Social network user group discovery method and system
By constructing a similarity matrix and AICD model, combining autoencoder and non-negative matrix decomposition technology, the community discovery method is optimized, and the problem of insufficient adaptability of community discovery methods to complex and dynamic networks in the existing technology is solved, and the problem of insufficient adaptability of community discovery methods to complex and dynamic networks is achieved. Efficient and accurate community discovery.
Patent Information
- Application Number
- CN202410928065.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-19
AI Technical Summary
The existing community discovery method relies on the adjacency matrix, cannot fully capture the overall similarity and complex correlation between nodes, is difficult to process high-dimensional data, adapt to dynamic networks and multiple relationships, resulting in inaccuracy and inefficiency of community discovery.
The similarity matrix is constructed, using the AICD model combined with autoencoder and non-negative matrix decomposition technology, and the Xi and Yq matrix is updated through objective function optimization and block coordinate descent method, and the Dice coefficient and sparse regularization term are used to ensure the accuracy of the model and Interpretability.
It improves the accuracy and efficiency of community discovery, can better capture deep relationships and dynamic changes in complex networks, enhances the understanding and representation of community structure, and is suitable for complex social network environments.
Smart Images

Figure CN118982434A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to but is not limited to the field of Internet technology, and in particular relates to a method and system for discovering user groups in a social network. Background Art
[0002] Common community discovery methods, such as spectral clustering and non-negative matrix factorization-based methods, are used to process high-dimensional adjacency matrices.
[0003] The prior art provides a modular non-negative matrix factorization (M-NMF) model that utilizes the consensus relationship between node representation and community structure, and jointly optimizes the NMF-based representation learning model and the modularity-based community detection model, so that the learned representation of the node retains both the microscopic and community structures, improving the comprehensiveness and comprehensiveness of the representation. Due to the need to optimize representation learning and community detection at the same time, the computational complexity of M-NMF is relatively high, and its generalization ability in dealing with complex networks or with more noise needs to be further verified.
[0004] The adjacency matrix only records the shallow connection relationship between nodes, resulting in insufficient evaluation of the overall similarity between nodes. This limitation will have an adverse impact on the community detection results. In addition, as the complexity of social networks continues to increase, the relationship between nodes will extend to more complex and abstract connections, which traditional community detection methods cannot accurately capture.
[0005] The existing technologies have the following technical problems and defects in community discovery tasks:
[0006] 1) Limitations of the adjacency matrix: Traditional community discovery methods, such as those based on spectral clustering and non-negative matrix factorization, mainly rely on the adjacency matrix for calculation. However, the adjacency matrix only records the direct connection relationship between nodes (i.e., shallow connection), and fails to fully reflect the overall similarity and complex associations between nodes. This limitation leads to insufficient evaluation of node relationships, which will affect the accuracy and effectiveness of community discovery.
[0007] 2) Insufficient ability to capture deep relationships: As the complexity of social networks continues to increase, the relationships between nodes will extend to more complex and abstract levels. Traditional community detection methods are obviously insufficient in capturing these deep relationships, resulting in the inability to accurately identify and divide communities when faced with complex network structures.
[0008] 3) Challenges in high-dimensional data processing: Data in social networks are usually high-dimensional. Traditional methods based on adjacency matrices have high computational complexity and low efficiency when processing high-dimensional data. At the same time, these methods are prone to data sparsity problems in high-dimensional data, which further affects the effectiveness of community discovery.
[0009] 4) Poor adaptability to dynamic networks: Social networks are dynamic, and the relationships between nodes and edges change over time. Traditional community discovery methods are usually static and difficult to adapt to dynamic changes in network structures. They cannot update community discovery results in real time, which affects their effectiveness in practical applications.
[0010] 5) Weak ability to handle multiple relationships: In actual social networks, there are multiple types of relationships between nodes (such as friendships, cooperation, etc.), and these multiple relationships constitute a complex network structure. Traditional community discovery methods are difficult to handle multiple types of relationships at the same time, resulting in incomplete understanding and analysis of the network structure, affecting the accuracy of community discovery.
[0011] In summary, the technical problems that the existing technologies face in the community discovery task are: the adjacency matrix only records the shallow connection relationship between nodes, resulting in insufficient evaluation of the overall similarity between nodes. To solve these problems, it is necessary to explore new methods and technical means to improve the accuracy and applicability of community discovery, so as to better capture the deep relationships and dynamic changes in complex networks. Summary of the invention
[0012] In view of the problems existing in the prior art, the present invention provides a method and system for discovering user groups in a social network.
[0013] The present invention is implemented as follows: a method for discovering user groups in a social network, comprising:
[0014] S1, Similarity matrix construction: In social networks, similarity matrices are used for clustering, community detection, and link prediction. The similarity measurement is based on different characteristics of the network, including node degree and shortest path;
[0015] S2, AICD model: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks;
[0016] S3, objective function optimization: In order to optimize the objective function, we need to find X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems;
[0017] S4,X i and Y q Update rule: According to the multiplication update rule, derive X i and Y qThe update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization;
[0018] S5, community discovery algorithm based on AICD: The input of the algorithm is the constructed similarity matrix, the regularization parameter α, and the maximum number of iterations T; first, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
[0019] Furthermore, S1 specifically includes:
[0020] Consider the neighbors of a node as a similarity measure and construct a similarity matrix S = [S ij ]∈R |V|*|V| , where the element S ij represents the similarity between node i and node j, V is the node set;
[0021] Since the Dice coefficient is usually used to compare the similarity between two sets, it is selected as a method to measure node similarity; assuming there are two sets A and B, the calculation of the Dice similarity coefficient between them is shown in formula (1):
[0022]
[0023] Where |A| and |B| represent the number of elements in sets A and B respectively, and A∩B represents the intersection of A and B. Similarly, the Dice similarity coefficient is used to represent the network, and the number of neighbors of a node is regarded as a similarity measure. Then the similarity between node i and node j is shown in formula (2):
[0024]
[0025] Where N represents the number of neighbors of a node, N(v i )∩N(v j ) represents the number of common neighbors of node i and node j; the value of Dice ranges from 0 to 1. If the two sets are the same, the Dice similarity coefficient is 1; if the two sets are disjoint, the Dice similarity coefficient is 0.
[0026] Furthermore, S2 specifically includes:
[0027] The similarity matrix S is the input of the model. The introduction of the encoder enables the similarity matrix to be mapped to a low-dimensional community member space. Specifically, the encoder extracts the potential features between nodes, so that the low-dimensional representation can more accurately reflect the community structure. The decoder uses the low-dimensional representation to restore the topological structure of the original network through inverse operations. The design of this structure enables the model to better understand the input data and capture the key information of the data in the process of dimensionality reduction.
[0028] In this model, the form of the objective function covers the idea of non-negative matrix factorization, which aims to learn a more compact and informative community member space representation by minimizing the difference between the similarity matrix and the product of two low-rank matrices; this process is carried out under the non-negativity constraint, thus allowing the results to be explained intuitively. The form of the objective function is shown in formula (3):
[0029]
[0030] This objective function is the sum of three different terms, each of which has a different role. D is the reconstruction error term used to quantify the difference between the original matrix S and the matrix X1X2…X q Y q The difference between the approximations derived from the product; this term ensures the fidelity of the model to the original data, thus providing an accurate representation of the network structure;
[0031] The second item J E is a sparse regularization term that encourages the features learned by the model to be sparse. This sparsity plays a key role in the model because it allows for easier feature selection and interpretation. With sparser features, each community is expected to have a set of significant features, making the community structure more interpretable.
[0032] Item 3J reg is a smoothing regularization term, which multiplies the hyperparameter α to ensure that adjacent features have similar values, which is crucial for preserving the local structure of the data; specifically, since adjacent nodes are more likely to belong to the same community, and αJ reg Ensure that the model does not produce overly fragmented community structures by requiring adjacent nodes to have similar feature representations;
[0033] These three items guide the AICD model to accurately reflect the network community structure while maintaining interpretability and structural coherence; the balance between them can be adjusted using the hyperparameter α, allowing the performance of the model to be fine-tuned according to the characteristics of the network.
[0034] Furthermore, S3 specifically includes:
[0035] Given X iMatrix, we can solve the following optimization problem by formula (4) to update Y q :
[0036]
[0037] Among them, Y q ≥0;
[0038] Similarly, each X can be updated by solving the following optimization problem i matrix:
[0039]
[0040] For all i, it is required that X i ≥0;
[0041] Various optimization techniques can be used to solve these subproblems, such as multiplicative update rules derived from the Karush-Kuhn-Tucker (KKT) conditions for non-negative matrix factorization; these update rules ensure that X i and Y q The non-negativity of the matrix and decreases the value of the objective function in each iteration.
[0042] Furthermore, S4 specifically includes:
[0043] The objective function without constraints is defined as F(X i ,Y q ), then the Lagrangian function of the problem can be written as follows:
[0044]
[0045] Among them, H i and G q is the Lagrange multiplier associated with the non-negativity constraint, and <·,·> represents the Frobenius inner product. The KKT condition for this problem is:
[0046]
[0047] According to these conditions, we can derive X i and Y q The multiplicative update rule for :
[0048]
[0049] Here ⊙ represents element-wise multiplication, and the division symbol represents element-wise division.
[0050] Another object of the present invention is to provide a social network user group discovery system for implementing the social network user group discovery method, comprising:
[0051] Similarity matrix building module: In social networks, similarity matrix is used for clustering and community detection. The similarity measurement is based on different characteristics of the network, including node degree and shortest path;
[0052] AICD model building module: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks;
[0053] Objective function optimization module: In order to optimize the objective function, we need to find the X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems;
[0054] Matrix update module: derive X according to the multiplication update rule i and Y q The update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization;
[0055] Community discovery algorithm module based on AICD: The input of the algorithm is the constructed similarity matrix, regularization parameter α, and the maximum number of iterations T; first, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
[0056] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor executes the steps of the social network user group discovery method.
[0057] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the social network user group discovery method.
[0058] Another object of the present invention is to provide an information data processing terminal, which includes the social network user group discovery system.
[0059] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:
[0060] First, the present invention constructs a similarity matrix that reflects the similarity of nodes in the network, and proposes an AutoEncoder-based community discovery model to learn the low-dimensional representation of the network to reconstruct the matrix; finally, an optimization algorithm is proposed for clustering using the node similarity in the reconstructed matrix, thereby achieving high-quality community division.
[0061] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are:
[0062] Accurate user group identification: It can help social networking platforms accurately identify and understand the structure and behavioral characteristics of user groups, thereby optimizing personalized recommendations and advertising targeting, and improving user satisfaction and engagement.
[0063] Enhance platform stickiness: By better understanding the community structure among users, social networking platforms can design more attractive social functions and activities, thereby enhancing user stickiness and activity.
[0064] Optimize marketing strategies: Effective user group discovery can provide marketing teams with valuable market insights.
[0065] Help them optimize their advertising strategies and promotional activities to improve advertising conversion rates.
[0066] Third, the present invention proposes a method for discovering user groups in social networks based on the AICD model, which overcomes many difficulties in the prior art. First, in the construction of the similarity matrix, a similarity measurement method based on the Dice coefficient is introduced to calculate the similarity between nodes by the number of node neighbors and the number of common neighbors. This method can more accurately capture the relationship between nodes in social networks, provide more accurate input data, and solve the limitations of traditional methods in similarity measurement.
[0067] Secondly, the present invention adopts an autoencoder structure in the AICD model and combines it with non-negative matrix decomposition technology to enable the similarity matrix to be mapped to a low-dimensional community member space. The encoder extracts the potential features between nodes so that the low-dimensional representation more accurately reflects the community structure, and the decoder restores the topological structure of the original network through inverse operations. This design enables the model to capture key information of the data during the dimensionality reduction process, significantly improving the model's understanding and representation of the community structure.
[0068] In terms of objective function optimization, the present invention minimizes the difference between the product of the similarity matrix and the low-rank matrix, combined with sparse regularization and smooth regularization terms, to ensure that the model can not only maintain high fidelity to the original data, but also learn sparse and explanatory features while retaining the local structure of the data. By adjusting the balance of the three terms by hyperparameters, the model can be fine-tuned according to the network characteristics to achieve optimal performance. This innovative optimization method improves the accuracy and interpretability of the model.
[0069] Finally, the present invention updates the Xi and Yq matrices through the multiplication update rule, ensuring the non-negativity of the matrices and reducing the value of the objective function in each iteration. This optimization technique ensures the stability and convergence of the model, allowing the community discovery algorithm to still operate effectively in a complex social network environment. Overall, the present invention has made significant progress in algorithm design, model structure, and optimization methods, providing an efficient and accurate solution for discovering user groups in social networks.
[0070] Fourth, the social network user group discovery method based on the AICD model proposed in the present invention solves many technical problems in the prior art in social network analysis and community detection. First, traditional similarity measurement methods are usually unable to accurately capture the complex relationships between nodes. The present invention introduces the Dice coefficient and uses the number of node neighbors and the number of common neighbors to calculate node similarity, thereby constructing a more accurate and effective similarity matrix. This improvement improves the accuracy of node similarity measurement in social networks and lays a solid foundation for subsequent group discovery.
[0071] Secondly, existing social network analysis methods are often inefficient when processing high-dimensional data, and it is difficult to effectively extract and represent community structures. To solve this problem, the present invention adopts the autoencoder structure and non-negative matrix decomposition technology in the AICD model, maps the similarity matrix to the low-dimensional community member space through the encoder, and then restores the topological structure of the original network by the decoder. This method can not only capture the potential features between nodes, but also retain the key information of the data during the dimensionality reduction process, greatly improving the expression ability and computational efficiency of the model.
[0072] In terms of optimizing the objective function, the present invention combines three different terms: reconstruction error, sparse regularization, and smooth regularization, ensuring that the model can maintain high fidelity to the original data during the learning process, learn sparse and interpretable features, and retain the local structure of the data. This multi-constrained optimization design enables the model to more accurately reflect the community structure in social networks, significantly improving the effect and interpretability of community discovery.
[0073] Finally, in order to ensure the stability and convergence of the model, the present invention adopts the multiplication update rule to update the Xi and Yq matrices, ensure their non-negativity, and reduce the value of the objective function in each iteration. This optimization technique effectively solves the problems of poor convergence and unstable results in traditional methods, so that the community discovery algorithm can still operate efficiently in a complex social network environment. In general, the present invention has made significant technical progress in algorithm design, model structure and optimization method, and provides an efficient, accurate and well-interpreted solution for social network user group discovery.
[0074] Fifth, the present invention proposes a solution based on the AICD model to address the technical problems existing in the specific application of prior art in the discovery of social network user groups. In traditional social network analysis methods, the measurement of node similarity usually has the problem of insufficient precision and cannot accurately reflect the complex relationship between nodes, resulting in unsatisfactory community detection results. In order to solve this problem, the present invention introduces the Dice coefficient, uses the number of node neighbors and the number of common neighbors to calculate node similarity, constructs a more accurate similarity matrix, and improves the accuracy of similarity measurement, thereby providing a reliable data basis for subsequent community discovery.
[0075] Existing social network analysis methods often face the problems of low efficiency and huge consumption of computing resources when processing high-dimensional and large-scale data. The present invention adopts the autoencoder structure and non-negative matrix decomposition technology in the AICD model, maps the high-dimensional similarity matrix to the low-dimensional community member space through the encoder, and then restores the topological structure of the original network by the decoder. Such a design not only effectively reduces the data dimension and computational complexity, but also retains key community structure information, thereby improving the computational efficiency and expression ability of the model.
[0076] In terms of optimizing the objective function, traditional methods often find it difficult to balance the interpretability of the results and the local structure of the data while ensuring the accuracy of the model. To this end, the present invention introduces three different terms: reconstruction error, sparse regularization, and smoothing regularization. Through multi-constrained optimization design, the model can learn sparse and interpretable features while maintaining high fidelity and retaining the local structure of the data. This method significantly improves the accuracy of community detection and the interpretability of the results, making the model more practical in industrial applications.
[0077] Finally, traditional methods often perform poorly in terms of model convergence and result stability, and are difficult to operate stably in complex social network environments. The present invention uses a multiplication update rule to update the Xi and Yq matrices to ensure their non-negativity and effectively reduce the value of the objective function in each iteration. This optimization technique not only ensures the stability and convergence of the model, but also enables the algorithm to operate efficiently in a complex social network environment, significantly improving the effect and reliability of community discovery. Overall, the present invention has made significant progress in solving specific application problems in the prior art, and provides the industry with an efficient, accurate and highly interpretable social network user group discovery solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0078] Figure 1 is a flow chart of a method for discovering user groups in a social network provided by an embodiment of the present invention;
[0079] Figure 2 is a structural diagram of a social network user group discovery system provided by an embodiment of the present invention;
[0080] Figure 3 is a schematic diagram of constructing a similarity matrix provided by an embodiment of the present invention;
[0081] Figure 4 is a structural diagram of an AICD provided by an embodiment of the present invention;
[0082] Figure 5 is a parameter sensitivity analysis diagram provided by an embodiment of the present invention; wherein, (a) KIT data set, (b) ERRON data set;
[0083] Figure 6 is a schematic diagram of convergence provided by an embodiment of the present invention; wherein, (a) KIT data set, (b) ERRON data set;
[0084] Figure 7 is the NMI result of non-overlapping community detection provided by the embodiment of the present invention
[0085] Figure 8 is the ARI result of non-overlapping community detection provided by the embodiment of the present invention
[0086] Fig. 9 The NMI result of the overlapping community detection provided by the embodiment of the present invention is
[0087] Fig.10 The ARI result of overlapping community detection provided by the embodiment of the present invention is
[0088] Fig.113 is a visualization diagram of community discovery results provided by an embodiment of the present invention; wherein, (a) ERRON, (b) KIT, (c) GN_data, and (d) LFR_10 (k=0.5).
[0089] Fig.12 1. The embodiment of the present invention provides a rendering before excavation (A); and a rendering after excavation (B). DETAILED DESCRIPTION
[0090] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0091] like Figure 1 As shown, the social network user group discovery method provided by the embodiment of the present invention includes:
[0092] S1, Similarity matrix construction: In social networks, similarity matrices are used for clustering, community detection, and link prediction. The similarity measurement is based on different characteristics of the network, including node degree and shortest path;
[0093] S2, AICD model: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks;
[0094] S3, objective function optimization: In order to optimize the objective function, we need to find X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems;
[0095] S4,X i and Y q Update rule: According to the multiplication update rule, derive X i and Y q The update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization;
[0096] S5, community discovery algorithm based on AICD: The input of the algorithm is the constructed similarity matrix, the regularization parameter α, and the maximum number of iterations T; first, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q, until convergence.
[0097] like Figure 2 As shown, the social network user group discovery system provided by the embodiment of the present invention includes:
[0098] Similarity matrix building module: In social networks, similarity matrix is used for clustering, community detection and link prediction. The similarity measurement is based on different characteristics of the network, including node degree and shortest path;
[0099] AICD model building module: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks;
[0100] Objective function optimization module: In order to optimize the objective function, we need to find the X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems;
[0101] Matrix update module: derive X according to the multiplication update rule i and Y q The update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization;
[0102] Community discovery algorithm module based on AICD: The input of the algorithm is the constructed similarity matrix, regularization parameter α, and the maximum number of iterations T; first, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
[0103] 1. Discovery of social network user groups
[0104] Social network user group discovery refers to identifying and discovering user groups with similar interests, characteristics or goals through in-depth analysis of user behavior on social network platforms. This task is usually completed using community discovery algorithms, with the goal of more comprehensively understanding the relationships between users in social networks and discovering potential influential nodes, providing strong support for social network analysis, information dissemination research, etc.
[0105] When discovering social network user groups on the Weibo platform, we first need to collect relevant data, including the interactive relationships of Weibo users, which constitute the edges in the network, and the interactions between a large number of users form a complex network structure. By mining the network through community discovery algorithms, we can identify social network user groups with similar interests or similar topics. By analyzing the user's interactive behavior and information dissemination path, we can identify those users who have an important position in the user group and play a key role in the information dissemination of the entire social network.
[0106] 2. Community Discovery Model Based on AutoEncoder
[0107] First, the construction process of the similarity matrix is introduced in detail, which involves how to quantify the similarity between nodes and provides necessary data support for subsequent community discovery.
[0108] After constructing the similarity matrix, the AICD model is proposed. This model is a community discovery model based on autoencoders. Its loss function takes into account the idea of non-negative matrix decomposition. It captures the potential features between nodes or entities by learning the low-dimensional representation of the data, thereby more effectively dividing the community structure.
[0109] Finally, an optimization algorithm for the AICD model is proposed, and the principle of the optimization algorithm, the parameter adjustment strategy, and how to effectively converge to the optimal solution are introduced. Optimization algorithms play a vital role in the model training process. They help adjust the model parameters to better fit the data and improve the performance of community detection.
[0110] 2.1 Construction of Similarity Matrix
[0111] In social networks, similarity matrix is a powerful tool for analyzing the similarity between nodes in the network, which can be used for tasks such as clustering, community detection and link prediction. The similarity measurement can be based on different characteristics of the network, such as node degree, shortest path, etc. Here, the neighbors of the node are considered as the similarity measure, and a similarity matrix S = [S ij ]∈R |V|*|V| , where the element S ij represents the similarity between node i and node j, and V is a set of nodes.
[0112] Since the Dice coefficient is usually used to compare the similarity between two sets, it is selected as a method to measure node similarity. Assume that there are two sets A and B, and the calculation of the Dice similarity coefficient between them is shown in formula (1):
[0113]
[0114] Where |A| and |B| represent the number of elements in sets A and B respectively, and A∩B represents the intersection of A and B. Similarly, the Dice similarity coefficient is used to represent the network, and the number of neighbors of a node is regarded as a similarity measure. Then the similarity between node i and node j is shown in formula (2):
[0115]
[0116] Where N represents the number of neighbors of a node, N(v i )∩N(v j ) represents the number of common neighbors of node i and node j. The value of Dice ranges from 0 to 1. If the two sets are the same, the Dice similarity coefficient is 1; if the two sets are disjoint, the Dice similarity coefficient is 0. Next, a simple example will be provided to illustrate the construction process of the similarity matrix, such as Figure 3 shown.
[0117] 2.2 AICD model
[0118] The AICD (AutoEncoder for Improved Community Detection) model proposed in this paper is based on the autoencoder structure and takes into account the non-negative matrix decomposition technology, providing an effective method for community detection tasks. Figure 4 shown.
[0119] The similarity matrix S is the input of the model. The introduction of the encoder enables the similarity matrix to be mapped into a low-dimensional community member space. Specifically, the encoder extracts the potential features between nodes, so that the low-dimensional representation can more accurately reflect the community structure. The decoder uses the low-dimensional representation to restore the topological structure of the original network through inverse operations. The design of this structure enables the model to better understand the input data and capture the key information of the data during the dimensionality reduction process.
[0120] In this model, the form of the objective function encompasses the idea of non-negative matrix factorization, which aims to learn a more compact and informative community member space representation by minimizing the difference between the similarity matrix and the product of two low-rank matrices. This process is carried out under the non-negativity constraint, thus allowing the results to be explained intuitively. The form of the objective function is shown in formula (3):
[0121]
[0122] This objective function is the sum of three different terms, each of which has a different role. D is the reconstruction error term used to quantify the difference between the original matrix S and the matrix X1X2…X q Y qThe difference between the approximations derived from the product. This term ensures the fidelity of the model to the original data, thus providing an accurate representation of the network structure.
[0123] The second item J E is a sparse regularization term that encourages the features learned by the model to be sparse. This sparsity plays a key role in the model because it allows for easier feature selection and interpretation. With sparser features, each community is expected to have a set of salient features, making the community structure more interpretable.
[0124] Item 3J reg is a smoothing regularization term, which, when multiplied by the hyperparameter α, ensures that adjacent features have similar values, which is crucial for preserving the local structure of the data. Specifically, since adjacent nodes are more likely to belong to the same community, and αJ reg Ensure that the model does not produce overly fragmented community structures by requiring adjacent nodes to have similar feature representations.
[0125] These three factors guide the AICD model to accurately reflect the network community structure while maintaining interpretability and structural coherence. The balance between them can be adjusted using the hyperparameter α, allowing the performance of the model to be fine-tuned according to the characteristics of the network.
[0126] 2.3 Objective Function Optimization
[0127] In order to optimize the objective function, we need to find X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints. Since the non-negativity constraints in the objective function and the product of multiple matrices make this a non-convex optimization problem, traditional convex optimization methods cannot be directly applied to this problem.
[0128] A general approach to solving this type of non-convex optimization problem is to use the block coordinate descent (BCD) method, which alternately updates X i and Y q matrices while keeping the other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems.
[0129] Given X i Matrix, we can solve the following optimization problem by formula (4) to update Y q :
[0130]
[0131] Among them, Y q ≥0.
[0132] Similarly, each X can be updated by solving the following optimization problem i matrix:
[0133]
[0134] For all i, it is required that X i ≥0.
[0135] Various optimization techniques can be used to solve these subproblems, such as multiplicative update rules derived from the Karush-Kuhn-Tucker (KKT) conditions for non-negative matrix factorization. These update rules ensure that X i and Y q The non-negativity of the matrix and decreases the value of the objective function in each iteration.
[0136] 2.4X i and Y q The update rule
[0137] X can be derived according to the multiplication update rule i and Y q The update rules are derived from the Karush-Kuhn-Tucker (KKT) conditions for non-negative matrix factorization. The objective function without constraints is defined as F(X i ,Y q ), then the Lagrangian function of the problem can be written as follows:
[0138]
[0139] Among them, H i and G q is the Lagrange multiplier associated with the non-negativity constraint, and <·,·> represents the Frobenius inner product. The KKT condition for this problem is:
[0140]
[0141] According to these conditions, we can derive X i and Y q The multiplicative update rule for :
[0142]
[0143] Here ⊙ represents element-wise multiplication, and the division symbol represents element-wise division.
[0144] 2.5 Community Discovery Algorithm Based on AICD
[0145] The present invention constructs an AICD model through an autoencoder structure for community discovery tasks. The overall algorithm of the model is shown below.
[0146] Algorithm 1: Community discovery algorithm based on AICD
[0147] Input: similarity matrix S, regularization parameter α, maximum number of iterations T
[0148] Output: Matrix X i and the matrix Y q
[0149]
[0150] The input of the algorithm is the similarity matrix constructed by the present invention, the regularization parameter α, and the maximum number of iterations T. First, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
[0151] The above algorithm has convergence guarantees due to the properties of the multiplicative update rule. Specifically, the update rule is derived from the Main Minimax (MM) algorithm, which ensures that the objective function J is monotonically decreasing in each iteration. As a result, the algorithm will converge to a stable point, which is a local minimum due to the non-convexity of the problem. In practice, the algorithm usually converges quite fast, although it can be slower in theory in the worst case. However, a good initialization can significantly speed up the convergence process.
[0152] 3. Experiment and result analysis
[0153] This section first introduces the experimental dataset and comparison algorithm used. The dataset is a public dataset, and the selected comparison algorithm is representative in community discovery. The evaluation indicators of the experiment use the indicators commonly used in community discovery, standardized mutual information and adjusted Rand coefficient. Then the setting of experimental parameters is explained, and the convergence and parameter sensitivity are proved. Then, comparative experiments are carried out under the same experimental environment. Finally, the experimental results are presented and the community structure is visualized. The results show that the AICD model achieves good results for community discovery.
[0154] 3.1 Experimental Dataset
[0155] Since users under the social network event topic have no labels, that is, the real communities to which users belong, the present invention uses four public data sets (KIT, Enron, GN_data, LFR_data) for experiments to verify the effectiveness of the model proposed in the present invention. The data sets selected by the present invention have been widely studied and used in the field of community discovery, and have high availability and representativeness. Moreover, their community structures are relatively obvious, which is convenient for comparison and verification.
[0156] The KIT dataset is obtained from the KIT information system and records the interaction relationships of users within the department over a period of time. By identifying 170 active users belonging to 35 communities, their interaction relationships are constructed into an adjacency matrix, generating a network containing 897 edges.
[0157] The Enron dataset is a record of emails sent by employees of Enron Corporation. A node represents a user's email address. If user i sends an email to user j, an undirected edge between user i and user j is established. A simplified dataset with 150 nodes, 34 edges, and 68 communities is used.
[0158] GN_data is based on the GN network benchmark proposed by Girvan and Newman, containing 128 nodes, 3365 edges and 4 communities, with the average degree set to 52.
[0159] LFR_data is a complex network dataset based on the benchmark dataset generator LFR_benchmark_graph. In this dataset, k represents the average degree of connectivity between each node and its external nodes in the community, and the range of k is set to {0.1, 0.2, 0.3, 0.4, 0.5}. Despite the different connectivity, the number of nodes, edges, and communities in each dataset is the same, which are 1000, 7692, and 36, respectively.
[0160] 3.2 Introduction to comparison method
[0161] In order to prove that the model proposed in this paper can learn better community similarities between nodes, that is, the classification of communities is more accurate, six representative methods are selected for comparison, namely NNSED, LabelPropagation, BigClam, DYNMOGA, Modularity-Aware GAE / VGAE and DANMF, as shown below:
[0162] NNSED (Non-negative Symmetric Encoder-Decoder): The NNSED model uses a non-negative symmetric encoder-decoder approach for community detection. It explicitly integrates the decoder and encoder into a unified loss function, and implicitly achieves the sparsity of node representation through symmetry and non-negative properties.
[0163] LabelPropagation: LabelPropagation is a semi-supervised learning algorithm that predicts the labels of unlabeled nodes by propagating known labels to neighboring nodes based on the connectivity of graph data.
[0164] BigClam: The BigClam algorithm is based on maximum likelihood estimation and reveals the community structure in the network by maximizing the membership of a node in the community.
[0165] DYNMOGA: The DYNMOGA algorithm combines the concepts of genetic algorithm and multi-objective optimization. By designing effective evolutionary operators and fitness evaluation methods, it can search for different community structures and find the optimal solution through multi-objective optimization.
[0166] MGV (Modularity-Aware GAE / VGAE): The MGV model is based on the GAE and VGAE encoders. It considers the embedding space based on the graph structure and modularity prior communities, and then mixes the GAE and VGAE encoders for training and optimization.
[0167] DANMF: The DANMF model is a method similar to deep autoencoders that considers an encoder component and extracts community members from the adjacency matrix, but is unable to learn deeper node representations.
[0168] AICD: The AICD model is an AutoEncoder-based model proposed in this paper. It combines non-negative matrix factorization to minimize the difference between the similarity matrix and the product of two low-rank matrices, and learns a more compact and information-rich community member space representation.
[0169] 3.3 Evaluation indicators
[0170] To evaluate the quality of community discovery, the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) were used, both of which are included in the sklearn toolkit.
[0171] (1)NMI
[0172] NMI is an indicator to measure the similarity between two sets. Its value ranges from 0 to 1. The closer the value is to 1, the closer the community division result is to the real community structure. Its definition is as follows:
[0173]
[0174] Among them, A represents the clustering result, that is, the result of community division, and B represents the real community label. I(A,B) represents the mutual information between A and B, and H(A) and H(B) represent the entropy of community A and the real label B respectively. The calculation formula of mutual information I(A,B) is as follows (10):
[0175]
[0176] Where k represents the number of clusters and m represents the number of categories. ijIt is the number of samples whose clustering result is category i and whose true label is category j divided by the total number of samples. i and p j They represent the number of samples whose clustering results are category i and whose true labels are category j divided by the total number of samples. The entropy H(A) is calculated as shown in formula (11):
[0177]
[0178] (2) ARI
[0179] ARI (Automated Readability Index), also known as the Adjusted Rand Index, is another evaluation indicator used to measure clustering results. The value range of ARI is between -1 and 1, and a larger value indicates that the clustering result is closer to the actual situation. Its definition is as shown in formula (12):
[0180]
[0181] Among them, RI is the Rand Index and Expected is the mathematical expectation.
[0182] 3.4 Experimental parameter setting
[0183] The experimental parameters mainly include the dimensions of each layer of the model, the hyperparameter α, and the random seed. For layers, the dimensions are set differently depending on the number of nodes and communities. For example, for a KIT network with 170 nodes and 35 communities, layers are usually set to [170, 128, 64, 35].
[0184] As for the hyperparameter α, it is a regularization parameter for the graph regularizer. The graph regularizer can improve the utilization of graph structure information in community detection tasks by imposing regularization constraints on graph data. A graph regularizer that is too large will blur the boundaries of the community, causing the node features within the community to be too similar, while the differences between communities are not obvious, which makes the model unable to accurately identify the true community structure. On the contrary, a graph regularizer that is too small imposes minimal constraints on the similarity or smoothness between nodes, resulting in overfitting of the training data and inability to effectively model and learn the graph structure. Based on multiple experiments, the range of adjusting α is set to {10 -4 ,10 -3 ,10 -2 ,10 -1}.
[0185] from Figure 5 In the figure, we can see that for the KIT and Enron datasets, the hyperparameter α has little effect on the NMI and ARI indicators. This shows that the AICD model is not sensitive to the hyperparameter α, indicating its robustness. Similar results were obtained for the other two datasets.
[0186] At the same time, in order to verify whether the AICD model gradually approaches the optimal solution or stable state during the training process and to evaluate the quality of the optimization algorithm, a convergence analysis was performed and a loss function curve was plotted. To ensure the consistency of the experimental conditions, the hyperparameter α was set to 0.01 and trained for 100 cycles. Figure 6 It can be observed that both the KIT network and the ERRON network converged. In terms of convergence speed, the KIT network basically converged in about 20 cycles, while the ERRON network showed similar results, converging at about 25 cycles, and the other two data sets were similar.
[0187] 3.5 Experimental Results Analysis
[0188] Experiments were conducted on the introduced datasets in the same experimental environment. In order to fairly compare the performance of each algorithm, each experiment was run 20 times and the results were averaged.
[0189] Table 1 Experimental results of non-overlapping communities
[0190]
[0191] The standardized mutual information NMI and adjusted Rand coefficient ARI scores of each model in the overlapping community dataset are shown in Table 1. It can be seen that the AICD model achieved better results.
[0192] Figure 7 and 8 The results of NMI and ARI of different models for non-overlapping community detection are shown. For the NMI index, the AICD model outperforms other algorithms on the Enron, KIT and GN_data datasets, indicating that the method proposed in this paper is more accurate in mining community structures. However, for the ARI index, the performance of the AICD model on the Enron dataset is slightly lower than that of the DYNMOGA algorithm, with a difference of only 0.07.
[0193] Fig. 9 and Fig.10The evaluation results of NMI and ARI indicators on the overlapping community LFR network are shown respectively. The x-axis of the line graph represents the average connectivity k of the LFR network, ranging from 0.1 to 0.5. The AICD algorithm performs very well in both NMI and ARI indicators. Although almost all algorithms show a significant decline when k=0.5, the AICD model still achieves good results, but slightly lower than MGV. This is because the model takes into account the modularity structure of the graph, and better considers the modularity structure of the graph when learning the representation of the graph, which makes the model more accurate in overlapping network detection, especially when there are obvious overlapping communities in the graph.
[0194] Fig. 9 The community discovery results of AICD are shown. Nodes with the same community use the same color. The degree of a node is related to its size. The larger the degree of a node, the larger the node size is set.
[0195] For the GN_data dataset, the average degree of each node is very high, so we choose to use a spring layout to simulate the behavior of a spring system and adjust the position of the node through the spring force and repulsion between the nodes. For the overlapping community LFR dataset, the overlap degree ranges from 0.1 to 0.5, and one of them is selected for the result display.
[0196] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of a method for discovering user groups in a social network.
[0197] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of a method for discovering user groups of social networks.
[0198] An application embodiment of the present invention provides an information data processing terminal, which includes a social network user group discovery system.
[0199] Social media platforms: used to analyze and understand the community structure and interaction patterns of users on social media, so as to optimize content recommendations, advertising placement and social function design.
[0200] E-commerce platform: Help e-commerce platforms identify and utilize user group information, improve personalized recommendation systems, and enhance user purchasing experience and transaction conversion rates.
[0201] Online education platform: used to analyze student group structure and learning behavior, improve personalized teaching plans, and improve learning outcomes and user retention rates.
[0202] Market research and social analysis: used for sociological research, market analysis and public opinion monitoring, revealing group behaviors and trends, and providing decision support and forecasting capabilities.
[0203] The method was compared on the non-overlapping community Enron, KIT and GN_data datasets and the overlapping community LFR network. On the non-overlapping community, for the NMI index, the AICD model performs better than other algorithms, indicating that the method proposed in this invention is more accurate in mining community structure. The performance of the ARI index on the Enron dataset is slightly lower than that of the DYNMOGA algorithm, with a difference of only 0.07.
[0204] On the overlapping community LFR network, AICD's algorithm performs very well in terms of both NMI and ARI indicators. Although almost all algorithms show a significant decline when k = 0.5, the AICD model still achieves good results, but slightly lower than MGV. This is because the model takes into account the modularity structure of the graph and better considers the modularity structure of the graph when learning the representation of the graph, which makes the model more accurate in overlapping network detection, especially when there are obvious overlapping communities in the graph.
[0205] The following are two specific application examples showing how to use the above social network user group discovery method:
[0206] Example 1: Social network platform user community detection
[0207] 1) Data collection and preprocessing:
[0208] Collect user data on a social networking platform, including follow-up relationships and interactions between users (such as likes and comments).
[0209] A similarity matrix between users is constructed, the Dice coefficient is used as the similarity measure, and the users’ neighbors (i.e., followers and followed) are considered as the basis for similarity measure.
[0210] 2) Apply AICD model:
[0211] The similarity matrix is used as the input of the AICD model.
[0212] Set the regularization parameter and the maximum number of iterations.
[0213] Initialize the non-negative matrices Xi and Yq to random positive values.
[0214] The Xi and Yq matrices are iteratively updated using the block coordinate descent method until convergence.
[0215] 3) Community discovery and result analysis:
[0216] Based on the output of the AICD model, different user communities in the social network are identified.
[0217] Analyze the characteristics of each community, such as user interests, behavior patterns, etc.
[0218] This information is used to provide strategic support for user recommendations, content distribution, etc. on social networking platforms.
[0219] Example 2: Segmentation of e-commerce platform user groups
[0220] 1) Data collection and preprocessing:
[0221] Collect user purchase history, browsing behavior, evaluation and other data on e-commerce platforms.
[0222] Construct a similarity matrix between users and use the Dice coefficient to consider the similarity of users' purchasing behavior and browsing history.
[0223] 2) Apply AICD model:
[0224] The similarity matrix is input into the AICD model.
[0225] Set appropriate regularization parameters and maximum number of iterations.
[0226] Initialize the non-negative matrices Xi and Yq and start the iterative update process.
[0227] Monitor the convergence of the objective function to ensure that the model reaches a stable state.
[0228] 3) User group segmentation and marketing strategy formulation:
[0229] Based on the output of the AICD model, e-commerce platform users are divided into different groups.
[0230] Analyze each group's purchasing preferences, consumption habits and other characteristics.
[0231] Develop targeted marketing strategies, such as personalized recommendations, coupon distribution, etc., to improve user satisfaction and purchase conversion rate.
[0232] These two embodiments demonstrate how to apply the above user group discovery method on different social networks and e-commerce platforms, and realize the effective division and utilization of user groups by constructing a similarity matrix, applying the AICD model, and performing result analysis and strategy formulation.
[0233] In the description of the present invention, unless otherwise specified, "plurality" means two or more than two; the orientations or positional relationships indicated by the terms "upper", "lower", "left", "right", "inner", "outer", "front end", "rear end", "head", "tail", etc. are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", "third", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0234] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.
[0235] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.
Claims
1. A method for discovering user groups in a social network, characterized in that: include: S1, Similarity matrix construction: In social networks, similarity matrices are used for clustering, community detection, and link prediction. The similarity measurement is based on different characteristics of the network, including node degree and shortest path; S2, AICD model: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks; S3, objective function optimization: In order to optimize the objective function, we need to find X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems; S4,X i and Y q Update rule: According to the multiplication update rule, derive X i and Y q The update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization; S5, community discovery algorithm based on AICD: The input of the algorithm is the constructed similarity matrix, the regularization parameter α, and the maximum number of iterations T; first, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
2. The method for discovering user groups in a social network according to claim 1, characterized in that: S1 specifically includes: Consider the neighbors of a node as a similarity measure and construct a similarity matrix S = [S ij ]∈R |V|*|V| , where the element S ij represents the similarity between node i and node j, V is the node set; Since the Dice coefficient is usually used to compare the similarity between two sets, it is selected as a method to measure node similarity; assuming there are two sets A and B, the calculation of the Dice similarity coefficient between them is shown in formula (1): Where |A| and |B| represent the number of elements in sets A and B respectively, and A∩B represents the intersection of A and B. Similarly, the Dice similarity coefficient is used to represent the network, and the number of neighbors of a node is regarded as a similarity measure. Then the similarity between node i and node j is shown in formula (2): Where N represents the number of neighbors of a node, N(v i )∩N(v j ) represents the number of common neighbors of node i and node j; the value of Dice ranges from 0 to 1; if the two sets are the same, the Dice similarity coefficient is 1; if the two sets are disjoint, the Dice similarity coefficient is 0.
3. The method for discovering user groups in a social network according to claim 1, characterized in that: S2 specifically includes: The similarity matrix S is the input of the model. The introduction of the encoder enables the similarity matrix to be mapped into a low-dimensional community member space. Specifically, the encoder extracts the potential features between nodes, so that the low-dimensional representation can more accurately reflect the community structure. The decoder uses the low-dimensional representation to restore the topological structure of the original network through inverse operations. The form of the objective function is shown in formula (3): The first item J D is the reconstruction error term used to quantify the difference between the original matrix S and the matrix X1X2…X q Y q Differences between product-derived approximations; The second item J E is a sparse regularization term that encourages the features learned by the model to be sparse. This sparsity plays a key role in the model because it allows for easier feature selection and interpretation. With sparser features, each community is expected to have a set of significant features, making the community structure more interpretable. Item 3J reg is a smoothing regularization term, which multiplies the hyperparameter α to ensure that adjacent features have similar values, which is crucial for preserving the local structure of the data; specifically, since adjacent nodes are more likely to belong to the same community, and αJ reg Ensure that the model does not produce overly fragmented community structures by requiring adjacent nodes to have similar feature representations.
4. The method for discovering user groups in a social network according to claim 1, characterized in that: S3 specifically includes: Given X i Matrix, we can solve the following optimization problem by formula (4) to update Y q : Among them, Y q ≥0; Similarly, each X can be updated by solving the following optimization problem i matrix: For all i, it is required that X i ≥0; Various optimization techniques can be used to solve these subproblems, such as multiplicative update rules derived from the Karush-Kuhn-Tucker (KKT) conditions for non-negative matrix factorization; these update rules ensure that X i and Y q The non-negativity of the matrix and decreases the value of the objective function in each iteration.
5. The method for discovering user groups in a social network as claimed in claim 1, characterized in that: S4 specifically includes: The objective function without constraints is defined as F(X i ,Y q ), then the Lagrangian function of the problem can be written as follows: Among them, H i and G q is the Lagrange multiplier associated with the non-negativity constraint, <·,·> represents the Frobenius inner product, and the KKT condition for this problem is: According to these conditions, we can derive X i and Y q The multiplicative update rule for : Where A represents element-wise multiplication and the division symbol represents element-wise division.
6. A social network user group discovery system implementing the social network user group discovery method according to any one of claims 1 to 5, characterized in that: include: Similarity matrix building module: In social networks, similarity matrix is used for clustering, community detection and link prediction. The similarity measurement is based on different characteristics of the network, including node degree and shortest path; AICD model building module: The AICD model is based on the autoencoder structure and considers non-negative matrix factorization technology to provide a method for community discovery tasks; Objective function optimization module: In order to optimize the objective function, we need to find the X that minimizes the objective function. i and Y q The optimal value of the matrix while satisfying the given constraints; using the block coordinate descent method, alternately update X i and Y q matrices while keeping other matrices fixed. This approach transforms the original complex problem into a series of more easily solvable subproblems; Matrix update module: derive X according to the multiplication update rule i and Y q The update rules for , which are derived from the Karush-Kuhn-Tucker conditions for non-negative matrix factorization; Community discovery algorithm module based on AICD: The input of the algorithm is the constructed similarity matrix, regularization parameter α, and the maximum number of iterations T; First, the non-negative matrix X i and Y q Initialize with random positive values, and then iteratively update X according to the proposed optimization algorithm i and Y q , until convergence.
7. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the social network user group discovery method according to any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the method for discovering user groups in a social network as claimed in any one of claims 1 to 5.
9. An information data processing terminal, comprising the social network user group discovery system according to claim 6.
Citation Information
Cited By
Advertisement data statistical method based on deep learning
CN120258905A
