user node at different time points are obtained, which are expressed as x i (t n )=concat(h i (t n ),PE(t n )) Among them, concat(·) represents the vector concatenation operation, x i (t n ) represents user node i at time point t n The input vector, h i (t n ) represents user node i at time point t n The structural characteristics of PE(t n ) represents the time point t n The temporal position embedding vector of S44. Calculate the weight vector of each user node at different time points according to the input vector; S45. Input the weighted vector of each user node at all time points into the feedforward neural network to obtain the feature vector of the user node.

5. According to claim 4, a method for discovering key influential users in a time-varying network is characterized in that: Step S44 calculates the weighted vector of each user node at different time points according to the input vector, expressed as in, Represents user node i at time point t n The input vector of user node i at time point t k The attention weight of the input vector, Q i (t n ) represents the time point t according to the user node i n The query vector K is calculated from the input vector i (t k ) represents the time point t according to the user node i k The key vector calculated from the input vector, d k represents the dimension of the key vector; V i (t k ) represents the time point t according to the user node i k The value vector calculated from the input vector, N represents the total number of time points, z i (t n ) represents user node i at time point t n The weight vector of .

6. The method for discovering key influential users in a time-varying network according to claim 1, characterized in that: The dual Transformer structure includes a shallow Transformer and a deep Transformer, the shallow Transformer uses a 2-layer encoder, and the deep Transformer uses a 6-layer encoder.

7. The method for discovering key influential users in a time-varying network according to claim 6, characterized in that: The process of obtaining the interest vector of any user node through the dual Transformer structure in step S6 includes: S61. Input the accumulated interaction sequence into the deep Transformer and output the user's long-term preference; S62. Input the recent interaction sequence into the shallow Transformer and output the user's short-term preference; S63. The user's long-term preference and the user's short-term preference are integrated to obtain an interest vector.

Images