A social network group discovery method based on user preference

Abstract

The application belongs to the field of guided topic network analysis, and particularly relates to a social network group discovery method based on user preferences, comprising: obtaining relevant information from a social network platform; calculating the interaction degree and reputation degree of users; defining a random walk strategy of IBT2vec algorithm, and representing the user topic network as a feature vector matrix; using RoBERTa to obtain the long-term and short-term preferences of users, fusing the same with the initial feature vector to form an enhanced user node representation; compensating for the network space features through meta-transfer learning to alleviate the data sparsity problem; classifying the user features through a spectral clustering method, and obtaining key influential user groups with similar interests and behaviors after full connection layer and normalization processing. The application not only considers the similarity between users, but also comprehensively considers the influence of users, so that the group discovery is more accurate and practical, and reliable support is provided for applications such as public opinion monitoring, information dissemination and marketing promotion.

Description

Technical Field

[0001] This invention belongs to the field of guided topic network analysis, specifically involving a method for discovering social network groups based on user preferences. Background Technology

[0002] With the rapid development of social networks, people widely participate in discussions and dissemination of various topics. Among these discussions, some groups exploit platform mechanisms and their own influence to create public opinion and guide trends for various purposes. In recent years, scholars have found that identifying influential user groups in areas such as influence analysis, public opinion management, and target market positioning has become a research hotspot. Monitoring potential key groups within influential online topics often has a positive effect, and efforts are being made to explore and address the challenges involved. Therefore, identifying key groups within influential online topics is of great significance for the current governance of the online environment.

[0003] Currently, research on the early-stage identification of online groups influencing trending topics on social networks mainly focuses on three aspects: social network user analysis, social network structure analysis, and dynamic evolution. Research on social network users primarily includes the identification of opinion leaders and the discovery of user communities. Research on social network structure analysis mainly focuses on the relationships, interactive behaviors, and information dissemination paths among members within a group. Regarding dynamic evolution, researchers use time-series analysis of social network data to reveal changes in group structure and member relationships over time, in order to predict the group's evolutionary trends.

[0004] In summary, while academic research on trending topics is already quite extensive, identifying key groups in the early stages of trending topics on social networks still faces the following challenges: 1. Sparsity of data in the early stages of trending topic outbreaks. User engagement is low in the early stages, resulting in a lack of effective data, and the data distribution may be uneven within local areas, making it difficult to represent global characteristics. Therefore, effectively mitigating data sparsity is a challenge. 2. The evolution of topic content. Topic evolution involves various complex forms such as topic updates, decay, revival, and content diversity, making direct analysis and modeling difficult. Therefore, effectively mining the correlations in topic evolution is a challenge. 3. The complexity of relationships between topic elements. The relationships between topic elements are complex, involving social relationships between users, transitional relationships between topics, and user participation and interaction with topics. Effectively mining the complex relationships between topic elements and capturing user groups with similar behaviors and connections urgently needs to be addressed. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for discovering social network groups based on user preferences, comprising:

[0006] S1: Obtain basic historical attributes, user relationship characteristics, and topic data from social networks;

[0007] The topic data includes: original topics and derived topics;

[0008] S2: Calculate the user's interaction degree and reputation degree based on the user's basic attribute characteristics and user relationship characteristics;

[0009] S3: Define a random walk strategy based on user interaction and credibility to better capture user behavior in social networks;

[0010] S4: Based on the random walk strategy, the IBT2vec algorithm is defined to represent the user's behavior in the topic network as a vector, thus obtaining the original user feature vector;

[0011] S5: Based on topic data, the RoBERTa model is used to obtain users' long-term and short-term preferences, which are then concatenated and fused with the original user feature vectors to obtain an enhanced user feature representation of the original topic domain;

[0012] S6: Map the enhanced user feature representation of the original topic domain to the derived topic domain to obtain the enhanced user feature representation of the derived topic domain;

[0013] S7: Construct a Laplacian matrix based on the enhanced user feature representations of the original topic domain and the derived topic domain to calculate the user feature vector, and use the k-means algorithm to perform spectral clustering on the user feature vector to obtain the user clustering results;

[0014] S8: Input the clustering results into the fully connected layer, use the softmax function to normalize the output of the fully connected layer to obtain the probability output, and determine whether the user is a key user based on the probability output. A clustering result of 0 indicates that the user is not important, and a clustering result of 1 indicates that the user is important. The set of users with a clustering result of 1 is the key group.

[0015] The beneficial effects of this invention are:

[0016] This invention calculates user interaction and credibility by comprehensively understanding user behavior patterns and social relationships, accurately assessing user activity and trustworthiness, and providing effective data support for group discovery. It defines a random walk strategy to capture user behavior patterns in social networks and uses the IBT2vec algorithm to represent the user topic network as a feature vector matrix. RoBERTa is used to obtain users' long-term and short-term preferences, which are then concatenated and fused with the original user feature vectors to form an enhanced user feature representation, improving the expressive power and accuracy of user features. Meta-transfer learning is used to learn shared meta-knowledge on multiple native topics and then transfer it to derived topics, rapidly adapting to new topic changes through fine-tuning, enhancing adaptability to new topics. A Laplacian matrix is ​​constructed to calculate user feature vectors, forming a user feature vector matrix. This matrix is ​​then processed and normalized in a fully connected layer to obtain a probability output, which is used to determine whether a user is a key user. Compared to existing technologies, this invention not only considers the similarity between users but also comprehensively considers user influence, making group discovery more accurate and practical, providing reliable support for applications such as public opinion monitoring, information dissemination, and marketing promotion. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a social network group discovery method based on user preferences according to the present invention;

[0018] Figure 2 This is a schematic diagram illustrating the process of obtaining the user feature vector matrix in a social network group discovery method based on user preferences according to the present invention.

[0019] Figure 3 This is a schematic diagram of spectral clustering in a social network group discovery method based on user preferences according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention proposes a method for discovering social network groups based on user preferences, such as... Figure 1 As shown, it includes:

[0022] S1: Obtain basic historical attributes, user relationship characteristics, and topic data from social networks;

[0023] The topic data includes: original topics and derived topics;

[0024] S2: Calculate the user's interaction degree and reputation degree based on the user's basic attribute characteristics and user relationship characteristics;

[0025] S3: Define a random walk strategy based on user interaction and credibility to better capture user behavior in social networks;

[0026] S4: Based on the random walk strategy, the IBT2vec algorithm is defined to represent the user's behavior in the topic network as a vector, thus obtaining the original user feature vector;

[0027] S5: Based on topic data, the RoBERTa model is used to obtain users' long-term and short-term preferences, which are then concatenated and fused with the original user feature vectors to obtain an enhanced user feature representation of the original topic domain;

[0028] S6: Map the enhanced user feature representation of the original topic domain to the derived topic domain to obtain the enhanced user feature representation of the derived topic domain;

[0029] S7: Construct a Laplacian matrix based on the enhanced user feature representations of the original topic domain and the derived topic domain to calculate the user feature vector, and use the k-means algorithm to perform spectral clustering on the user feature vector to obtain the user clustering results;

[0030] S8: Input the clustering results into the fully connected layer, use the softmax function to normalize the output of the fully connected layer to obtain the probability output, and determine whether the user is a key user based on the probability output. A clustering result of 0 indicates that the user is not important, and a clustering result of 1 indicates that the user is important. The set of users with a clustering result of 1 is the key group.

[0031] Methods for obtaining topic-related web data include using social media APIs, web crawling techniques, public dataset platforms, or subscribing to third-party data services to obtain advanced social media data.

[0032] In this embodiment, calculating the user's interaction level includes:

[0033]

[0034] Among them, UserInt(u i ,u j ,t) represents user u i and user u j The user interaction level at time t, userInt(u i ,u j ) represents user u i and user u j The degree of user interaction within the time period T = (t1, t2), L i,j Represents user node ui and user node u j The number of relational levels, N represents user u i and user u j The total number of original topic messages published in the topic, where M represents the M types of interactive behaviors between users in the topic space. Indicates the kth m Weighting coefficients for different interaction methods Representing user u at time t within the topic space respectively i u j With all users on the network, the kth m The total number of interactions, U i,j (k m ,t) represents user u at time t in the topic space. i and user u j Perform the kth m The total number of interactions.

[0035] k m The weight coefficients for each interaction method are expressed as follows:

[0036]

[0037] In this embodiment, calculating user reputation includes:

[0038]

[0039] Among them, UserCred(u j ) represents user u j Reputation level, PlatCred(u i ) represents user u i Account level on this social media platform, Fans(u i ) represents user u i The number of followers, Follow(u i ) represents user u i Number of followers, w i ∈[0,1] represents user u i The weight of its historical reputation.

[0040] In this embodiment, the user's own attributes and the relationships between users in the topic network space contain key group behavior information. Therefore, this invention leverages the advantages of random walk experiments in network topology representation learning, combines topic network node attributes to vectorize nodes, and mines the hidden relationships between nodes.

[0041] During the random walk, a length of l is set, and nodes u are defined. i to node u j Random walk strategies include:

[0042]

[0043] Wherein, P(u j =x|u i ) represents node u i to node u j The transition probability, W i (u i ,u j ) represents two nodes u i ,u j The edge weights between them, UserCred(u j ) represents user u j The credibility, Z represents the normalization coefficient, α p,q (u i ,u j ) represents the wandering bias term, E u This represents all the relationship edges between users who participate in the spread of a topic on a social network.

[0044] The wandering bias term includes:

[0045]

[0046] in, Represents node u i to node u j The shortest path length between them, p and q represent the first and second walking parameters used to control the bias during the walking process.

[0047] Two nodes u i ,u j The edge weights between them include:

[0048]

[0049] Here, U represents all users who participated in spreading the topic.

[0050] In this embodiment, the IBT2vec algorithm, defined according to the random walk strategy, represents the user's behavior in the topic network as a vector, resulting in the original user feature vector, including:

[0051] D u ={D1,D2,...,D N}

[0052] Among them, D u D represents the original user feature vector. N This represents the node vector representation of the Nth user.

[0053] In this embodiment, the RoBERTa model is used to obtain users' short-term and long-term preferences based on topic data, including:

[0054] By processing user interaction sequences in topic data using the RoBERTa model, we can obtain trends in user interest changes across different topics, derive short-term and long-term user preferences, and concatenate and fuse these preferences with the original user feature vectors to form an enhanced user feature representation. Figure 2 As shown.

[0055] To measure user similarity, user interests are used as probability distributions, and the KL divergence is used to evaluate the difference between the two probability distributions, thus measuring user similarity. Within the time period T = (t1, t2), the similarity between users is defined as S(u i ,u j The formula is defined as follows:

[0056]

[0057] Where p(u) i k) represent user u i The probability distribution of the k-th topic in the interest preference sequence.

[0058] The objective function for optimizing the IBT2vecc method is defined as:

[0059] max∑ u logP(N S (u)|φ(u))

[0060] Where, N S (u) represents the set of nearest neighbors of node u in the network obtained through the walking strategy, P(N) S (u)|φ(u)) represents observing node u and its neighborhood N simultaneously, given the feature representation mapping φ. S The probability of (u).

[0061] In this embodiment, user features from the original topic domain are mapped to derived topic domains, and the feature representation of the derived topic domains is enhanced, including:

[0062]

[0063] in, T represents the parameters of the meta-learning model, ← indicates parameter update, α and β represent the internal and external learning rates respectively, and T i Let p(T) represent the task, and p(T) represent the sampling task distribution. Represents the vector differential operator. Indicates task T i The loss function f onθ This represents a model that takes parameter θ as input.

[0064] In this embodiment, the user feature vector is calculated by constructing a Laplacian matrix based on the enhanced user feature representations of the original topic domain and the derived topic domain, including:

[0065] Calculate the similarity between users to construct a similarity matrix;

[0066] Construct the Laplacian matrix using the similarity matrix: L = DS, where L represents the Laplacian matrix, D represents the degree matrix, and D(i,j) = ∑ j S(i,j), where S represents the similarity matrix;

[0067] Perform eigenvalue decomposition on the Laplacian matrix L to obtain the user feature vectors corresponding to the first q smallest eigenvalues. in, This represents a user feature vector matrix of size n×q.

[0068] In this embodiment, calculating the similarity between users includes:

[0069]

[0070] Where S(i,j) represents user u i with u j The similarity between them, E i E j Let represent the embedding vectors of the i-th and j-th users, respectively, where · represents the inner product and |||| represents the norm of the vector.

[0071] In this embodiment, the k-means algorithm is used to perform spectral clustering on the user feature vectors to obtain the user clustering results, such as... Figure 3 As shown, it includes:

[0072] clusters = k-means(U,k)

[0073] Where clusterers represents the clustering results of users, k represents the cluster size in the k-means algorithm, and U represents all users who participated in the topic propagation.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for discovering social network groups based on user preferences, characterized in that, include: S1: Obtain basic historical attributes, user relationship characteristics, and topic data from social networks; The topic data includes: original topics and derived topics; S2: Calculate the user's interaction degree and reputation degree based on the user's basic attribute characteristics and user relationship characteristics; Calculating user reputation includes: Among them, UserCred(u j ) represents user u j Reputation level, PlatCred(u i ) represents user u i Account levels on social media platforms, Fans(u i ) represents user u i The number of followers, Follow(u i ) represents user u i Number of followers, w i ∈[0,1] represents user u i The weight of historical reputation; S3: Define a random walk strategy based on user interaction and credibility to better capture user behavior in social networks; S4: Based on the random walk strategy, the IBT2vec algorithm is defined to represent the user's behavior in the topic network as a vector, thus obtaining the original user feature vector; S5: Based on topic data, the RoBERTa model is used to obtain users' long-term and short-term preferences, which are then concatenated and fused with the original user feature vectors to obtain an enhanced user feature representation of the original topic domain; S6: Map the enhanced user feature representation of the original topic domain to the derived topic domain to obtain the enhanced user feature representation of the derived topic domain; By mapping user features from the original topic domain to the derived topic domain, an enhanced user feature representation of the derived topic domain is obtained, including: in, T represents the parameters of the meta-learning model, ← indicates parameter update, α and β represent the internal and external learning rates respectively, and T i Let p(T) represent the task, and p(T) represent the sampling task distribution. Represents the vector differential operator. Represents task T i The loss function f on θ This represents a model that takes parameter θ as input. S7: Construct a Laplacian matrix based on the enhanced user feature representations of the original topic domain and the derived topic domain to calculate the user feature vector, and use the k-means algorithm to perform spectral clustering on the user feature vector to obtain the user clustering results; S8: Input the clustering results into the fully connected layer, use the softmax function to normalize the output of the fully connected layer to obtain the probability output, and determine whether the user is a key user based on the probability output. A clustering result of 0 indicates that the user is not important, and a clustering result of 1 indicates that the user is important. The set of users with a clustering result of 1 is the key group.

2. The method for discovering social network groups based on user preferences according to claim 1, characterized in that, Calculating user interaction includes: Among them, UserInt(u i ,u j ,t) represents user u i and user u j The user interaction level at time t, userInt(u i ,u j ) represents user u i and user u j The degree of user interaction within the time period T = (t1, t2), L i,j Represents user node u i and user node u j The number of relational levels, N represents user u i and user u j The total number of original topic messages published in the topic, where M represents the M types of interactive behaviors between users in the topic space. Indicates the kth m Weighting coefficients for different interaction methods Representing user u at time t within the topic space. i u j With all users on the network, the kth m The total number of interactions, U i,j (k m ,t) represents user u at time t in the topic space. i and user u j Perform the kth m The total number of interactions.

3. The method for discovering social network groups based on user preferences according to claim 1, characterized in that, The random walk strategy is defined based on user interaction and reputation, including: Wherein, P(u j =x|u i ) represents node u i to node u j The transition probability, W i (u i ,u j ) represents two nodes u i ,u j The edge weights between them, UserCred(u j ) represents user u j The credibility, Z represents the normalization coefficient, α p,q (u i ,u j ) represents the wandering bias term, E u This represents all the relationship edges between users who participate in the spread of a topic on a social network.

4. The method for discovering social network groups based on user preferences according to claim 1, characterized in that, Based on topic data, the RoBERTa model is used to obtain users' short-term and long-term preferences, including: By processing user interaction sequences in topic data using the RoBERTa model, we can obtain the trend of user interest changes under different topics and obtain users' long-term and short-term preferences.

5. The method for discovering social network groups based on user preferences according to claim 1, characterized in that, To compute user feature vectors, a Laplacian matrix is ​​constructed based on the enhanced user feature representations of the original and derived topic domains, including: Calculate the similarity between users to construct a similarity matrix; Construct the Laplacian matrix using the similarity matrix: L = DS, where L represents the Laplacian matrix, D represents the degree matrix, and D(i,j) = ∑ j S(i,j), where S represents the similarity matrix; Perform eigenvalue decomposition on the Laplacian matrix L to obtain the user feature vectors corresponding to the first q smallest eigenvalues. in, This represents a user feature vector matrix of size n×q.

6. The method for discovering social network groups based on user preferences according to claim 5, characterized in that, Calculating similarity between users includes: Where S(i,j) represents user u i with u j The similarity between them, E i E j Let represent the embedding vectors of the i-th and j-th users, respectively, where · represents the inner product and |||| represents the norm of the vector.

7. The method for discovering social network groups based on user preferences according to claim 1, characterized in that, The k-means algorithm is used to perform spectral clustering on user feature vectors to obtain the user clustering results, including: clusters = k-means(U,k) Where clusterers represents the clustering results of users, k represents the cluster size in the k-means algorithm, and U represents all users who participated in the topic propagation.

Images