Social network analysis method combining relationship perception and high-order heterogeneous graph convolution
By combining relation-aware and high-order heterogeneous graph convolutional social network analysis methods, a social network analysis model is constructed. This model solves the problems of structural redundancy, insufficient relation semantic modeling, and inadequate learning of high-order dependency information in heterogeneous graph neural networks in social network analysis. It achieves effective learning of heterogeneous graph structure and feature information, and improves the effectiveness of node representation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-03-27
AI Technical Summary
Existing heterogeneous graph neural networks suffer from structural redundancy or incompleteness, insufficient semantic modeling of relationships, inadequate learning of higher-order dependency information, and insufficient modeling of topological structure in social network analysis.
This paper proposes a social network analysis method that combines relation awareness and high-order heterogeneous graph convolution. By constructing a social network analysis model, including an input layer, a relation awareness nonlinear modeling module, a semantic awareness fusion module, a structure enhancement module, and a high-order heterogeneous graph convolution module, the graph structure is optimized using KAN networks and Markov processes to learn and predict node features.
It improves the effectiveness of node representation, solves the problems of incomplete structure, weak semantic modeling of relations, and insufficient learning of high-order features in social networks, and realizes effective learning of heterogeneous graph structure and feature information.
Smart Images
Figure CN121745929A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of graph neural networks, in particular to a social network analysis method combining relationship perception and high-order heterogeneous graph convolution. BACKGROUND
[0002] Graph neural networks have become an important tool for graph data mining due to their strong structure modeling capability. On this basis, heterogeneous graph neural networks (HGNNs) are proposed to support the modeling of complex semantic information of graph data with multiple types of nodes and multiple types of edges. Compared with traditional homogeneous graph neural networks, HGNNs have the following advantages: 1) strong relationship modeling capability: HGNNs can effectively integrate different types of relationship information through specialized relationship modeling mechanisms (such as relationship convolution and attention mechanism); 2) effective modeling of high-order semantic information: when dealing with complex relationships such as "user-item-tag" or "entity-relation-entity", HGNNs can learn the high-order semantic information implied by different complex relationships, thereby learning more effective node representations; 3) adaptable to multiple scene tasks: HGNNs have been widely used in node classification, link prediction, recommendation ranking, entity alignment, etc., showing good generalization ability and practicality; 4) strong scalability: due to its propagation mechanism based on graph structure, HGNNs can flexibly learn features of multi-hop neighbors, meta paths, and meta graphs, thereby learning multiple aspects of graph data features.
[0003] Although HGNNs have shown great potential in heterogeneous graph data modeling, there are still the following challenges in practical applications: 1) structure may be redundant or incomplete: real-world heterogeneous graphs often have problems such as missing connections or noisy edges, which can lead to insufficient feature learning or introduce noise, affecting the effectiveness of node representation learning; 2) insufficient relationship semantic modeling: some methods simplify the processing of different relationship types and fail to fully exploit the information implied by different relationship types; 3) insufficient learning of high-order dependency information: traditional graph convolution methods are usually limited to shallow propagation and cannot effectively model long-distance node features and high-order dependency information, affecting the representation learning effect; 4) deficiencies in topology structure modeling: although some methods try to enhance the graph structure through relationship models to improve the model effect, these topology structure optimization and enhancement methods still lack effective mechanisms to enhance the connections between similar nodes. SUMMARY
[0004] The purpose of the present application is to provide a social network analysis method combining relationship perception and high-order heterogeneous graph convolution, which aims to solve or improve at least one of the above technical problems.
[0005] To achieve the above purpose, the present application provides the following scheme: A social network analysis method combining relationship perception and high-order heterogeneous graph convolution, comprising: acquiring historical social network data and constructing heterogeneous graph data; the heterogeneous graph data is composed of nodes and adjacency matrices of different types of edges; wherein the nodes are users in the social network, and the adjacency matrices are relationships between different users, including friend relationships, family relationships and interaction relationships; constructing a social network analysis model combining relationship perception and high-order heterogeneous graph convolution; the social network analysis model comprises an input layer, a relationship perception nonlinear modeling module, a semantic perception fusion module, a structure enhancement module, a high-order heterogeneous graph convolution module and an output module; wherein the input layer is connected with the relationship perception nonlinear modeling module and the structure enhancement module respectively, the relationship perception nonlinear modeling module is connected with the semantic perception fusion module, the semantic perception fusion module and the structure enhancement module are also connected with the high-order heterogeneous graph convolution module, and the high-order heterogeneous graph convolution module is also connected with the output module; inputting the heterogeneous graph data into the social network analysis model for training, using the graph node classification indicators F1-macro and F1-micro to evaluate the feature learning performance of the model, using the trained model to predict the to-be-tested social network data, and obtaining a social network distribution diagram.
[0006] Optionally, the acquiring historical social network data and constructing heterogeneous graph data specifically comprises: acquiring historical social network data and performing standardization processing, extracting a node feature matrix of the heterogeneous graph and a set of adjacency matrices of different types of edges from the data after the standardization processing respectively ; wherein, is a relationship type serial number, t is the number of types of heterogeneous graph edges, is the number of nodes, N is the node feature dimension, F is a real number.
[0007] Optionally, in the social network analysis model: the input layer is configured to receive heterogeneous graph data; the relationship-aware nonlinear modeling module is configured to learn node features by constructing a nonlinear graph convolution layer using a KAN network based on adjacency matrices of different types of edges of the heterogeneous graph; the semantic-aware fusion module is configured to fuse the learned node representations based on different edge types to obtain fused node features; the structure enhancement module is configured to first transform the set of adjacency matrices, aggregate adjacency matrices of multiple types of edges, and then encode the overall structure of the heterogeneous graph using a Markov process to generate an adjacency matrix that fuses information of multiple types of edges of the heterogeneous graph; the high-order heterogeneous graph convolution module is configured to first use multiple propagation layers to aggregate features of the fused node features and the adjacency matrix that fuses information of multiple types of edges of the heterogeneous graph to extract information of different order neighbor nodes, then aggregate the representations output by different propagation layers by weighting, and finally generate a prediction result by linear transformation and a softmax function; and the output module is configured to output the prediction result.
[0008] Optionally, the processing procedure of the relationship-aware nonlinear modeling module includes: In the heterogeneous graph, each type of edge or relationship has a corresponding adjacency matrix due to the heterogeneity of the edges, which constitutes a set of adjacency matrices based on different types of edges , is equal to the number of types of edges of the heterogeneous graph, and the node feature matrix is represented as , is a real number, N is the number of nodes, F is the dimension of the node feature, and the subscript t represents the relationship type serial number; First, the node type feature transformation matrix is used to transform the features of different types of nodes to generate the transformed node features , d is the embedding dimension, and the transformation uses the formula: ; Subsequently, the KAN neural network is used to extract local structure and feature information according to the adjacency matrix of each relationship type A t ; wherein the KAN neural network constructs a relationship-aware mapping function by function decomposition to aggregate node features, generate node embeddings with rich relationship semantics, and the calculation formula is: wherein, is a cubic spline function, is a linear function for processing the mapped features, is an activation function RELU, and the subscripts q and p represent the identities of the functions, is an adjacency matrix of the relation type t , is a node feature matrix after transformation, is the output of the relation-aware nonlinear modeling module.
[0009] Optionally, the processing procedure of the semantic-aware fusion module includes: learning its corresponding weight for each relation type representation , which is formalized as: wherein, is a neural network for realizing weight learning, and the steps include: In order to learn the weight information of each relation type representation, each input representation is first linearly transformed by three learnable parameter matrices , , respectively, and the calculation formula is: , wherein, , and are outputs of the input representation obtained by linear transformation of the three parameter matrices, then the attention coefficients of and are calculated using the dot-product attention mechanism, and then the attention coefficients are normalized using softmax function to obtain the attention score of each relation type representation : Here, T is a matrix transposition operation, d is an embedding dimension, and then the attention scores of different type representations calculated by are used to fuse the representations of different relation types to obtain the fused node feature representation : .
[0010] Optionally, the processing procedure of the structure enhancement module includes: First, the adjacency matrix set of different relation types is transformed to obtain a global adjacency matrix : Then, the degree matrix H of Dright H Normalization is performed to obtain the normalized matrix. The structure enhancement module iteratively optimizes the topology of the graph through a Markov process to enhance the connectivity between nodes with similar structures. The calculation process is as follows: in, D yes H The degree matrix, It is after the first iteration k The adjacency matrix obtained is then expanded nonlinearly by introducing an expansion factor, as shown in the following formula: in, a As an extension factor, [ i , j ] represents the matrix index; then set the threshold. To reduce redundant information in the graph and retain the more meaningful connections: Finally, an enhanced adjacency matrix is generated. .
[0011] Optionally, the processing procedure of the higher-order heterogeneous graph convolution module includes: First, use the enhanced adjacency matrix. Propagating node features yields node representations at different layers, let , No. l The nodes of a layer are represented as follows: in, It is the first l Layer node representation, concatenating layer outputs to facilitate the aggregation of information from different layers: Here, Concat() represents the concatenation operation, and then a learnable parameter matrix is used. right Z Perform transformations to extract more meaningful representations: Then use the learnable parameter matrix and Aggregating information from different layers: in, tanh and ReLU It is a non-linear activation function, and is used last. softmax Normalizing the function yields the node embedding. : Embedding the resulting nodes into a feedforward network composed of two linear layers to output a distribution over the classes P : where, is a linear layer with learnable weights, denotes the predicted outcome.
[0012] Optionally, the social network analysis model employs a cross-entropy loss function to measure the error between the predicted outcome and the true label during the training process and updates the model parameters.
[0013] According to the specific embodiments provided by the present application, the following technical effects are disclosed: The application discloses a social network analysis method combining relationship perception and high-order heterogeneous graph convolution, which comprises the following steps: acquiring historical social network data and constructing heterogeneous graph data; constructing a social network analysis model combining relationship perception and high-order heterogeneous graph convolution; the social network analysis model comprises an input layer, a relationship perception nonlinear modeling module, a semantic perception fusion module, a structure enhancement module, a high-order heterogeneous graph convolution module and an output module; the heterogeneous graph data is input into the social network analysis model for training, the feature learning performance of the model is evaluated by using a graph node classification index F1-macro and F1-micro, the trained model is used to predict the to-be-tested social network data, and a social network distribution graph is obtained. The application can better solve the problems of incomplete structure, weak relationship semantic modeling and insufficient high-order feature learning in the social network, effectively learn the heterogeneous graph structure and feature information, and thus improve the effectiveness of node representation. BRIEF DESCRIPTION OF DRAWINGS
[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.
[0015] Figure 1 It is a step flow diagram of the social network analysis method of the present application; Figure 2 It is an operation logic diagram of the social network analysis model in the embodiment. DETAILED DESCRIPTION
[0016] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.
[0017] The present application aims to provide a social network analysis method combining relationship awareness and high-order heterogeneous graph convolution, aiming to solve or improve at least one of the above technical problems.
[0018] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.
[0019] As shown in Figures 1-2 , the present application provides a social network analysis method combining relationship awareness and high-order heterogeneous graph convolution, comprising: Step 100: obtaining historical social network data and constructing heterogeneous graph data; the heterogeneous graph data is composed of nodes and adjacency matrices of different types of edges; wherein the nodes are users in the social network, and the adjacency matrices are relationships between different users, including friend relationship, family relationship and interaction relationship.
[0020] The step specifically comprises: The present method selects the heterogeneous graph public benchmark data set IMDB in the field of social network analysis for model training, verification and testing. The data set is divided into three independent parts of training set, verification set and test set according to the ratio of 1:1:8, and is uniformly standardized. The node feature matrix of the heterogeneous graph and the adjacency matrix set of different types of edges are extracted from the data set, where t represents the serial number of the relationship type, is the type number of the heterogeneous graph edge, N is the number of nodes, F is the node feature dimension. The adjacency matrix set is represented by a tensor .
[0021] Step 200: constructing a social network analysis model combining relationship awareness and high-order heterogeneous graph convolution; the social network analysis model comprises an input layer, a relationship awareness nonlinear modeling module, a semantic awareness fusion module, a structure enhancement module, a high-order heterogeneous graph convolution module, and an output module; wherein the input layer is connected with the relationship awareness nonlinear modeling module and the structure enhancement module respectively, the relationship awareness nonlinear modeling module is connected with the semantic awareness fusion module, the semantic awareness fusion module and the structure enhancement module are both connected with the high-order heterogeneous graph convolution module, and the high-order heterogeneous graph convolution module is further connected with the output module.
[0022] The step specifically comprises: constructing a social network analysis model combining relationship awareness and high-order heterogeneous graph convolution according to the input layer, the relationship awareness nonlinear modeling module, the semantic awareness fusion module, the structure enhancement module, the high-order heterogeneous graph convolution module, and the output module.
[0023] 1) the relationship awareness nonlinear modeling module uses a Kolmogorov-Arnold Network (KAN) network to construct a nonlinear graph convolution layer on the basis of the adjacency matrix of different types of edges in the heterogeneous graph to learn the node features; 2) the semantic awareness fusion module fuses the learned node representations based on different edge types to obtain more comprehensive node features; 3) the structure enhancement module first transforms the adjacency matrix set to aggregate the adjacency matrices of multiple types of edges, and then uses a Markov process to encode the overall structure of the heterogeneous graph to generate an adjacency matrix that fuses the information of multiple types of edges of the heterogeneous graph; 4) the high-order heterogeneous graph convolution module first uses multiple propagation layers to aggregate the features of the heterogeneous graph to extract the information of different-order neighbor nodes, and then aggregates the representations output by different propagation layers by weighting to enable the model to effectively model long-distance node features and structures. Finally, the output of the high-order heterogeneous graph convolution module is linearly transformed and subjected to a softmax function to generate a predicted category.
[0024] 1) implementation of the relationship awareness nonlinear modeling module: In the heterogeneous graph, each type of edge or relationship has a corresponding adjacency matrix, which constitutes an adjacency matrix set based on different types of edges The adjacency matrix set can be simply written as a tensor , represents a real number, N represents the number of nodes, is equal to the number of types of edges of the heterogeneous graph. The node feature matrix can be represented as , F is the dimension of the node feature, and the subscript t represents the identification of different relationship types.
[0025] The heterogeneous relationship-aware KAN network is used for efficiently learning the complex interaction relationship between multiple types of nodes in a heterogeneous graph. The features of different types of nodes are transformed to extract more effective node representations, and the transformed node features are generated , which are transformed as follows: Then, according to the adjacency matrix of each relationship type , the KAN neural network is used to extract local structure and feature information, so as to achieve the effect of relationship-aware modeling. The KAN neural network constructs a relationship-aware mapping function by function decomposition to aggregate node features, thereby generating node embeddings with rich relationship semantics, and the calculation process is as follows: wherein is a unary mapping function, and here a cubic spline function is used, is a linear function for processing the mapped features, is an activation function RELU, and subscripts q and p represent the identity of the function, n represents the network hidden layer dimension. is the adjacency matrix of the relationship type t . is the node feature matrix after transformation, is the output of the relationship-aware nonlinear modeling module.
[0026] 2) Implementation of the semantic-aware fusion module: The relationship-aware nonlinear modeling module can learn the node representation under different relationship types, and these representations need to be fused to learn more comprehensive information of the heterogeneous graph. For this purpose, the semantic-aware fusion module is designed, which can adaptively learn the weight of each relationship type representation. Specifically, we learn the corresponding weight } for each relationship type representation }, which is formally represented as: wherein is a neural network for realizing weight learning, and the implementation steps of the neural network are described in detail as follows.
[0027] In order to learn the weight information of each relationship type representation, each input representation is first linearly transformed by three learnable parameter matrices , , respectively. The calculation process is as follows: , in , and The input representation is the output obtained through a linear transformation of three parameter matrices. Then, a dot product attention mechanism is used to compute... and The attention coefficient, and then use the attention coefficient. softmax The function is normalized to obtain the attention score for each relation type representation. : here, T This is a matrix transpose operation. d It's the embedding dimension. Then, attention scores are calculated based on different types of representations. The representations of different relation types are merged to obtain the merged node representation. : Through this process, the model can learn more comprehensive information about the heterogeneous graph based on the importance of each relationship, thereby obtaining more effective node representations.
[0028] 3) Implementation of the structural enhancement module: Since the original heterogeneous graph structure inevitably contains noise, graph structure enhancement is needed to optimize the structural information of the heterogeneous graph. To this end, a Structure Enhancement Module (SEM) is designed, which enhances the topological structure information through Markov processes.
[0029] In this module, the adjacency matrices of different relation types are first set. The transformation yields a global adjacency matrix. : Then use H degree matrix D right H Normalization yields The structure enhancement module iteratively optimizes the topology of the graph using a Markov process, thereby enhancing the connections between structurally similar nodes. The computation process is as follows: in, D yes H The degree matrix, It is after the first iteration kThe adjacency matrix is obtained. Next, an expansion factor is introduced to perform a nonlinear expansion operation on the adjacency matrix of the graph. The formula is as follows: wherein a is an expansion factor, i , j is the matrix index, that is, the value of the matrix in the first i row and the first j column. This operation is beneficial to enhance the connection between topologically similar nodes in the graph. Next, a threshold is set to reduce the redundant information in the graph and retain more meaningful connections in the graph: The structure enhancement module generates an enhanced adjacency matrix , which provides a basis for the implementation of the subsequent high-order heterogeneous graph convolution module.
[0030] 4) Implementation of the high-order heterogeneous graph convolution module: To further improve the modeling ability of the model on high-order node features, the present application constructs a high-order heterogeneous graph convolution module after completing semantic perception fusion and structure enhancement. This module aggregates the representations output by different propagation layers by weighting, so that the model can effectively model the dependency relationship of long-distance node features. The module first uses the enhanced adjacency matrix to propagate node features to obtain node representations of different layers, let , and the node representation of the first l layer is: wherein is the node representation of the first l layer. The outputs of the layers are spliced to facilitate the aggregation of information of different layers: wherein Concat() represents a splicing operation. Then, a learnable parameter matrix is used to transform Z to extract more meaningful representations: Then, a learnable parameter matrix and are used to aggregate information of different layers: wherein tanh and ReLU are nonlinear activation functions. Finally, the softmax function is used for normalization to obtain node embedding : The final node embeddings can be widely applied to downstream tasks of heterogeneous graph analysis. The final node embeddings are input into a feedforward network composed of two linear layers to obtain the distribution P over output classes where is a linear layer with learnable weights, denotes the predicted outcome. During training, we use cross-entropy loss as the loss function to measure the performance of the model and optimize the model parameters.
[0031] Step 300: input the heterogeneous graph data into the social network analysis model for training, and use the graph node classification indicators F1-macro and F1-micro to evaluate the feature learning performance of the model, use the trained model to predict the social network data to be tested, and obtain the social network distribution map.
[0032] The specific steps include: For the input heterogeneous graph data G = { V , E}, where V is a set of nodes, E is a set of edges, after S1, the node feature matrix X and the adjacency matrix set of different types of edges of the heterogeneous graph network are extracted are input into the model constructed by S2, and finally the predicted results are output by the linear layer of the model and softmax the loss function . During the entire model training process, by adjusting the hyperparameters such as dropout, weight-decay, learning rate, and mixed residual ratio, the optimal combination of hyperparameters that make the model perform best on the test data set is found, and the model is optimized and constructed.
[0033] During the entire model training process, by adjusting the hyperparameters such as dropout, weight-decay, learning rate, and mixed residual ratio, the optimal combination of hyperparameters that make the model perform best on the test data set is found, and the model is optimized and constructed. The total number of iterations is 50, the optimizer uses Adam, the learning rate is set to 0.001, and the weight-decay is set to 0.001.
[0034] The established model is evaluated on the social network dataset IMDB using the graph node classification indicators F1-macro and F1-micro, and the feature learning performance of the model is evaluated, and the results show that the established model can learn effective heterogeneous graph node representation, and the experimental results exceed the mainstream heterogeneous graph node classification model.
[0035] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.
[0036] The principles and implementation manners of the present application are described by applying specific examples herein, and the above description of the examples is only for helping to understand the core idea of the present application; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the specification should not be understood as the limitation of the present application.
Claims
1. A social network analysis method combining relation-awareness and high-order heterogeneous graph convolution, characterized in that, include: Historical social network data is acquired and heterogeneous graph data is constructed. The heterogeneous graph data consists of adjacency matrices of nodes and different types of edges. Nodes represent users in the social network, and the adjacency matrix represents the relationships between different users, including friend relationships, kinship relationships, and interaction relationships. A social network analysis model combining relation-aware and high-order heterogeneous graph convolution is constructed. The social network analysis model includes an input layer, a relation-aware nonlinear modeling module, a semantic-aware fusion module, a structure enhancement module, a high-order heterogeneous graph convolution module, and an output module. The input layer is connected to both the relation-aware nonlinear modeling module and the structure enhancement module. The relation-aware nonlinear modeling module is connected to the semantic-aware fusion module. Both the semantic-aware fusion module and the structure enhancement module are also connected to the high-order heterogeneous graph convolution module, which is also connected to the output module. The heterogeneous graph data is input into the social network analysis model for training, and the feature learning performance of the model is evaluated using graph node classification metrics F1-macro and F1-micro. The trained model is then used to predict the social network data to be tested, resulting in a social network distribution map.
2. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The acquisition of historical social network data and the construction of heterogeneous graph data specifically include: Historical social network data is acquired and standardized. Node feature matrices of heterogeneous graphs are then extracted from the standardized data. and the adjacency matrix set of different types of edges. ;in, t For relation type sequence number, The number of edge types in the heterogeneous graph. N For the number of nodes, F For node feature dimensions, It is a real number.
3. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, In the social network analysis model: the input layer is used to receive heterogeneous graph data; the relation-aware nonlinear modeling module is used to construct a nonlinear graph convolutional layer based on the adjacency matrix of different types of edges in the heterogeneous graph and to learn node features using the KAN network; The semantic-aware fusion module is used to fuse the learned node representations based on different edge types to obtain fused node features; The structure enhancement module first transforms the adjacency matrix set, aggregating adjacency matrices of various edge types. Then, it uses a Markov process to encode the overall structure of the heterogeneous graph, generating an adjacency matrix that integrates the edge information of various heterogeneous graph types. The high-order heterogeneous graph convolution module first uses multiple propagation layers to perform feature aggregation on the fused node features and the adjacency matrix that integrates the edge information of various heterogeneous graph types to extract information of neighboring nodes of different orders. Then, it performs weighted aggregation on the representations output by different propagation layers. Finally, it generates prediction results through linear transformation and the softmax function. The output module outputs the prediction results.
4. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The processing steps of the relation-aware nonlinear modeling module include: In heterogeneous graphs, due to the heterogeneity of edges, each type of edge or relation has a corresponding adjacency matrix, forming a set of adjacency matrices based on different types of edges. , Equal to the number of edge types in the heterogeneous graph, the node feature matrix is represented as , Represent real numbers, N Indicates the number of nodes. F The dimension of the node feature, index t Indicates the relation type number; First, transform the node type feature matrix. The features of different types of nodes are transformed to generate transformed node features. d is the embedding dimension, and the transformation uses the formula: ; Then, based on the adjacency matrix for each relation type... A t The KAN neural network is used to extract local structure and feature information. Specifically, the KAN neural network constructs a relation-aware mapping function through function decomposition to aggregate node features, generating node embeddings rich in relational semantics. The calculation formula is as follows: in, It is a cubic spline function. It is a linear function used to process the mapped features. It is the activation function ReLU, subscript q and p The identifier representing the function. n Represents the dimension of the hidden layer in the network. For relation type t The adjacency matrix, This is the transformed node feature matrix. It is the output of the relation-aware nonlinear modeling module.
5. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The processing procedure of the semantic awareness fusion module includes: Representation for each relation type { Learn its corresponding weights { }, formally represented as: in, To implement a neural network that learns weights, the steps include: To learn the weight information for each relation type representation, each input representation { First, use three learnable parameter matrices. , , Perform linear transformations separately, and the calculation formulas are as follows: , in, , and The input represents the output obtained through a linear transformation of three parameter matrices, which is then computed using a dot product attention mechanism. and The attention coefficient, and then use the attention coefficient. softmax The function is normalized to obtain the attention score for each relation type representation. : here, T This is a matrix transpose operation. d For the embedding dimension, attention scores are then calculated using different types of representations. The representations of different relation types are fused to obtain the fused node feature representation. : 。 6. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The processing steps of the structural enhancement module include: First, set up adjacency matrices of different relation types. The transformation yields a global adjacency matrix. : Then use H degree matrix D right H Normalization is performed to obtain the normalized matrix. The structure enhancement module iteratively optimizes the topology of the graph through a Markov process to enhance the connectivity between nodes with similar structures. The calculation process is as follows: in, D yes H The degree matrix, It is after the first iteration k The adjacency matrix obtained is then used to perform a nonlinear expansion operation on the adjacency matrix of the graph by introducing an expansion factor, as shown in the following formula: in, a As an extension factor, [ i , j ] represents the matrix index; then set the threshold. To reduce redundant information in the graph and retain the more meaningful connections: Finally, an enhanced adjacency matrix is generated. .
7. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The processing steps of the higher-order heterogeneous graph convolution module include: First, use the enhanced adjacency matrix. Propagating node features yields node representations at different layers, let , No. l The nodes of a layer are represented as follows: in, It is the first l Layer node representation, concatenating layer outputs to facilitate the aggregation of information from different layers: Here, Concat() represents the concatenation operation, and then a learnable parameter matrix is used. right Z Perform transformations to extract more meaningful representations: Then use the learnable parameter matrix and Aggregating information from different layers: in, tanh and ReLU It is a non-linear activation function, and is used last. softmax Normalizing the function yields the node embedding. : Embed the final node The input is fed into a feedforward network consisting of two linear layers to obtain the distribution of the output class. P : in, For a linear layer with learnable weights, This indicates the result of the prediction.
8. The social network analysis method combining relation awareness and high-order heterogeneous graph convolution as described in claim 1, characterized in that, The social network analysis model uses the cross-entropy loss function to measure the error between the predicted results and the true labels during the training process and updates the model parameters accordingly.