A social network community discovery method based on user values

Abstract

This invention relates to the fields of complex network analysis and social network data mining, and provides a community discovery method based on user values, comprising the following steps: S1: constructing user-topic dual-encoding representations; S2: decoupling representations based on value hierarchy structure; S3: value semantic alignment constraints; S4: user value representation generation and optimization; S5: hybrid similarity construction and node centrality calculation; S6: label update strategy based on hybrid centrality ranking; S7: label propagation driven by multi-level influence; S8: community attribute center construction and fuzzy node identification; S9: fuzzy node re-attribution correction based on community attribute centers to solve the problems of difficulty in obtaining stable user value representations under low / no corpus conditions, and difficulty in effectively introducing value orientation as node attributes into community partitioning, thereby achieving stable community partitioning based on user values.

Description

Technical Field

[0001] This invention relates to the technical fields of complex network analysis and social network data mining, and particularly to a method for discovering social network communities based on user values. By modeling and utilizing the semantic attribute representations of user nodes, it achieves automated discovery and analysis of user group structures, applicable to scenarios such as social networks, public opinion analysis, and user behavior analysis. Background Technology

[0002] Community detection in complex networks is a crucial task for analyzing the structure of systems such as social networks, web networks, and citation networks. Traditional topology-based community detection methods (such as the KL algorithm, the Girvan-Newman algorithm, and the Louvain method) are effective in scenarios with clear network structures and high modularity, but they often fail to achieve stable partitioning results when the network is sparse or the community boundaries are ambiguous (Kernighan and Lin, 1970; Girvan and Newman, 2002; Blondel et al., 2008). The label propagation algorithm (LPA) is widely used due to its near-linear time complexity, but its reliance on majority voting of neighbor labels makes the results highly sensitive to the initial state and update order, resulting in insufficient stability in networks with many boundary nodes or weak structural signals (Raghavan et al., 2007). With the widespread emergence of networks with node semantic attributes, some studies have attempted to introduce node attribute information into the community detection process to improve the accuracy of community partitioning. For example, the SAS-LP algorithm enhances semantic consistency by fusing structural and attribute information during label propagation, while the NILPA algorithm further considers the combined impact of topology and node attributes on label updates during the propagation phase (Berahmand et al., 2020; Zhao et al., 2025). In addition, there are methods that simultaneously model network topology and node attributes through deep representation learning (such as CDBNE), but these methods typically rely on complex model training, incurring significant computational overhead and high parameter tuning costs (Zhou et al., 2023).

[0003] On the other hand, in the field of user modeling and public opinion analysis, most existing research on user morality and value modeling is based on user-generated text, social relationships, or historical behaviors, and has achieved certain results in scenarios with sufficient text (Lin et al., 2008; Liu et al., 2020; Qi et al., 2023). However, the above methods generally rely on explicit text expressions or observable social background information. When the user's speech frequency is low or there is a lack of stable text records, the modeling effect often drops significantly. In recent years, some research has begun to introduce factors such as values and moral orientations into the user modeling process (Bataeva et al., 2021; Jing et al., 2024; Paul et al., 2024), attempting to characterize the user's orientation on different value dimensions. However, most methods still mainly rely on indirect language cues and are difficult to stably learn the user value representation solely based on attitude response signals without relying on the user's text content.

[0004] In summary, there are mainly two limitations in the existing technology: 1) In the scenario of low or even no text data, it is difficult to stably obtain the user value representation, which limits the ability to characterize users at the value level; 2) The existing community discovery methods are difficult to effectively introduce the user value orientation as a node attribute into the community division process, resulting in a community structure that is difficult to form a clear, interpretable, and consistent division result in the value semantic space. Therefore, it is necessary to propose a method that can learn a stable user value representation without relying on the user's text content and use this representation as a node attribute to participate in community division, so as to achieve the mining of user group structure and community discovery oriented to the value dimension. Summary of the Invention

[0005] In view of the above problems, the purpose of the present invention is to provide a social network community discovery method based on user values, which can learn a stable and semantically interpretable user value representation under low or even no text data conditions, and effectively introduce the value representation as a node attribute into the community division process; through integrating the similarity of value attributes and network structure information for community propagation and update, it realizes the consistent community division of user groups in the value dimension, thereby solving the problems in the existing technology that it is difficult to stably obtain the user value representation and the value orientation is difficult to participate in the community division as a node attribute, and the community division method can also be applied to the general community division data set.

[0006] The above object of the present invention is achieved by the following technical solutions: A social network community discovery method based on user values, comprising the following steps: S1: Construct user-topic double-encoding representation: Based on the attitude response data of users to different topics, construct potential representations from the user perspective and the topic perspective respectively, and learn the initial representations of users and topics in the value space through a double-encoder structure; S2: Representation decoupling based on the value hierarchy structure: Based on the potential representations from the user perspective and the topic perspective, introduce the value hierarchy structure constraints to decouple and model the potential representations, so as to enhance the distinguishability between different value dimensions and maintain their semantic organizational relationships; S3: Value semantic alignment constraints: Based on the predefined value label system, align and constrain the topic potential representation with the explicit value semantics, guiding the model to learn potential representations consistent with the value semantics; S4: Generation and optimization of user value representation: Jointly model the potential representations from the user perspective and the topic perspective through the decoder structure, and overall optimize the model based on the attitude response prediction task to obtain a stable and semantically interpretable user value representation; S5: Construction of hybrid similarity and calculation of node centrality: Integrate the evidence of node topological structure and semantic attribute evidence to construct a node hybrid similarity index, and calculate the node centrality based on the hybrid similarity to characterize the comprehensive influence ability of nodes in the network, providing a sorting basis for the subsequent label propagation process; S6: Label update strategy based on sorting by hybrid centrality: Sort the nodes in the network according to the node centrality, and update the node labels in sequence according to the sorting results, enabling nodes with greater influence to participate in the propagation first, guiding the labels to stably evolve in regions with higher local consistency; S7: Multi-order influence-driven label propagation: Consider the influence of both the first-order neighbors and the second-order neighbors of nodes during the label propagation process. By introducing a multi-order influence propagation mechanism based on distance attenuation, regulate the propagation intensity of different-order neighborhoods, reduce the randomness in the label propagation process, and enhance the stability of the community structure; S8: Construction of community attribute centers and identification of fuzzy nodes: After the initial label propagation ends, construct the prototype representations of the attributes of each community, and identify fuzzy nodes such as small-scale community nodes and marginal nodes based on the community scale and node neighborhood characteristics; S9: Reassignment and correction of fuzzy nodes based on community attribute centers: By calculating the similarity between fuzzy nodes and the attribute centers of each community, rejudge and correct the community membership of fuzzy nodes, optimize the initial community division results, and improve the accuracy and stability of overall community discovery.

[0007] Furthermore, step S1 is specifically as follows: For each user , we model the distribution of their value orientations through their attitude responses to multiple topics , assume their true posterior distribution is , and establish an approximate posterior distribution through the inference network , where For parameters; We set up users Victory potential representation Satisfying the prior of a multivariate standard normal distribution: in for An identity matrix of order 1. Indicates the number of dimensions in the value label space; Its corresponding approximate posterior distribution is defined as: The mean with standard deviation By user encoder generate, The ordinal index for the dimension of the potential representation; Regarding the topic We base our analysis on the column vectors in the user-topic attitude response matrix. Constructing its underlying value representation This indicates the activation status of the topic across various value tags, among which To represent the number of value label dimensions, we assume... Follows a multidimensional Bernoulli prior distribution: in These are hyperparameters used to approximate the posterior distribution. We construct a reasoning network Used to model its variational posterior distribution in As a parameter, therefore, Victory potential representation The approximate posterior is defined as: in By topic encoder This generates potential representations of users and topics.

[0008] Furthermore, step S2 specifically includes: Introducing total relevance (TC) and KL divergence as regularization terms guides the latent dimensions to be as independent as possible, representing the latent user values. As input, its TC decoupling constraint Defined as: in, Let KL divergence be the KL divergence. This represents the posterior distribution after aggregation. This indicates the total number of users. Indicates user Potential representation The marginal posterior distribution on the k-th dimension constrains redundant information among latent factors, which helps improve the independence and interpretability of the representation; Meanwhile, to prevent latent variables from deviating far from the standard prior, we introduce a dimension-wise KL divergence regularization: in ; However, considering that value labels are not completely independent but organized in a hierarchical structure, completely equivalently constraining the independence of all potential dimensions would ignore this structural information. Therefore, we propose a hierarchical value structure decoupling strategy, which guides latent variables to maintain appropriate local relevance and global separation in semantic organization based on the partitioning structure of different levels in the label tree. Let the value label tree be the first A layer contains several non-overlapping subdomains, each subdomain corresponding to a set of latent variable dimensions. The structure of this layer is denoted as follows: Then the decoupling regularization term guided by this hierarchical structure is defined as: in, Indicates the first The number of subdomains included in the hierarchical value tag structure. Indicates the first The first in the layer Potential representations over each labeled subdomain Indicates user Potential representation In the The first in the layer Tag subdomain The marginal posterior distribution on the formula allows the latent representation to retain appropriate relevance within the same semantic subdomain, while maintaining strong structural independence between different value semantic subdomains, thereby enhancing the representation's ability to characterize the hierarchical structure of values.

[0009] Furthermore, step S3 specifically includes: A value tag alignment module is introduced to supervise and align the output of the topic value encoder with the explicit value tag system. A topic Using the topic-value tag matrix Q provided by the first A topic Corresponding explicit value label vector As a supervisory signal, it guides its potential representation. In semantic space, it is close to the real label, where For topic encoders targeting the first A topic Output Value-based label activation vector; The mean squared error (MSE) is used as the supervised loss function, defined as follows: Where M represents the total number of topics. The activation probabilities generated for the topic encoder. By assigning value labels to the model, this supervision term directly imposes semantic constraints on the latent value space, guiding the model to learn representations that conform to the value system structure, thereby improving the model's structural expressive ability and attitude response prediction performance.

[0010] Furthermore, step S4 specifically includes: After obtaining the latent value representations of users and topics, the decoder module is used to predict the user's attitude response to a specific topic, i.e., support or opposition, defined as a triple: ,in, Indicates user Topic The true attitude reaction labels are 1 for support and 0 for opposition. Our objective function is defined as: in, express function, This represents the binary cross-entropy loss; The design follows the monotonicity assumption in cognitive modeling in the field of intelligent learning, which states that the higher a user's activity level is on a certain cognitive dimension, the more likely they are to support topics and content that are consistent with their value orientation. By continuously minimizing the total loss function until the loss can no longer be reduced, the final effect is achieved, and the user's value representation is obtained. The total loss function is shown in the following formula: in, These represent all trainable parameters in the model, corresponding to the user encoder and topic encoder, respectively. , To control the hyperparameters of alignment supervision and prior regularization, For the first The weights of the layer structure decoupling loss This indicates the hierarchical depth of the value labeling system. The loss is the value decoupling loss corresponding to the i-th layer.

[0011] Furthermore, step S5 specifically includes: Node hybrid similarity The calculation process is shown in the following formula: in and Representing nodes respectively and nodes Structural and attribute evidence, This is a moderating coefficient for structural evidence, used to control the contribution of structural information to the overall similarity. Specifically: Structural evidence - node topological similarity Using extended neighborhood-based Similarity, as structural evidence, is calculated using the following formula: in, and Let u and v represent the sets of their neighbors, including themselves. express and The number of nodes in the intersection. express and The number of nodes in a cluster is a metric that effectively represents the consistency of a node's neighborhood. Attribute Evidence: Node Attribute Similarity By calculating the absolute differences between nodes across each attribute dimension, converting them into similarity scores, and then averaging them, the overall similarity of nodes at the attribute level is comprehensively measured. and nodes The attribute similarity between them is calculated using the following formula: in, Indicates the number of dimensions of node attributes. Represents a node Attribute vectors, Represents a node The first attribute vector Dimensional attribute features, Represents a node and In the first The similarity of attribute features across all dimensions is calculated using a formula that not only captures minute changes in attribute values ​​but also, by dividing by the total number of attribute dimensions... This ensured The value range of is strictly limited to [0,1], and larger values ​​are not allowed. This indicates that two nodes have a higher similarity in node semantic attributes.

[0012] Furthermore, step S6 specifically includes: By combining node similarity A node centrality metric is constructed to measure the relative importance of nodes in the similarity propagation process. The node centrality is calculated using the following formula: in Represents a node The formula considers the similarity between a node and its neighbors at both structural and attribute levels, thus reflecting the node's potential influence in the local network. Although the above constructs a hybrid similarity index by fusing topological and attribute information, when applied to real social networks, the scale-free nature of network topology must be taken into account. To eliminate computational interference caused by significant degree differences, the accumulated result must be optimized. Normalization corrections are performed to ensure fair comparison of nodes of different sizes on the same level, and the centrality of nodes is calculated using the following formula: in The maximum value of node centrality in the network is represented by the logarithmic transformation. The value of node centrality is compressed to a finite interval, thereby effectively alleviating the problem of high-connectivity nodes overdoing the propagation process. Thus, the value of each node can be calculated. Value, based on the node The value determines the final node update order.

[0013] Furthermore, step S7 specifically includes: After obtaining the centrality order of each node, considering the varying distances between nodes during influence propagation, and recognizing that the connection strength weakens as the distance between nodes increases, the algorithm assumes the connection strength between two nodes is related to their distance and defaults to a node importance of 1. The connection strength halves with each unit increase in distance. The connection weight based on the shortest path between nodes is defined by the following formula and denoted as... : in, Represents two nodes and The shortest path length between nodes shows that when the distance between nodes is relatively large, their connection weight is small. Based on this, the propagation influence of neighbors on the target node is defined by the following formula: Represents a node The degree, For nodes centrality, For nodes With nodes Mixed similarity between them; From the above formula, the propagation influence of first-order neighbors on the target node can be defined as follows: in, Represents a node It is a node The first-order neighbor, i.e. , Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node The degree of a first-order neighbor is such that if the first-order neighbor has a high degree of similarity to the target node, then this neighbor node has a greater influence on the target node. The propagation influence of second-order neighbors on the target node is defined by the following formula: in, Represents a node It is a node The second-order neighbor, i.e. That is, the target node The neighbor's neighbor, Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node As described in the above description of the calculation of node influence propagation, the degree of the connection between nodes gradually weakens as the distance between them increases. Therefore, the influence of second-order neighbors on the target node will be further weakened. The target node's new community label is selected based on the influence of its first-order and second-order neighbors, taking the label of the neighborhood with the highest influence as its new community label, after the maximum number of iterations. After that, the initial community was formed.

[0014] Furthermore, step S8 specifically includes: The initial community was obtained by using the community attribute center. To represent the overall attribute characteristics within a community, the following formula is defined: in The attribute vector representing node u. Represents any community. Used to describe the overall feature representation of a community in the attribute space, the formula constructs the community's attribute prototype vector by aggregating the attribute vectors of all nodes within the community. This community attribute center can comprehensively reflect the common features of nodes within the community at the attribute level and, to a certain extent, reduce the impact of individual node attribute noise on the community representation. By determining the threshold of small communities To identify all the small community nodes and add them to... In the list, to identify community edge nodes, the community boundary nodes are defined as follows: in, Represents a node The set of first-order neighbors, Representative node With nodes The shortest distance between them, as can be seen from the formula above, means that edge nodes are those nodes that belong to and are connected to other community nodes of that node. All edge nodes will be added to... In the list.

[0015] Furthermore, step S9 specifically includes: From the obtained community attribute center Further calculation of nodes With the community The attribute similarity between nodes is used to measure the degree to which a node belongs to a community in the attribute space. Cosine similarity is used for measurement, and its calculation is shown in the following formula: in The value range is [-1, 1]. The larger the value, the more likely it is to be a node. Attributes and community The more similar the overall attribute distribution, the more effective the method is in directly characterizing the relationship between nodes and the overall attribute features of the community, unlike the node-to-node similarity calculation method. This effectively reduces the impact of community size differences on the similarity calculation results. By using a fuzzy node reassignment strategy based on community attribute centers, the node-community attribute similarity is calculated, and the community with the highest similarity value is finally selected as the reassignment community for the target node. The calculation process is shown in the following formula: in It is a node and community The formula selects the community tag with the highest similarity value as the node based on the similarity of the community attribute centers. The re-attribution community is then used to obtain the final community division result after the re-attribution mechanism.

[0016] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) User value representation modeling based on attitude response: Through user-topic dual encoder, value decoupling and semantic alignment mechanism, user value representation is constructed using only user attitude response information to topic. This effectively solves the problem that traditional methods rely heavily on text content and background information and are difficult to cover users with low corpus. It provides a stable and semantically interpretable node attribute representation for user group structure analysis.

[0017] (2) Community discovery enhanced by community-centered representation of label propagation: By introducing community-centered representation in the label propagation process and combining multi-level neighborhood influence propagation and fuzzy node reassignment correction mechanism, the problem of easy misclassification of nodes in the community boundary area is effectively alleviated, the error accumulation caused by mispropagation is reduced, and the divided communities have higher internal consistency and clearer community separation structure in terms of attribute characteristics. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the social network community discovery method based on user values ​​of the present invention; Figure 2 This is a schematic diagram of the method for obtaining user value representation in steps S1-S4 of the present invention; Figure 3 This is a schematic diagram of the community division method in steps S5-S9 of the present invention; Figure 4 This is a schematic diagram of the A-LPA ablation experiment results of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0021] First Embodiment like Figure 1-3 As shown, this embodiment provides a method for discovering social network communities based on user values, including the following steps: S1: Construct user-topic dual-encoding representations: Based on user attitude response data to different topics, construct latent representations from the user's perspective and the topic's perspective respectively, and learn the initial representations of users and topics in the value space through the dual encoder structure.

[0022] In this embodiment, step S1 specifically includes: For each user We observed their attitudes and reactions on multiple topics. To model the distribution of its value orientation, we assume that its true posterior distribution is And through reasoning networks Establish an approximate posterior distribution ,in For parameters; We set up users Victory potential representation Satisfying the prior of a multivariate standard normal distribution: in for An identity matrix of order 1. Indicates the number of dimensions in the value label space; Its corresponding approximate posterior distribution is defined as: The mean with standard deviation By user encoder generate, The ordinal index for the dimension of the potential representation; Regarding the topic We base our analysis on the column vectors in the user-topic attitude response matrix. Constructing its underlying value representation This indicates the activation status of the topic across various value tags, among which To represent the number of value label dimensions, we assume... Follows a multidimensional Bernoulli prior distribution: in These are hyperparameters used to approximate the posterior distribution. We construct a reasoning network Used to model its variational posterior distribution in As a parameter, therefore, Victory potential representation The approximate posterior is defined as: in By topic encoder This generates potential representations of users and topics.

[0023] Step S1, "Constructing a User-Topic Dual Encoding Representation," is based on user attitude response data for different topics. For each user, their value orientation is characterized by their attitudes towards multiple topics. First, an initial abstract feature conforming to standard patterns is assigned to the user. Then, a corresponding encoding tool is used to generate specific features that reflect the user's values. For each topic, based on the user group's attitude data, abstract value features corresponding to the topic are constructed (to reflect the topic's tendency across various value dimensions). Similarly, an initial pattern distribution is assigned to the topic, and then a specialized encoding tool is used to generate the topic's value features. Ultimately, this approach allows for the simultaneous acquisition of initial abstract features for both users and topics in the value dimension, providing a foundation for subsequent value modeling.

[0024] S2: Value Hierarchy-Based Representation Decoupling: Based on the potential representations from the user's perspective and the topic's perspective, a value hierarchy constraint is introduced to decouple and model the potential representations, thereby enhancing the distinguishability between different value dimensions and maintaining their semantic organization.

[0025] In this embodiment, step S2 specifically includes: Introducing total relevance (TC) and KL divergence as regularization terms guides the latent dimensions to be as independent as possible, representing the latent user values. As input, its TC decoupling constraint Defined as: in, Let KL divergence be the KL divergence. This represents the posterior distribution after aggregation. This indicates the total number of users. Indicates user Potential representation The marginal posterior distribution on the k-th dimension constrains redundant information among latent factors, which helps improve the independence and interpretability of the representation; Meanwhile, to prevent latent variables from deviating far from the standard prior, we introduce a dimension-wise KL divergence regularization: in ; However, considering that value labels are not completely independent but organized in a hierarchical structure, completely equivalently constraining the independence of all potential dimensions would ignore this structural information. Therefore, we propose a hierarchical value structure decoupling strategy, which guides latent variables to maintain appropriate local relevance and global separation in semantic organization based on the partitioning structure of different levels in the label tree. Let the value label tree be the first A layer contains several non-overlapping subdomains, each subdomain corresponding to a set of latent variable dimensions. The structure of this layer is denoted as follows: Then the decoupling regularization term guided by this hierarchical structure is defined as: in, Indicates the first The number of subdomains included in the hierarchical value tag structure. Indicates the first The first in the layer Potential representations over each labeled subdomain Indicates user Potential representation In the The first in the layer Tag subdomain The marginal posterior distribution on the formula allows the latent representation to retain appropriate relevance within the same semantic subdomain, while maintaining strong structural independence between different value semantic subdomains, thereby enhancing the representation's ability to characterize the hierarchical structure of values.

[0026] Step S2, "representation decoupling based on value hierarchy," further optimizes the initial user and topic value features. First, two types of constraint rules are introduced: one guides the features of each dimension of the value system to be as independent as possible, reducing information redundancy and making each dimension more distinctive; the other prevents these features from deviating from a reasonable range, ensuring their standardization. However, because values ​​themselves have a hierarchical structure (e.g., a large value category contains multiple sub-dimensions), making all dimensions completely independent would ignore this hierarchical relationship. Therefore, a hierarchical optimization strategy is designed: based on the hierarchical division of values, features within the same level of sub-category maintain appropriate correlation, while features between different sub-categories are separated as much as possible. This optimization preserves the semantic hierarchy of the values ​​themselves while enhancing the distinctiveness of features across different dimensions, making the value representation more reasonable and clear.

[0027] S3: Value Semantic Alignment Constraint: Based on a predefined value labeling system, the latent representation of the topic is aligned with the explicit value semantics, guiding the model to learn a latent representation consistent with the value semantics.

[0028] In this embodiment, step S3 specifically includes: A value tag alignment module is introduced to supervise and align the output of the topic value encoder with the explicit value tag system. A topic Using a topic-value tag matrix The first one provided in the middle A topic Corresponding explicit value label vector As a supervisory signal, it guides its potential representation. In semantic space, it is close to the real label, where For topic encoders targeting the first A topic Output Value-based label activation vector; The mean squared error (MSE) is used as the supervised loss function, defined as follows: Where M represents the total number of topics. The activation probabilities generated for the topic encoder. By assigning value labels to the model, this supervision term directly imposes semantic constraints on the latent value space, guiding the model to learn representations that conform to the value system structure, thereby improving the model's structural expressive ability and attitude response prediction performance.

[0029] The “value semantic alignment constraint” corresponding to step S3 is to make the value features of the topic fit the pre-set value label system: a special alignment module will be introduced here to match the value features generated by each topic with the explicit labels of the corresponding topics in the value label system. Using these explicit labels as reference standards, the value features of the topic are guided to be semantically closer to the real value definition.

[0030] Simultaneously, a bias-measuring method is used to supervise this process, minimizing the difference between topic features and standard labels. Through this constraint, the topic value features learned by the model will better conform to the pre-defined value system structure, strengthening the ability to express value semantics and improving the accuracy of subsequent predictions of user attitude responses.

[0031] S4: Generation and optimization of user value representation: The latent representations from the user perspective and the topic perspective are jointly modeled through the decoder structure, and the model is optimized as a whole based on the attitude response prediction task to obtain a stable and semantically interpretable user value representation.

[0032] In this embodiment, step S4 specifically includes: After obtaining the latent value representations of users and topics, the decoder module is used to predict the user's attitude response to a specific topic, i.e., support or opposition, defined as a triple: ,in, Indicates user Topic The true attitude reaction labels are 1 for support and 0 for opposition. Our objective function is defined as: in, express function, This represents the binary cross-entropy loss; The design follows the monotonicity assumption in cognitive modeling in the field of intelligent learning, which states that the higher a user's activity level is on a certain cognitive dimension, the more likely they are to support topics and content that are consistent with their value orientation. By continuously minimizing the total loss function until the loss can no longer be reduced, the final effect is achieved, and the user's value representation is obtained. The total loss function is shown in the following formula: in, These represent all trainable parameters in the model, corresponding to the user encoder and topic encoder, respectively. , To control the hyperparameters of alignment supervision and prior regularization, For the first The weights of the layer structure decoupling loss This indicates the hierarchical depth of the value labeling system. The loss is the value decoupling loss corresponding to the i-th layer.

[0033] Step S4, "User Value Representation Generation and Optimization," is the final step in obtaining usable user value features. First, using the previously obtained user and topic value features, a decoder module predicts the user's attitude (support or opposition) towards each topic. This prediction follows a cognitive logic: the stronger a user's inclination in a certain value dimension, the more likely they are to support topics aligned with that value orientation. Then, the entire model is optimized by integrating various errors: these include deviations between the predicted attitudes and the actual situation, as well as errors from constraints such as value alignment and hierarchical decoupling in previous steps. By continuously reducing these combined errors until the model's performance no longer improves, stable user value features that clearly reflect value semantics are finally obtained.

[0034] S5: Hybrid Similarity Construction and Node Centrality Calculation: By integrating node topological evidence and semantic attribute evidence, a hybrid similarity index for nodes is constructed, and node centrality is calculated based on the hybrid similarity to characterize the comprehensive influence of nodes in the network and provide a ranking basis for the subsequent label propagation process.

[0035] In this embodiment, step S5 specifically includes: Node hybrid similarity The calculation process is shown in the following formula: in and Representing nodes respectively and nodes Structural and attribute evidence, This is a moderating coefficient for structural evidence, used to control the contribution of structural information to the overall similarity. Specifically: Structural evidence - node topological similarity Using extended neighborhood-based Similarity, as structural evidence, is calculated using the following formula: in, and Let u and v represent the sets of their neighbors, including themselves. express and The number of nodes in the intersection. express and The number of nodes in a cluster is a metric that effectively represents the consistency of a node's neighborhood. Attribute Evidence: Node Attribute Similarity By calculating the absolute differences between nodes across each attribute dimension, converting them into similarity scores, and then averaging them, the overall similarity of nodes at the attribute level is comprehensively measured. and nodes The attribute similarity between them is calculated using the following formula: in, Indicates the number of dimensions of node attributes. Represents a node Attribute vectors, Represents a node The first attribute vector Dimensional attribute features, Represents a node and In the first The similarity of attribute features across all dimensions is calculated using a formula that not only captures minute changes in attribute values ​​but also, by dividing by the total number of attribute dimensions... This ensured The value range of is strictly limited to [0,1], and larger values ​​are not allowed. This indicates that two nodes have a higher similarity in node semantic attributes.

[0036] Step S5, "Construction of Hybrid Similarity and Calculation of Node Centrality," combines two types of node information to build a comprehensive index: on the one hand, it considers the nodes' connectivity (topology) within the network, measuring their structural similarity by comparing the overlap of their neighbor ranges; on the other hand, it considers the nodes' value-based semantic features (attributes), calculating the similarity of nodes across various value dimensions and averaging the results to obtain attribute-level similarity. These two types of similarity are then fused in a certain proportion to obtain the comprehensive similarity between nodes. Based on this comprehensive similarity, the comprehensive influence (centrality) of each node in the network is calculated, which serves as the ranking criterion for subsequent label propagation, prioritizing nodes with higher influence in the propagation process.

[0037] S6: Label update strategy based on mixed centrality ranking: Nodes in the network are ranked according to their centrality, and node labels are updated sequentially according to the ranking results, so that nodes with greater influence participate in the propagation first, and the labels are guided to evolve stably in regions with high local consistency.

[0038] In this embodiment, step S6 specifically includes: By combining node similarity A node centrality metric is constructed to measure the relative importance of nodes in the similarity propagation process. The node centrality is calculated using the following formula: in Represents a node The formula considers the similarity between a node and its neighbors at both structural and attribute levels, thus reflecting the node's potential influence in the local network. Although the above constructs a hybrid similarity index by fusing topological and attribute information, when applied to real social networks, the scale-free nature of network topology must be taken into account. To eliminate computational interference caused by significant degree differences, the accumulated result must be optimized. Normalization corrections are performed to ensure fair comparison of nodes of different sizes on the same level, and the centrality of nodes is calculated using the following formula: in The maximum value of node centrality in the network is represented by the logarithmic transformation. The value of node centrality is compressed to a finite interval, thereby effectively alleviating the problem of high-connectivity nodes overdoing the propagation process. Thus, the value of each node can be calculated. Value, based on the node The value determines the final node update order.

[0039] Step S6, corresponding to the "label update strategy based on mixed centrality ranking," first determines the propagation priority of nodes. It combines the previous comprehensive similarity score, measuring a node's influence (centrality) within the local network by assessing the overall similarity between the node and its neighbors. However, in real-world networks, the number of connections between nodes varies greatly. To prevent nodes with excessive connections from dominating propagation, this influence value is processed by using a logarithmic transformation to compress it to a reasonable range, allowing for a fair comparison of the influence of different nodes. Finally, nodes are ranked according to their processed influence, prioritizing nodes with higher influence in label propagation. This guides labels to spread more stably in highly similar local areas.

[0040] S7: Multi-order influence-driven label propagation: The influence of first-order and second-order neighbors of a node is considered simultaneously during the label propagation process. By introducing a multi-order influence propagation mechanism based on distance decay, the propagation intensity of different order neighborhoods is regulated, reducing the randomness in the label propagation process and enhancing the stability of the community structure.

[0041] In this embodiment, step S7 specifically includes: After obtaining the centrality order of each node, considering the varying distances between nodes during influence propagation, and recognizing that the connection strength weakens as the distance between nodes increases, the algorithm assumes the connection strength between two nodes is related to their distance and defaults to a node importance of 1. The connection strength halves with each unit increase in distance. The connection weight based on the shortest path between nodes is defined by the following formula and denoted as... : in, Represents two nodes and The shortest path length between nodes shows that when the distance between nodes is relatively large, their connection weight is small. Based on this, the propagation influence of neighbors on the target node is defined by the following formula: Represents a node The degree, For nodes centrality, For nodes With nodes Mixed similarity between them; From the above formula, the propagation influence of first-order neighbors on the target node can be defined as follows: in, Represents a node It is a node The first-order neighbor, i.e. , Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node The degree of a first-order neighbor is such that if the first-order neighbor has a high degree of similarity to the target node, then this neighbor node has a greater influence on the target node. The propagation influence of second-order neighbors on the target node is defined by the following formula: in, Represents a node It is a node The second-order neighbor, i.e. That is, the target node The neighbor's neighbor, Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node As described in the above description of the calculation of node influence propagation, the degree of the connection between nodes gradually weakens as the distance between them increases. Therefore, the influence of second-order neighbors on the target node will be further weakened. The target node's new community label is selected based on the influence of its first-order and second-order neighbors, taking the label of the neighborhood with the highest influence as its new community label, after the maximum number of iterations. After that, the initial community was formed.

[0042] Step S7, "Multi-order Influence-Driven Label Propagation," considers the influence of directly connected neighbors (first-order) and their neighbors' neighbors (second-order) simultaneously during label propagation. It adjusts the influence strength based on the distance between nodes; the greater the distance, the weaker the mutual influence. The influence of a neighbor on the target node is calculated by combining the strength attenuation due to distance, the neighbor's own network influence, and the combined similarity between the two. First-order neighbors (directly connected nodes) have their influence calculated according to corresponding rules; second-order neighbors (neighbors' neighbors) have a further weakened influence due to their greater distance. By simultaneously considering the influence of these two orders of neighbors, random fluctuations during label propagation can be reduced, resulting in a more stable final community structure.

[0043] S8: Community Attribute Center Construction and Fuzzy Node Identification: After the initial label propagation is completed, construct the attribute prototype representation of each community, and identify fuzzy nodes such as small-scale community nodes and edge nodes based on community size and node neighborhood characteristics.

[0044] In this embodiment, step S8 specifically includes: The initial community was obtained by using the community attribute center. To represent the overall attribute characteristics within a community, the following formula is defined: in The attribute vector representing node u. Represents any community. Used to describe the overall feature representation of a community in the attribute space, the formula constructs the community's attribute prototype vector by aggregating the attribute vectors of all nodes within the community. This community attribute center can comprehensively reflect the common features of nodes within the community at the attribute level and, to a certain extent, reduce the impact of individual node attribute noise on the community representation. By determining the threshold of small communities To identify all the small community nodes and add them to... In the list, to identify community edge nodes, the community boundary nodes are defined as follows: in, Represents a node The set of first-order neighbors, Representative node With nodes The shortest distance between them, as can be seen from the formula above, means that edge nodes are those nodes that belong to and are connected to other community nodes of that node. All edge nodes will be added to... In the list.

[0045] Step S8, "Community Attribute Center Construction and Fuzzy Node Identification," further refines the community structure after the initial label propagation. First, an attribute center is constructed for each community. By aggregating the attribute vectors of all nodes within the community, an attribute prototype representing the overall characteristics of the community is obtained. This more clearly reflects the commonalities of the community at the attribute level and reduces the interference of individual node attribute noise on the overall community characteristics. Next, two special types of fuzzy nodes are identified: one type is nodes in small communities, where a preset threshold is used to determine which communities are too small, and nodes within these small communities are marked; the other type is community edge nodes, which are nodes that belong to one community but whose directly connected neighbors belong to other communities. Identifying these fuzzy nodes provides a more accurate basis for subsequent community optimization.

[0046] S9: Fuzzy Node Reassignment Correction Based on Community Attribute Centers: By calculating the similarity between fuzzy nodes and each community attribute center, the community affiliation of fuzzy nodes is re-determined and corrected, optimizing the initial community division results and improving the accuracy and stability of overall community discovery.

[0047] In this embodiment, step S9 specifically includes: From the obtained community attribute center Further calculation of nodes With the community The attribute similarity between nodes is used to measure the degree to which a node belongs to a community in the attribute space. Cosine similarity is used for measurement, and its calculation is shown in the following formula: in The value range is [-1, 1]. The larger the value, the more likely it is to be a node. Attributes and community The more similar the overall attribute distribution, the more effective the method is in directly characterizing the relationship between nodes and the overall attribute features of the community, unlike the node-to-node similarity calculation method. This effectively reduces the impact of community size differences on the similarity calculation results. By using a fuzzy node reassignment strategy based on community attribute centers, the node-community attribute similarity is calculated, and the community with the highest similarity value is finally selected as the reassignment community for the target node. The calculation process is shown in the following formula: in It is a node and community The formula selects the community tag with the highest similarity value as the node based on the similarity of the community attribute centers. The re-attribution community is then used to obtain the final community division result after the re-attribution mechanism.

[0048] Step S9, "Fuzzy Node Reassignment Correction Based on Community Attribute Centers," is a step to precisely optimize the initial community division. After identifying fuzzy nodes, we calculate the attribute similarity between each fuzzy node and each community attribute center. This similarity measures the degree to which the node's own attributes match the overall attribute characteristics of the community. Then, each fuzzy node is reassigned to the community with the highest attribute similarity. This reassignment correction allows the community division to better reflect the true attribute distribution of the nodes, effectively reducing the interference of differences in community size on the division results, ultimately leading to a more accurate and stable community division.

[0049] Second Embodiment This embodiment aims to verify whether the method can segment stable and distinguishable user communities in the representation space, thereby illustrating the effectiveness of the method for "user group structure mining". The X-Stance dataset does not provide explicit connections between users, but the community segmentation method of this patent requires a graph structure as input. To conduct experiments on this dataset, this paper uses the user value representation vectors learned in steps S1-S3 as node attributes, and employs... The k-nearest neighbor (kNN) strategy constructs a user similarity graph: cosine similarity is used to measure the proximity between user representations, and for each user, its nearest neighbor (kNN) is connected to the nearest neighbor (kNN) graph. ( =5) most similar neighbors, thus generating weighted similarity edges. Considering that the edge relationship is induced by the representation space, there may be repetition of source information and the influence of noisy edges. In the experiment, the weight of the topological term in the structure-attribute fusion similarity is set to a small value ( =0.01), ensuring that community partitioning is primarily driven by node attribute relationships while retaining weak neighborhood consistency constraints to improve stability. The comparison method chosen is K-means, with the same number of communities set for the output partitioning results. Since X-Stance does not include real community labels, this paper uses the unsupervised metric Silhouette coefficient to evaluate partitioning quality. The distance metric used is cosine distance; a larger Silhouette coefficient indicates more compact intra-cluster clusters and better inter-cluster separability.

[0050] On the X-Stance dataset, our method and K-means achieved Silhouette scores of 0.6382 and 0.4103, respectively. Our method achieved a higher Silhouette score, indicating that it can form a more compact and discriminative user community structure within the user value representation space. This is mainly due to two factors: First, using KNN-based local neighborhoods provides a reasonable candidate range for label propagation, allowing label updates to focus on users with similar representations, thus making it easier to aggregate groups with similar value orientations. Second, the fuzzy node reassignment mechanism proposed in our method can re-determine and correct users at community boundaries with unclear membership after propagation, reducing the possibility of boundary nodes being misassigned to adjacent communities. This further enhances consistency within communities and clarifies separation between communities, ultimately reflecting an improved Silhouette score.

[0051] This invention compares the performance of the general community partitioning method described in steps S5 to S9 using six real-world network datasets widely used in the field of community detection. Table 1 shows the specific information of the six real-world datasets.

[0052] (1) Cornell, Texas, Washington, and Wisconsin: The above four datasets are all derived from the WebKB University Webpage Network. Nodes in the network represent webpages on the official websites of universities, and edges between nodes represent hyperlinks between webpages. Each node is represented by a 1703-dimensional text attribute vector to characterize the content information of the webpage. Based on the functional category to which the webpage belongs, the nodes are divided into 5 different communities. The node size of the four datasets ranges from 187 to 265, and the network structure is relatively sparse, making it suitable for evaluating the performance of community detection algorithms in small-scale networks with node attributes.

[0053] (2) Cora: Cora is a widely used citation network dataset where nodes represent academic papers and edges represent citation relationships between papers. The dataset contains 2708 papers and 5429 citation links. Each paper is represented by a 1433-dimensional keyword vector, and its community label corresponds to the research field to which the paper belongs, containing a total of 7 different communities.

[0054] (3) Citeseer: Citeseer is also a citation network dataset, where nodes represent papers and edges represent citation relationships. This dataset contains 3312 papers and 4732 citation links. Each node is represented by a 3703-dimensional text attribute vector, and the papers are divided into 6 different research areas, corresponding to 6 communities.

[0055] Table 1: Detailed Information on Six Real-World Datasets Experiments were conducted on the above dataset and comparison methods, and the results are shown in Tables 2 and 3.

[0056] Table 2: Experimental results of different methods on different real-world network datasets for NMI Table 3: Experimental results of different methods on different real-world network datasets for ARI The experimental results in Tables 2 and 3 show that A-LPA(ours) performs stably across different types of real-world networks, achieving significantly better results than the comparison methods on some networks. Compared to traditional structural methods such as KL, GN, Louvain, and Infomap, our method maintains high levels of NMI and ARI on all datasets, demonstrating the advantages of joint structural and attribute modeling. In the Wisconsin network, our method shows a clear advantage, significantly outperforming NILPA and other comparison methods in both NMI and ARI. This is because the Wisconsin network has low modularity and relatively ambiguous community boundaries, leaving many nodes with uncertain affiliations after the propagation phase. Our method effectively adjusts boundary nodes and small-scale community nodes through re-allocation correction based on community attribute centers, thus significantly improving community detection performance. In the Cora and Citeseer citation networks, our method also achieves better results than other methods, significantly outperforming traditional structural methods and basic label propagation methods. Citation networks have larger node scales, more communities, and richer semantic information in node attributes, making them more prone to generating ambiguous nodes during propagation. A-LPA effectively mitigates the accumulation of structural propagation errors by performing attribute-driven reassignment correction on fuzzy nodes, demonstrating better stability and accuracy in complex network scenarios. However, in the Cornell, Texas, and Washington WebKB networks, the results of our proposed method are slightly lower than NILPA, but still show significant advantages compared to methods such as LPA and SAS-LP. This is mainly because these networks are relatively small in scale, highly modular, and have a clear community structure at the topological level, allowing for relatively stable partitioning results during the propagation phase. The number of fuzzy nodes is also relatively small. The fuzzy node reassignment mechanism introduced by A-LPA after propagation has limited gain in these types of networks, thus resulting in slightly inferior overall performance compared to NILPA.

[0057] To further verify the contribution of the "fuzzy node reassignment based on community attribute center" module in the A-LPA algorithm to the final community partitioning effect, this paper designs an ablation experiment. After removing the reassignment step based on community attribute center (denoted as NO-R), the NMI and ARI values ​​decreased to varying degrees on all datasets. The experimental results are as follows: Figure 4As shown in the figure, the decline was more pronounced on the Wisconsin network and citation networks such as Cora and Citeseer, with NMI values ​​decreasing by approximately 10.43%, 5.11%, and 6.47%, respectively, and ARI values ​​decreasing by approximately 7.56%, 4.51%, and 6.33%, respectively. In contrast, the performance decline was relatively smaller on the Cornell, Texas, and Washington networks, with NMI values ​​decreasing by approximately 2.78%, 3.53%, and 4.13%, respectively, and ARI values ​​decreasing by approximately 1.81%, 2.99%, and 4.2%, respectively.

[0058] The above results indicate that the "fuzzy node reassignment based on community attribute centers" mechanism significantly improves the quality of community partitioning. Especially when the network modularity is low and community boundaries are ambiguous, the propagation phase is more prone to generating boundary nodes with uncertain affiliations and small-scale community nodes. Therefore, the lack of reassignment correction leads to the accumulation of errors caused by mispropagation, resulting in a more significant performance degradation. Conversely, when the network modularity is high and community boundaries are clear, a relatively stable partitioning result can be obtained during the propagation phase, with fewer ambiguous nodes, thus limiting the gains brought by the reassignment module. Overall, ablation experiments show that the fuzzy node reassignment mechanism based on community attribute centers can effectively improve the stability and accuracy of A-LPA in scenarios with ambiguous community structures. Furthermore, even after removing this module from a structurally clear network, the algorithm still maintains good performance, indicating that the overall framework of A-LPA has a certain degree of robustness. The reassignment module is mainly responsible for improving the fine-grained boundary characterization capability in complex scenarios.

[0059] A computer-readable storage medium stores computer code that, when executed, performs the methods described above. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0060] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for discovering social network communities based on user values, characterized in that, Includes the following steps: S1: Construct user-topic dual-encoding representations: Based on user attitude response data to different topics, construct latent representations from the user's perspective and the topic's perspective respectively, and learn the initial representations of users and topics in the value space through the dual encoder structure; S2: Value hierarchy-based representation decoupling: Based on the potential representations from the user's perspective and the topic's perspective, a value hierarchy constraint is introduced to decouple and model the potential representations, thereby enhancing the distinguishability between different value dimensions and maintaining their semantic organization. S3: Value Semantic Alignment Constraint: Based on a predefined value labeling system, the latent representation of the topic is aligned with the explicit value semantics, guiding the model to learn a latent representation consistent with the value semantics; S4: Generation and optimization of user value representation: The latent representations of user perspective and topic perspective are jointly modeled through the decoder structure, and the model is optimized as a whole based on the attitude response prediction task to obtain a stable and semantically interpretable user value representation. S5: Hybrid Similarity Construction and Node Centrality Calculation: Integrate node topological structure evidence and semantic attribute evidence to construct a hybrid similarity index for nodes, and calculate node centrality based on the hybrid similarity to characterize the comprehensive influence of nodes in the network and provide a ranking basis for the subsequent label propagation process; S6: Label update strategy based on mixed centrality ranking: Rank the nodes in the network according to node centrality and update the node labels in turn according to the ranking results, so that the nodes with greater influence participate in the propagation first, and guide the labels to evolve stably in areas with high local consistency. S7: Multi-order influence-driven label propagation: The influence of first-order and second-order neighbors of a node is considered simultaneously during the label propagation process. By introducing a multi-order influence propagation mechanism based on distance decay, the propagation intensity of different order neighborhoods is regulated, reducing the randomness in the label propagation process and enhancing the stability of the community structure. S8: Community Attribute Center Construction and Fuzzy Node Identification: After the initial label propagation is completed, construct the attribute prototype representation of each community, and identify fuzzy nodes such as small-scale community nodes and edge nodes based on community size and node neighborhood characteristics; S9: Fuzzy Node Reassignment Correction Based on Community Attribute Centers: By calculating the similarity between fuzzy nodes and each community attribute center, the community affiliation of fuzzy nodes is re-determined and corrected, optimizing the initial community division results and improving the accuracy and stability of overall community discovery.

2. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S1 is as follows: For each user We observed their attitudes and reactions on multiple topics. To model the distribution of its value orientation, we assume that its true posterior distribution is And through reasoning networks Establish an approximate posterior distribution ,in For parameters; We set up users Victory potential representation Satisfying the prior of a multivariate standard normal distribution: in for An identity matrix of order 1. Indicates the number of dimensions in the value label space; Its corresponding approximate posterior distribution is defined as: The mean with standard deviation By user encoder generate, The ordinal index for the dimension of the potential representation; Regarding the topic We base our analysis on the column vectors in the user-topic attitude response matrix. Constructing its underlying value representation This indicates the activation status of the topic across various value tags, among which To represent the number of value label dimensions, we assume... Follows a multidimensional Bernoulli prior distribution: in These are hyperparameters used to approximate the posterior distribution. We construct a reasoning network Used to model its variational posterior distribution in As a parameter, therefore, Victory potential representation The approximate posterior is defined as: in By topic encoder This generates potential representations of users and topics.

3. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S2 is as follows: Introducing total relevance (TC) and KL divergence as regularization terms guides the latent dimensions to be as independent as possible, representing the latent user values. As input, its TC decoupling constraint Defined as: in, Let KL divergence be the KL divergence. This represents the posterior distribution after aggregation. This indicates the total number of users. Indicates user Potential representation The marginal posterior distribution on the k-th dimension constrains redundant information among latent factors, which helps improve the independence and interpretability of the representation; Meanwhile, to prevent latent variables from deviating far from the standard prior, we introduce a dimension-wise KL divergence regularization: in ; However, considering that value labels are not completely independent but organized in a hierarchical structure, completely equivalently constraining the independence of all potential dimensions would ignore this structural information. Therefore, we propose a hierarchical value structure decoupling strategy, which guides latent variables to maintain appropriate local relevance and global separation in semantic organization based on the partitioning structure of different levels in the label tree. Let the value label tree be the first A layer contains several non-overlapping subdomains, each subdomain corresponding to a set of latent variable dimensions. The structure of this layer is denoted as follows: Then the decoupling regularization term guided by this hierarchical structure is defined as: in, Indicates the first The number of subdomains included in the hierarchical value tag structure. Indicates the first The first in the layer Potential representations over each labeled subdomain Indicates user Potential representation In the The first in the layer Tag subdomain The marginal posterior distribution on the formula allows the latent representation to retain appropriate relevance within the same semantic subdomain, while maintaining strong structural independence between different value semantic subdomains, thereby enhancing the representation's ability to characterize the hierarchical structure of values.

4. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S3 is as follows: A value tag alignment module is introduced to supervise and align the output of the topic value encoder with the explicit value tag system. A topic Using a topic-value tag matrix The first one provided in the middle A topic Corresponding explicit value label vector As a supervisory signal, it guides its potential representation. In semantic space, it is close to the real label, where For topic encoders targeting the first A topic Output Value-based label activation vector; The mean squared error (MSE) is used as the supervised loss function, defined as follows: Where M represents the total number of topics. The activation probabilities generated for the topic encoder. By assigning value labels to the model, this supervision term directly imposes semantic constraints on the latent value space, guiding the model to learn representations that conform to the value system structure, thereby improving the model's structural expressive ability and attitude response prediction performance.

5. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S4 is as follows: After obtaining the latent value representations of users and topics, the decoder module is used to predict the user's attitude response to a specific topic, i.e., support or opposition, defined as a triple: ,in, Indicates user Topic The true attitude reaction labels are 1 for support and 0 for opposition. Our objective function is defined as: in, express function, This represents the binary cross-entropy loss; The design follows the monotonicity assumption in cognitive modeling in the field of intelligent learning, which states that the higher a user's activity level is on a certain cognitive dimension, the more likely they are to support topics and content that are consistent with their value orientation. By continuously minimizing the total loss function until the loss can no longer be reduced, the final effect is achieved, and the user's value representation is obtained. The total loss function is shown in the following formula: in, These represent all trainable parameters in the model, corresponding to the user encoder and topic encoder, respectively. , To control the hyperparameters of alignment supervision and prior regularization, For the first The weights of the layer structure decoupling loss This indicates the hierarchical depth of the value labeling system. The loss is the value decoupling loss corresponding to the i-th layer.

6. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S5 is as follows: Node hybrid similarity The calculation process is shown in the following formula: in and Representing nodes respectively and nodes Structural and attribute evidence, This is a moderating coefficient for structural evidence, used to control the contribution of structural information to the overall similarity. Specifically: Structural evidence - node topological similarity Using extended neighborhood-based Similarity, as structural evidence, is calculated using the following formula: in, and Let u and v represent the sets of their neighbors, including themselves. express and The number of nodes in the intersection. express and The number of nodes in a cluster is a metric that effectively represents the consistency of a node's neighborhood. Attribute Evidence: Node Attribute Similarity By calculating the absolute differences between nodes across each attribute dimension, converting them into similarity scores, and then averaging them, the overall similarity of nodes at the attribute level is comprehensively measured. and nodes The attribute similarity between them is calculated using the following formula: in, Indicates the number of dimensions of node attributes. Represents a node Attribute vectors, Represents a node The first attribute vector Dimensional attribute features, Represents a node and In the first The similarity of attribute features across all dimensions is calculated using a formula that not only captures minute changes in attribute values ​​but also, by dividing by the total number of attribute dimensions... This ensured The value range of is strictly limited to [0,1], and larger values ​​are not allowed. This indicates that two nodes have a higher similarity in node semantic attributes.

7. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S6 is as follows: By combining node similarity A node centrality metric is constructed to measure the relative importance of nodes in the similarity propagation process. The node centrality is calculated using the following formula: in Represents a node The formula considers the similarity between a node and its neighbors at both structural and attribute levels, thus reflecting the node's potential influence in the local network. Although the above constructs a hybrid similarity index by fusing topological and attribute information, when applied to real social networks, the scale-free nature of network topology must be taken into account. To eliminate computational interference caused by significant degree differences, the accumulated result must be optimized. Normalization corrections are performed to ensure fair comparison of nodes of different sizes on the same level, and the centrality of nodes is calculated using the following formula: in The maximum value of node centrality in the network is represented by the logarithmic transformation. The value of node centrality is compressed to a finite interval, thereby effectively alleviating the problem of high-connectivity nodes overdoing the propagation process. Thus, the value of each node can be calculated. Value, based on the node The value determines the final node update order.

8. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S7 is as follows: After obtaining the centrality order of each node, considering the varying distances between nodes during influence propagation, and recognizing that the connection strength weakens as the distance between nodes increases, the algorithm assumes the connection strength between two nodes is related to their distance and defaults to a node importance of 1. The connection strength halves with each unit increase in distance. The connection weight based on the shortest path between nodes is defined by the following formula and denoted as... : in, Represents two nodes and The shortest path length between nodes shows that when the distance between nodes is relatively large, their connection weight is small. Based on this, the propagation influence of neighbors on the target node is defined by the following formula: Represents a node The degree, For nodes centrality, For nodes With nodes Mixed similarity between them; From the above formula, the propagation influence of first-order neighbors on the target node can be defined as follows: in, Represents a node It is a node The first-order neighbor, i.e. , Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node The degree of a first-order neighbor is such that if the first-order neighbor has a high degree of similarity to the target node, then this neighbor node has a greater influence on the target node. The propagation influence of second-order neighbors on the target node is defined by the following formula: in, Represents a node It is a node The second-order neighbor, i.e. That is, the target node The neighbor's neighbor, Represents a node centrality, Represents a node and nodes Mixed similarity between Represents a node As described in the above description of the calculation of node influence propagation, the degree of the connection between nodes gradually weakens as the distance between them increases. Therefore, the influence of second-order neighbors on the target node will be further weakened. The target node's new community label is selected based on the influence of its first-order and second-order neighbors, taking the label of the neighborhood with the highest influence as its new community label, after the maximum number of iterations. After that, the initial community was formed.

9. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S8 is as follows: The initial community was obtained by using the community attribute center. To represent the overall attribute characteristics within a community, the following formula is defined: in The attribute vector representing node u. Represents any community. Used to describe the overall feature representation of a community in the attribute space, the formula constructs the community's attribute prototype vector by aggregating the attribute vectors of all nodes within the community. This community attribute center can comprehensively reflect the common features of nodes within the community at the attribute level and, to a certain extent, reduce the impact of individual node attribute noise on the community representation. By determining the threshold of small communities To identify all the small community nodes and add them to... In the list, to identify community edge nodes, the community boundary nodes are defined as follows: in, Represents a node The set of first-order neighbors, Representative node With nodes The shortest distance between them, as can be seen from the formula above, means that edge nodes are those nodes that belong to and are connected to other community nodes of that node. All edge nodes will be added to... In the list.

10. The social network community discovery method based on user values ​​according to claim 1, characterized in that, Step S9 is as follows: From the obtained community attribute center Further calculation of nodes With the community The attribute similarity between nodes is used to measure the degree to which a node belongs to a community in the attribute space. Cosine similarity is used for measurement, and its calculation is shown in the following formula: in The value range is [-1, 1]. The larger the value, the more likely it is to be a node. Attributes and community The more similar the overall attribute distribution, the more effective the method is in directly characterizing the relationship between nodes and the overall attribute features of the community, unlike the node-to-node similarity calculation method. This effectively reduces the impact of community size differences on the similarity calculation results. By using a fuzzy node reassignment strategy based on community attribute centers, the node-community attribute similarity is calculated, and the community with the highest similarity value is finally selected as the reassignment community for the target node. The calculation process is shown in the following formula: in It is a node and community The formula selects the community tag with the highest similarity value as the node based on the similarity of the community attribute centers. The re-attribution community is then used to obtain the final community division result after the re-attribution mechanism.

Images