Method for generating an interaction network based on users and user behavior

Abstract

The application discloses a kind of interactive network generation methods based on user and user behavior, first is with social interaction behavior connection weight e i→j =[l,c,r] as the directed connection edge between users, a social graph FIG=(V,e j→i ) based on social user attributes and interaction behavior is constructed;Then considering the actual problem that it is difficult to obtain the privacy attribute information of each user in user set USER and it is difficult to obtain sufficient data in the early propagation process of hot event, each user is mapped from the interaction space FIG=(V,e j→i ) to latent variable space FIG_BH using graph variational autoencoder, the attribute information of user is enhanced using user behavior information containing likes, comments and forwards, and the smoothness of social information features is improved by the method of enhancing the correlation of multi-source information.

Description

Technical Field

[0001] This invention relates to social network data mining technology, and more particularly, to a method for generating interactive networks based on users and user behavior. Background Technology

[0002] Compared to the "point-to-multi" information dissemination method of traditional news media, the "point-to-point" dissemination method of social network news facilitates the rapid spread of information through user interaction. While it is possible to gather public opinions and needs from interactive information, this dissemination method struggles to constrain the spread of misinformation through simple review mechanisms. Most existing public opinion monitoring algorithms focus on extracting features of fake news from massive amounts of content and reviewing online comments one by one. These algorithms face the following challenges in information collection and feature integration:

[0003] (1) Due to the massive scale of network information and the adversarial behavior of malicious users, the algorithm not only fails to meet the speed requirements due to the excessive number of monitoring tasks, but also fails to meet the monitoring accuracy requirements due to the lack of sufficient high-quality datasets.

[0004] (2) The computational algorithm may fail to complete the multi-step reasoning process due to the weak correlation between information and the lack of common sense knowledge.

[0005] (3) Social networks are subject to various types of concept drift, which may completely change the monitoring results of the model.

[0006] Given the current lack of comprehensive research on public opinion monitoring based on a four-pronged approach encompassing user attributes, user behavior, text content, and social networks, it is impossible to conduct accurate data analysis and extract key technologies for social networks used in public opinion monitoring. Summary of the Invention

[0007] In order to target key users in existing social networks ( This not only accelerates the spread of trending events but also influences user perceptions of these events. In monitoring public opinion surrounding trending events, quickly and accurately identifying the aforementioned... This invention plays a positive role in guiding public opinion. It proposes a method for generating interactive networks based on users and user behavior. In this invention, the focus user... A key characteristic is that individual attributes and interactive behaviors have significant influence. Therefore, in order to accurately identify the focal users during trending events... The interactive network generation method of the present invention first uses social interaction behavior to connect weights. As directed edges connecting users, a social graph based on social user attributes and interaction behaviors is constructed. Then, considering the difficulty in obtaining the user set... Addressing the challenges of obtaining sufficient data on various user privacy attributes and the difficulty in acquiring data during the early stages of trending events, a graph variational autoencoder is used to extract data from each user's privacy attributes. Interaction space mapped to latent variable space In this study, user behavior information (including likes, comments, and reposts) is used to enhance user attribute information, thereby improving the smoothness of social information features by enhancing the correlation between multiple information sources.

[0008] The present invention provides a method for generating interactive networks based on users and user behavior, which is applied to registered users on social networking sites; and includes the following steps:

[0009] Step 1: Collect user and social behavior information from any trending event on any social networking site;

[0010] Step one recorded user sets on social networking sites based on trending events. And social behavior information between any two users The connection weight of social interaction behavior between any two users is denoted as ,user The person who posted the message.

[0011] Step two: Construct a social network based on the connection weights of social interaction behaviors among users;

[0012] The social network constructed in step two This not only formally represents the ontological attributes of social media users but also integrates representations of user interaction behaviors. However, both user attributes and interaction information on social networking sites exhibit sparsity, leading to a semantic gap in the reasoning process of deep learning.

[0013] Step 3: Spatial mapping using a graph variational autoencoder;

[0014] Step 301, set up a single-layer latent variable model;

[0015] Step 302: Set up the encoding and decoding network model;

[0016] Step 303, processing of the encoding network model;

[0017] Step 304, Decoding the network model;

[0018] In step three, the graph variational encoder used introduces an encoder network model. and decoder network model Ensuring the potential space is continuous and smooth further enables it to maintain a certain degree of robustness in the sparse space to adapt to more domain data spaces.

[0019] Step four, identification of focus users;

[0020] Using methods to measure the influence of key users in rebuilding social networks Focus users The identification.

[0021] The advantages of the interactive network generation method based on users and user behavior in this invention are:

[0022] ① This invention uses connection weights based on social interaction behavior in generating social networks. As directed connection edges between users, users in social networking sites are used as user nodes in the social network.

[0023] ②This invention utilizes a graph variational autoencoder to separate each user from the... Interaction space mapped to latent variable space In this study, user behavior information (including likes, comments, and reposts) is used to enhance user attribute information, thereby improving the smoothness of social information features by enhancing the correlation between multiple information sources.

[0024] ③This invention is about rebuilding social networks Focus users in the process Identification is based on the fact that only when the external information received by participating users is sufficiently close to their own internal beliefs will they be likely to abandon their own viewpoints and follow the focus user. That's the viewpoint. Attached Figure Description

[0025] Figure 1 This is a diagram of the network structure of user social behavior interaction.

[0026] Figure 2 This is a flowchart of the interactive network generation process based on users and user behavior in this invention.

[0027] Figure 3 This is a comparison chart of training time and test accuracy for relation prediction.

[0028] Figure 4 This is a comparison chart of AUC and mAP performance for graph models with different numbers of layers. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0030] See Figure 1 , Figure 2 As shown, the present invention provides a method for generating an interactive network based on users and user behavior, which is applied to registered users on social networking sites; it includes the following steps:

[0031] Step 1: Collect user and social behavior information from any trending event on any social networking site;

[0032] In this invention, the user is referred to as The first user is recorded as The second user is recorded as , No. Each user is recorded as , No. Each user is recorded as , No. Each user is recorded as A total of [data] was collected from social networking sites. A user set is represented as a set. subscript Also known as the total number of users.

[0033] In this invention, user-to-user social behavior information includes "likes" (…). ),Comment( ) and / or forwarding ( The behavior of any two users. Social behavior information between any two users is recorded as follows: The connection weights of social interaction behaviors are denoted as .user The person who posted the message.

[0034] In this invention, hot events are denoted as... The first trending event is denoted as The second hot topic is denoted as , No. Each hot event is recorded as , No. Each hot event is recorded as , No. Each hot event is recorded as .

[0035] For example Figure 1 In the middle, users The posts included: the first trending event. , No. Hot topics and the Hot topics ;and and By user Likes ( ),and By user Comments ( );and By user Likes ( ); Explanation of user With users and users Social interaction behaviors existed. Specifically:

[0036] (1) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending event involved 2 likes, 1 comment, and 0 shares.

[0037] (2) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending event involved 1 like, 0 comments, and 0 shares.

[0038] For example Figure 1 In the middle, users The posts included: the first trending event. And the second hot topic ;and By user Comments ( );and and By user Comments ( ); Explanation of user With users and users Social interaction behaviors existed. Specifically:

[0039] (1) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending events involved had 0 likes, 1 comment, and 0 shares.

[0040] (2) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending events involved had 0 likes, 2 comments, and 0 shares.

[0041] For example Figure 1 In the middle, users The posts included: the first trending event. The second hot topic and the Hot topics ;and By user Comments ( );and By user Forwarded ( );and By user Likes ( ); Explanation of user With users and users Social interaction behaviors existed. Specifically:

[0042] (1) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending events involved had 0 likes, 1 comment, and 0 shares.

[0043] (2) User With users Social behavior information is recorded as The connection weights of social interaction behaviors are denoted as The Indicate user For users The trending event involved one like, zero comments, and one share.

[0044] For example Figure 1 In the middle, users Users who did not post are not considered as subjects of study in this invention patent. However, users Participating users and users Social behavior, therefore only applicable to users and users Recording social behaviors can reduce the generation of redundant social behavior information on social networking sites.

[0045] Step one recorded user sets on social networking sites based on trending events. subscript Also known as the total number of users; and social behavior information between any two users. The connection weight of social interaction behavior between any two users is denoted as ,user The person who posted the message.

[0046] Step two: Construct a social network based on the connection weights of social interaction behaviors among users;

[0047] In this invention, social networks based on trending events are denoted as... .

[0048] In this invention, the user set Any user in the network as a social network The nodes connect the social interaction behaviors between users with weights. As a social network Directed connection edges are used to build social networks. ,like Figure 1 As shown.

[0049] social networks Represented as .

[0050] In this invention, the social network constructed in step two This not only formally represents the ontological attributes of social media users but also integrates representations of user interaction behaviors. However, both user attributes and interaction information on social networking sites exhibit sparsity, leading to a semantic gap in the reasoning process of deep learning.

[0051] Step 3: Spatial mapping using a graph variational autoencoder;

[0052] Considering the difficulty in obtaining user sets Addressing the challenges of accessing sufficient data on individual user privacy attributes and the difficulty in obtaining adequate data during the early stages of trending events, this invention utilizes a graph variational autoencoder to extract user privacy information from social networks. Interaction space mapped to latent variable space In this study, user behavior information (including likes, comments, and reposts) is used to enhance user attribute information, thereby improving the smoothness of social information features by enhancing the correlation between multiple information sources.

[0053] Step 301, set up a single-layer latent variable model;

[0054] In this invention, the graph variational autoencoder combines Bayesian inference with the flexibility of neural networks to achieve robust representation learning. By applying reparameterization techniques, the graph variational autoencoder allows optimization of continuous random variables using standard backpropagation. In its simplest form, the graph variational autoencoder can be viewed as a single-layer latent variable model, as shown in Equation (1):

[0055] .

[0056] This represents a single-layer latent variable model.

[0057] Social networks One of the explicit features.

[0058] Representing the hidden variable space One of the hidden features.

[0059] Representing the hidden variable space One of the hidden feature variables.

[0060] Representing the hidden variable space middle Able to access social networks middle Perform optimal capture.

[0061] Step 302: Set up the encoding and decoding network model;

[0062] In this invention, the graph variational autoencoder comprises two parts: an encoder network model and a decoder network model. The encoder network model is denoted as... The decoder network model is denoted as Variational reasoning is used to make equation (1) yield the objective of maximizing the lower limit of evidence, denoted as . , This is the encoder's initial value. This is the initial value for the decoder.

[0063] In this invention, an encoder network model is first used. Receive social networks and input social networks Mapping to the latent variable space China and Israel The representation is then passed through the decoder network model. Restore social networks The objective function is optimized using a maximization function. This section uses a graph variational encoder with bidirectional encoding capabilities because such encoders add the direction of influence, fully representing the users who exert and receive the influence. The addition of influence direction information further refines the focus user representation. The interaction information makes the latent variable space A smoother surface is beneficial for improving models in cases of sparse data. Performance, Figure 2 The framework structure of the bidirectional graph encoder is shown.

[0064] In this invention, the graph variational encoder can improve the embedding representation capability of the graph posterior distribution, thereby enabling social networks... The structure and node attribute information can be optimally captured in the latent space. This applies to social networks. Mapping to the latent variable space This allows for the extraction of richer social interaction features. Compared to explicit methods... Latent variable space The features represented in the It contains stronger user-related information.

[0065] Step 303, processing of the encoding network model;

[0066] In this invention, the number of iterations of the network is denoted as... .

[0067] In this invention, social networks users in The poster, the user It serves as a reference for the social interaction behavior of posts.

[0068] social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as .

[0069] social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as .

[0070] social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as Subscript For users The identifier number.

[0071] social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as Subscript For users The identifier number.

[0072] social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as Subscript For users The identifier number.

[0073] In this invention, to make the latent variable space The space of user data representation is smoother, and a graph variational coding mechanism is introduced. Considering the bidirectional nature of user interaction, two variational forms in two directions are used as shown in equations (2) and (3). A graph neural network is used to obtain the aggregated and updated representations of each user node in the social graph, and a single gated loop is used as the update function of the forward and backward encoders.

[0074] In the In this iteration, forward information passing uses Calculation from user To users The forward message vector, denoted as . Then, information is transmitted using... Calculation from user To users The reverse message vector, denoted as Each graphic information direction is aggregated separately.

[0075] .

[0076] .

[0077] In this invention, the user is directly connected. The preceding and following terms, then the user The Second iteration Represented as:

[0078] .

[0079] In this invention, formula (4) not only captures social networks With the hidden variable space The interactive network topology also captured social networks. The information path of social interaction behavior within. In order to enable... In the encoder network model The output approximates the posterior distribution of a multivariate normal distribution, and is characterized by equation (5). Encoder network model The output after that is:

[0080] .

[0081] express Encoder network model The posterior distribution characteristics of the approximate multivariate normal distribution.

[0082] This represents the activation threshold function, the Choose the sigmoid function.

[0083] Represents a linear activation function, the Choose the Identity function.

[0084] This represents the total number of users.

[0085] For users The identifier number.

[0086] Step 304, Decoding the network model;

[0087] In this invention, because the graph variational autoencoder needs to generate representations of bidirectional nodes and directed connections between users, the generated target loss function is the superposition of four loss functions, as shown in equation (6): .

[0088] This represents the target loss function generated by the graphical variational autoencoder.

[0089] Indicates user To users The forward loss function.

[0090] Indicates user To users The inverse loss function.

[0091] Indicates user To users The forward loss function.

[0092] Indicates user To users The inverse loss function.

[0093] Decoder network model Based on the target loss function The hidden feature z is used as input to reconstruct the social network from two directions. Reconstruction of social networks . Representing the hidden variable space One of the hidden features.

[0094] This invention relies on forward and backward propagation, where each graph is iteratively constructed through a series of operations that add nodes and edges until a final node is generated. The combination of the forward and backward graphs constitutes the reconstructed output – a social network. .

[0095] After obtaining the initial user nodes and embedded representations, rebuild the social network. The edge weights are generated iteratively. The process is shared in the iterations until an end-user node is drawn. In each iteration, a new node is created and added to the social network. It takes the embedding representation and hidden features z as input, determines the type of the next missing node, and requires the newly added node to conform to the classification function. Parameters are generated for the distribution of the classified node types. The next user node is sampled from these parameters and then added to the existing embedding representation, as shown in Equation (7).

[0096] Following the After the nth iteration, the th The embedding representation of the next iteration is:

[0097] .

[0098] For each added user node, the decoder network model Create connecting edges from existing nodes based on a scoring function (where higher values ​​indicate possible edges). Using the embedded user nodes as input and combining them with the hidden feature z, the number of iterations will be obtained. The probability distribution of the connecting edges is given by equation (8), and follows a Bernoulli distribution.

[0099] .

[0100] This represents the Bernoulli distribution.

[0101] This represents a linear classification function.

[0102] A new set of directed edges is sampled from the distribution of Equation (8) to the end of the Vth user. After the new directed edges are generated, the connected nodes will be aggregated and updated, resulting in updated node embeddings. In the middle. These node embeddings are aggregated into a single graph representation. .

[0103] See Figure 2 As shown, in step three, the graph variational encoder used introduces an encoder network model. and decoder network model Ensuring the potential space is continuous and smooth further enables it to maintain a certain degree of robustness in the sparse space to adapt to more domain data spaces.

[0104] Step four, identification of focus users;

[0105] In rebuilding - social networks All network node influence measurement work assumes that adjacent nodes will influence each other.

[0106] In rebuilding - social networks In the context of user sets any user in When considering only the influence between individual user nodes, the influence between individual user nodes is formally represented by the amplification effect of nodes on feature information.

[0107] Influence between individual user nodes It can be represented as:

[0108] .

[0109] This represents hyperparameters.

[0110] This refers to the amplification effect of information between spatially adjacent neighbors.

[0111] This is used to measure potential neighbors.

[0112] Due to the enhancement effect of the latent variable space on features, many nodes with similar relationships but not spatially adjacent will appear. These nodes are called hidden neighbor nodes, and the specific calculation is shown in equations (10) and (11). Hyperparameters Used to adjust the importance of explicit and hidden neighbors.

[0113] .

[0114] .

[0115] Indicates in There are neighboring users with actual connecting edges.

[0116] Indicates in The central node in.

[0117] Indicates in Users who have no physical connection to each other but engage in social interaction.

[0118] Indicates belonging to The node centrality.

[0119] Indicates belonging to The node centrality.

[0120] Indicates belonging to The cosine similarity, and .

[0121] Indicates belonging to The cosine similarity, and .

[0122] Validation of test data

[0123] The interactive network generation method based on users and user behavior in this invention is called the SIFAF model.

[0124] The experimental dataset used in this embodiment is an extension of the FANG dataset. Since many parts of the FANG dataset lacked news content, this content was manually added based on URL information. Simultaneously, users and their profiles, such as followers, followed accounts, and gender, were located using the Twitter API based on their user IDs. After filtering out duplicate users, a total of 51,187 users and 1,011 related news items were collected. Through simple data preprocessing, some "bot accounts" and "zombie accounts" were removed, retaining only 43,278 nodes and 112,346 edges, resulting in 11,861,452 samples. However, by filtering users and news content with limited reach to improve data quality, 3,129,763 samples were ultimately obtained from the 43,278 users and 119,184 identical pieces of content. The dataset was randomly divided into training, augmentation, and test sets in a ratio of 8:1:1. Directed edges were established whenever there was any interaction, sharing, or commenting between users. Three types of directed edge attributes were considered: likes, comments, and shares. In the experiments, a graph neural network was implemented using PyTorch Geometric with 500 epochs and a learning rate of 0.005 for each training dataset. The experiment was run on a 10-core 3.5GHz GPU with 64GB of memory and the Adam optimizer. During hyperparameter tuning, if performance did not improve after 5 consecutive epochs, the learning rate was reduced by 0.5 times, and an early stopping technique with a window size of 10 epochs was introduced. One epoch represents all the data fed into the graph neural network, completing one forward computation and backward propagation cycle.

[0125] Commonly used models for leader prediction (GCN, GraphSAGE, AGC, HAT) are used as baseline models. Model performance is measured by accuracy, and AUC (Area Under Curve, defined as the area enclosed by the feature curve and the coordinate axes) and mAP (mean average precision) are used to verify the performance of the proposed method on the dataset. Social user influence can be obtained by quantifying the scope of a user's social influence (including explicit and implicit relationships). Therefore, measuring the model's performance in identifying focal users can be transformed into a problem of predicting edges in a graph. The trained graph variational coding model can transform the problem into a more latent variable space, in which the probability of interactive links under a certain hot event is predicted. Figure 3The results show the predictions of different models on the dataset. The horizontal axis shows the end-to-end training time for each epoch, and the vertical axis shows the test accuracy of the current model at the end of each epoch. The results show that although the inference time of the SIFAF model of this invention is relatively long, it has a significant advantage in accuracy performance.

[0126] The graph variational autoencoder of this invention was used on a social network dataset, and its performance improved continuously with the increase of the number of graph neural network layers. Figure 4 As can be seen, compared with other baseline models (GCN, GraphSAGE, AGC, HAT), the SIFAF model of this invention is less prone to getting trapped in local minima as the number of layers increases during the tuning process, further demonstrating the advantage of the SIFAF model of this invention in resisting overfitting. Based on Figure 4 Performance analysis was conducted, taking into account computational resource limitations. A 3-layer graph neural network was used to calculate the influence value, while simultaneously selecting 6 most likely potential neighbor users for each user. Through tuning, the hyperparameters were chosen as follows: =0.75.

[0127] When hot events occur, quickly identifying key users and promptly monitoring their social network interactions plays a very positive role in enabling cyberspace administration departments to proactively guide public opinion.

[0128] In specific trending events, effectively identifying relevant key users plays a crucial role in the focused monitoring and positive guidance of public opinion. This invention quantifies the amplification effect of information by extracting and integrating user interaction behavior features, thereby effectively identifying key users. To this end, a metric for measuring the information amplification rate of users in social networks is proposed. This metric is based on social user interaction behavior graphs and bidirectional variational graph autoencoders. By extracting user network interaction behavior features, the social network is transformed into a latent variable space to enhance user feature associations. This framework not only extracts user features but also integrates the strong correlations between social user behaviors, reducing dependence on labels. Therefore, this method has a certain robustness to sparse data with long-tail effects.

Claims

1. A method for generating interactive networks based on users and user behavior, applied to registered users on social networking sites; characterized in that... It includes the following steps: Step 1: Collect user and social behavior information from any trending event on any social networking site; Collected from social networking sites The set of registered users is represented as a set. subscript The total number of users; ~ Indicates the first to the second One registered user; User-to-user social behavior information includes actions such as liking, commenting, and / or sharing; Likes are recorded as Comments are recorded as Forwarding is recorded as ; Social behavior information between any two users is denoted as The connection weights of social interaction behaviors are denoted as ;user For the poster; Step two: Construct a social network based on the connection weights of social interaction behaviors among users; Social media posts based on trending events ; user set Any user in the network as a social network The nodes connect the social interaction behaviors between users with weights. As a social network Directed connection edges are used to build social networks. ; social networks Represented as ; Step 3: Spatial mapping using a graph variational autoencoder; Step 301, set up a single-layer latent variable model; The graphical variational autoencoder is considered as a single-layer latent variable model, as shown in equation (1): ； This represents a single-layer latent variable model; Social networks One of the explicit features; Representing the hidden variable space One of the hidden features; Representing the hidden variable space One of the hidden feature variables; Representing the hidden variable space middle Able to access social networks middle To achieve optimal capture; Step 302: Set up the encoding and decoding network model; Graph variational autoencoders include encoder network models. and decoder network model Two parts; using variational reasoning, equation (1) yields the objective of maximizing the lower limit of evidence. , This is the encoder's initial value. This is the initial value for the decoder; First, use the encoder network model. Receive social networks and input social networks Mapping to the latent variable space China and Israel The representation is then passed through the decoder network model. Restore social networks The objective function is optimized using a maximization function. Step 303, processing of the encoding network model; The number of iterations of the network is denoted as ; social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as subscript For users The identifier; social networks users in Initial Embedded Latent Variable Space The features in are represented as Using graph neural networks, user data is obtained through aggregation and update mechanisms. The The iteration is represented as subscript For users The identifier; In the In this iteration, forward information passing uses Calculation from user To users The forward message vector, denoted as . Then, information is transmitted using... Calculation from user To users The reverse message vector, denoted as Each graphic information direction is aggregated separately. ； ； Direct connection to users The preceding and following terms, then the user The iteration Represented as: ； Formula (4) not only captures social networks With the hidden variable space The interactive network topology also captured social networks. The information path of social interaction behavior in the middle; in order to enable In the encoder network model The output approximates the posterior distribution of a multivariate normal distribution, and is characterized by equation (5). Encoder network model The output after that is: ； express Encoder network model The posterior distribution characteristics of the approximate multivariate normal distribution; Indicates the activation threshold function, the Choose the sigmoid function; Represents a linear activation function, the Select the Identity function; Step 304, Decoding the network model; Because the graph variational autoencoder needs to generate representations of bidirectional nodes and directed edges between users, the target loss function is the superposition of four loss functions, as shown in equation (6): ； This represents the target loss function generated by the graphical variational autoencoder. Indicates user To users The forward loss function; Indicates user To users The inverse loss function; Indicates user To users The forward loss function; Indicates user To users The inverse loss function; Decoder network model Based on the target loss function The hidden feature z is used as input to reconstruct the social network from two directions. Reconstruction of social networks ; Representing the hidden variable space One of the hidden features; Following the After the nth iteration, the th The embedding representation of the next iteration is: ； For each added user node, the decoder network model Create connecting edges from existing nodes based on the scoring function; by Using the embedded user nodes as input and combining them with the hidden feature z, the number of iterations will be obtained. The probability distribution of the connecting edges is given by equation (8) and follows a Bernoulli distribution; ； Indicates the Bernoulli distribution; Represents a linear classification function; A new set of directed edges is sampled from the distribution of Equation (8) to the end of the Vth user. After the new directed edges are generated, the connected nodes will be aggregated and updated, resulting in updated node embeddings. In the middle; these node embeddings are aggregated into a single graph representation. ; Step four, rebuilding the social network Focus users Identification.

2. The method for generating interactive networks based on users and user behavior according to claim 1, characterized in that: In rebuilding - social networks All network node influence measurement work assumes that adjacent nodes will influence each other; In rebuilding - social networks In the context of user sets any user in When considering only the influence between individual user nodes, the influence between individual user nodes is formally represented by the amplification effect of nodes on feature information. Influence between individual user nodes It can be represented as: ； Indicates hyperparameters; This refers to the amplification effect of information between spatially adjacent neighbors; This is used to measure potential neighbors; Due to the enhancement effect of the latent variable space on features, many nodes with similar relationships but not spatially adjacent will appear. These nodes are called hidden neighbor nodes, and the specific calculation is shown in equations (10) and (11); hyperparameters Used to adjust the importance of explicit and hidden neighbors; ； ； Indicates in Neighboring users with actual connecting edges; Indicates in The central node in; Indicates in Users who have no physical connection to the network but engage in social interaction; Indicates belonging to The node centrality; Indicates belonging to The node centrality; Indicates belonging to The cosine similarity, and ; Indicates belonging to The cosine similarity, and .

3. The method for generating interactive networks based on users and user behavior according to claim 1, characterized in that: The users concentrated here are those on social networking sites who are influenced by trending events.

4. The method for generating interactive networks based on users and user behavior according to claim 1, characterized in that: Graph variational autoencoders combine Bayesian inference with the flexibility of neural networks to achieve robust representation learning.

Images