A social network relationship prediction method based on a multi-curvature hyperbolic graph neural network

Abstract

The application discloses a social network relationship prediction method based on a multi-curvature hyperbolic graph neural network, and belongs to the technical field of social networks, and comprises the following steps: constructing a social network graph; constructing a category prediction network comprising a random deletion layer, a multi-curvature hyperbolic space mapping layer, an improved hyperbolic graph convolutional neural network, a dynamic curvature fusion layer and a classifier; training the category prediction network to obtain a category prediction model; obtaining a social network graph to be identified, and obtaining a corresponding fusion feature matrix Zr through the category prediction model; processing Zr into a binary matrix, and determining an edge relationship according to element values in the binary matrix. The application introduces a multi-curvature learning mechanism, so that the model is more flexible and more accurate in matching the real geometric characteristics of substructures in the social network graph, a scaling aggregation factor is introduced during hyperbolic graph neural network feature extraction to adjust the contribution of neighbors, so as to adjust node features, maintain the hierarchical relationship during multi-layer feature extraction, and finally improve the accuracy of relationship prediction.

Description

Technical Field

[0001] This invention relates to the field of social network technology, and in particular to a method for predicting social network relationships based on a multi-curvature hyperbolic graph neural network. Background Technology

[0002] A social network is a professional graph structure based on graph theory. It uses people (or organizations) as nodes, draws "who is related to whom" as edges, and the whole consisting of nodes and edges is the social network.

[0003] Scientific research collaboration is a key factor driving technological progress. The relationships within such collaborations are highly diverse. Academic social networks are specific network structures of social networks within a research context. Their nodes can be researchers, papers, journals, conferences, keywords, institutions, etc., while edges represent various relationships between these nodes arising from research activities. For example, when a node is a researcher, the edge relationship can be defined as a "co-authorship relationship." If two nodes co-author a paper, an edge is established between them, and the edge weight can be defined as the "number of co-authorships" between the two nodes. Alternatively, the edge relationship can be defined as "same institution." If nodes belong to the same predefined institution (team, geographical region), an edge is established between them. Similarly, when a node is a paper, the edge relationship can be defined as a "citation relationship." For instance, if paper A cites paper B, a directed edge is established between them. Analyzing academic social networks and predicting potential collaborative relationships helps optimize the allocation of research resources and promote interdisciplinary innovation. Existing methods for predicting scientific research collaborations mainly revolve around network structures and can be divided into two categories:

[0004] (1) Similarity index-based and classical machine learning methods: These methods transform the prediction problem into a similarity calculation or binary classification task between nodes. Among them, structural similarity methods (such as common neighbors, Jaccard, Katz, etc.) use the topological information of the network to evaluate the potential connection probability between authors and have a certain degree of universality; while machine learning methods (such as logistic regression, support vector machine, etc.) use these similarity indices as features to train models to predict the existence of cooperative relationships. Since similarity indices such as common neighbors and Jaccard can only capture shallow information of local adjacency structure, they cannot model the hierarchical nature (such as community structure, power hierarchy) that is common in social networks.

[0005] (2) Graph Neural Network-Based Methods: These methods embed research collaboration networks into vector spaces for learning, especially graph neural networks operating in Euclidean space. They can automatically learn the high-order structural representations of nodes and are better able to capture nonlinear relationships compared to traditional methods. For example, patent application number 202411854409.X discloses a method for constructing a multi-layer research collaboration relationship network based on professional papers and patent documents. Through key data confirmation and multiple matching strategies, it improves the accuracy of identifying collaboration relationships. Although the above patent solves the problem of research prediction, the following problems still exist in actual operation:

[0006] (1) The failure to delve into the fine-grained semantic relationships between papers makes it difficult to effectively identify and confirm the collaborative relationship.

[0007] (2) It failed to explore the multi-level geometric characteristics between sub-communities with different disciplines and different degrees of cooperation. It used a single curvature space to represent the network structure, which made it difficult to globally and optimally characterize the complex structure of the entire network, thus limiting the model's expressive power and prediction accuracy.

[0008] (3) In hyperbolic space, distance increases exponentially with radius. Small changes in the representation vector of nodes far from the origin (which may correspond to network edges or high-level nodes) can lead to huge distance differences. Without a scaling factor, aggregating the features of neighbors with huge differences in distance will distort the hyperbolic metric.

[0009] (4) Using only a single curvature. Complex real-world social networks cannot conform to only one uniform geometric constraint. For example, social networks related to scientific research collaborations have both tight lab circles (highly hierarchical) and loose interdisciplinary connections (flatter). Using a single curvature would force all heterogeneous regions in the network (such as tight collaboration circles and loose interdisciplinary connections) to squeeze into the same geometric framework, resulting in a serious decrease in the embedding quality of some nodes, which cannot accurately reflect their true topological roles. Summary of the Invention

[0010] The purpose of this invention is to provide a social network relationship prediction method based on a multi-curvature hyperbolic graph neural network that solves the above-mentioned problems by introducing a multi-curvature learning mechanism, enabling the model to adaptively assign appropriate hyperbolic curvature to different nodes or substructures in the network, thereby more accurately modeling the heterogeneous hierarchical structure of scientific research networks and improving the accuracy of cooperative relationship prediction.

[0011] To achieve the above objectives, the technical solution adopted by this invention is as follows: a social network relationship prediction method based on a multi-curvature hyperbolic graph neural network, comprising the following steps:

[0012] S1, construct the social network graph G=(V,E);

[0013] The system predefines nodes, node categories, edge relationships, and node feature vectors. The nodes can be publications, authors, conferences, journals, or research fields. If there is an edge relationship between two nodes, an edge is generated between the two nodes.

[0014] Construct a set of nodes V and a set of edges E, respectively, and construct G=(V,E). Construct the node feature matrix X, adjacency matrix A, and category label set Y of G. The i-th row vector in X is the node feature vector of node i, 1≤i≤|V|, and |V| is the total number of nodes.

[0015] S2, construct a class prediction network, including a random deletion layer, a multi-curvature hyperbolic space mapping layer, an improved hyperbolic graph convolutional neural network, a dynamic curvature fusion layer, and a classifier;

[0016] The random deletion layer is used to randomly delete edges of some nodes within G, generate a preprocessed network graph G', and regenerate the preprocessed adjacency matrix A' according to G';

[0017] The multicurvature hyperbolic space mapping layer is used to learn and generate... Let X be a hyperbolic space, and let X be mapped to each hyperbolic space to obtain a mapping matrix, where the curvature of the θ-th hyperbolic space is c. θ The mapping matrix is ; 1≤θ≤ ;

[0018] The improved hyperbolic graph convolutional neural network includes K layers of improved hyperbolic graph convolutional layers, where the k-th layer is used to generate the hyperbolic graph convolutional features of the k-th layer according to the following formula. , 1≤k≤K;

[0019] ,

[0020] In the formula, It is a mapping from tangent space to hyperbolic space. It is a mapping from hyperbolic space to tangent space, where N(i) is the set of neighboring nodes of node i, and j is the j-th neighboring node in N(i). It is the scaling aggregation factor of the k-th layer. , These are the weight matrix and bias term of the k-th layer, respectively. The output features of the (k-1)th hyperbolic graph convolutional layer are... and These are addition and multiplication in hyperbolic space, respectively, and × represents the Hadamard product;

[0021] The dynamic curvature fusion layer is used to map the Kth layer hyperbolic graph convolutional features of each hyperbolic space to the tangent space, and then adaptively weighted and fused them to obtain the fused feature matrix Z.

[0022] The classifier takes Z as input and outputs a probability vector of each node in Z belonging to each category, and constructs a predicted category set from the probability vectors of all nodes. ;

[0023] S3, train the category prediction network with G=(V,E) and minimize The parameters of the category prediction network are adjusted to convergence using the Bayesian personalized ranking loss function Loss between Y, resulting in the category prediction model.

[0024] S4, obtain the social network graph Gr to be identified, where the relationships between edges are incomplete, input the category prediction model, and obtain the corresponding fusion feature matrix Zr;

[0025] S5. For each row vector of Zr, calculate the inner product of each pair to generate an inner product matrix, and then convert it to a binary matrix using 0-1. For each element in the binary matrix, if the value is 1, then there is an edge relationship between the corresponding two nodes; otherwise, there is no edge relationship.

[0026] As a preferred option Calculate according to the following formula:

[0027] ,

[0028] ,

[0029] In the formula, Let be the scaling aggregation factor for node i at layer k, e be the natural constant, and d be the scaling aggregation factor. j It is the number of neighboring nodes of node j. This is for splicing operations.

[0030] Preferably, in S2, the dynamic curvature fusion layer generates the fusion feature matrix Z, specifically including the following steps:

[0031] S21, map the Kth layer hyperbolic graph convolutional features of each hyperbolic space to the tangent space to obtain the tangent space features, where the Kth layer hyperbolic graph convolutional features of the θth hyperbolic space are... The tangent space features are ;

[0032] S22, generate the fusion feature vector for each node, where the fusion feature vector Z of node i is... i We obtain the following formula:

[0033] ,

[0034] ,

[0035] In the formula, Z i Let be the fused feature vector of node i. for The vector in the i-th row, for The weights, where exp(∙) is the exp function. For temperature parameters, Let λ be the i-th row vector of the tangent space feature;

[0036] S23, Z i As the i-th row vector of Z, the fused feature matrix Z is generated.

[0037] Preferably, in S3, one training session includes steps S31 to S33:

[0038] S31. Input G=(V,E) into the category prediction network, and randomly delete some edges through the random deletion layer to obtain the preprocessed network graph G' and the preprocessed adjacency matrix A'.

[0039] S32, mapping the node feature matrix X to A hyperbolic space is obtained. Each mapping matrix is ​​processed by an improved hyperbolic graph convolutional neural network to obtain... Each output is then passed through a dynamic curvature fusion layer to obtain a fused feature matrix Z. A classifier then outputs the predicted category for each node in Z, forming a set of predicted categories. ;

[0040] S33, Calculation The Bayesian personalized ranking loss function Loss for Y is used, and the parameters of the category prediction network are adjusted by minimizing the Loss.

[0041] Regarding the random deletion layer: It is used to randomly delete some edges from a known complete social network graph G and train a model to predict these deleted edges. After deleting the edges, the node feature matrix X in G=(V,E) remains unchanged, but the adjacency matrix A will change. The adjacency matrix obtained after each random deletion of edges is different.

[0042] Regarding multi-curvature hyperbolic space mapping layers: By setting... Hyperbolic spaces with different curvatures ~ The origin is o in sequence Γ =(0, 0, ...,0), This transforms the eigenvectors of each node in the node feature matrix X through an initialization transformation, mapping them to the vicinity of the corresponding hyperbolic space origin. This generates a mapping matrix for the node feature matrix X in each hyperbolic space, where the curvature of the θ-th hyperbolic space is c. θ The mapping matrix is The calculation formula is: In the formula, , The curvature c θ Learnable transformation matrix and bias vector, The curvature is c θ The hyperbolic space is an exponential mapping at the origin. Thus, a set of curvature values ​​c can be learned. θ ={c θ |θ= 1,2,…, }, mapping social graph networks to In hyperbolic spaces with different curvatures, generate It can represent the features of hyperbolic spaces and automatically learn which curvature is suitable for which structure, thereby expressing complex social networks more accurately and flexibly.

[0043] Regarding improvements to hyperbolic graph convolutional neural networks: Existing hyperbolic graph convolutional neural networks consist of multiple stacked hyperbolic graph convolutional layers. Each layer processes the input features through operations such as hyperbolic transformation, hyperbolic neighborhood aggregation, and hyperbolic nonlinear activation. While hyperbolic space excels at representing hierarchical structures, aggregation operations, which aggregate features from neighbors with significant differences in distance, distort the hyperbolic metric, affecting subsequent prediction accuracy. To overcome this deficiency, this invention:

[0044] (1) Construct a scaling aggregation factor for each hyperbolic graph convolutional layer, where the scaling aggregation factor for the k-th layer is: , , Let i be the scaling aggregation factor for node i in the k-th layer, according to the formula It can be calculated that, therefore, it is known that It is related to the number of neighbors of node i and can be regarded as the contribution of the neighbors of node i in the k-th hyperbolic graph convolutional layer.

[0045] (2) Introduce the scaling aggregation factor into the hyperbolic graph convolutional layer, see the following equation:

[0046] ,

[0047] in, It contains two layers of parentheses, with the innermost layer representing a hyperbolic transformation. For nonlinear activation, To map the result of the nonlinear transformation back to the tangent space to obtain all spatial features ,and The size is the same as X, then use for To perform aggregation scaling, specifically, use... The i-th element and Each element in the i-th row vector is multiplied sequentially. By adjusting the contributions of its neighbors, node i's position is adjusted. The internal characteristics are used to maintain the hierarchical relationship.

[0048] Regarding the dynamic curvature fusion layer: The node feature matrix X is obtained through a multi-curvature hyperbolic space mapping layer. mapping matrices ~ The improved hyperbolic graph convolutional neural network outputs hyperbolic graph convolutional features. ~ The scales are all the same as X, for ~ Each node is weighted and fused to obtain the fused feature vector of each node, which is then combined into a complete fused feature matrix Z.

[0049] Compared with the prior art, the advantages of the present invention are as follows:

[0050] (1) Stronger structural representation ability: By introducing a multi-curvature learning mechanism, the model can adaptively assign appropriate hyperbolic curvature to different nodes or substructures in the network, thereby more flexibly and accurately matching the real geometric characteristics of substructures in the social network graph, modeling the heterogeneous hierarchical structure of the social network graph, and overcoming the expression limitations of the single curvature model.

[0051] (2) Higher prediction accuracy: In the hyperbolic space with multiple curvature, the hyperbolic graph convolutional neural network is improved. In each layer of the existing hyperbolic graph convolutional neural network, a scaling aggregation factor is introduced to perform adaptive aggregation on each node. This ensures that in the hyperbolic space with multiple curvature, the node embedding can maintain the fine structure of the local close cooperative community and the large-scale hierarchical structure of the global network at the same time, so as to more accurately mine potential features and improve the accuracy of subsequent predictions.

[0052] (3) Better interpretability: The learned curvature parameters can be used as a measure of the heterogeneity of network structure. High curvature (small absolute value) regions may correspond to closely connected and domain-specific communities, while low curvature (large absolute value) regions may correspond to more abstract and interdisciplinary communication levels, providing a new perspective for relationship analysis in social networks. Attached Figure Description

[0053] Figure 1 This is a flowchart of the present invention;

[0054] Figure 2 A diagram of the network structure for category prediction. Detailed Implementation

[0055] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0056] Example 1: See Figure 1 and Figure 2 A social network relationship prediction method based on a multi-curvature hyperbolic graph neural network includes the following steps:

[0057] S1, construct the social network graph G=(V,E);

[0058] The system predefines nodes, node categories, edge relationships, and node feature vectors. The nodes can be publications, authors, conferences, journals, or research fields. If there is an edge relationship between two nodes, an edge is generated between the two nodes.

[0059] Construct a set of nodes V and a set of edges E, respectively, and construct G=(V,E). Construct the node feature matrix X, adjacency matrix A, and category label set Y of G. The i-th row vector in X is the node feature vector of node i, 1≤i≤|V|, and |V| is the total number of nodes.

[0060] S2, construct a class prediction network, including a random deletion layer, a multi-curvature hyperbolic space mapping layer, an improved hyperbolic graph convolutional neural network, a dynamic curvature fusion layer, and a classifier;

[0061] The random deletion layer is used to randomly delete edges of some nodes within G, generate a preprocessed network graph G', and regenerate the preprocessed adjacency matrix A' according to G';

[0062] The multicurvature hyperbolic space mapping layer is used to learn and generate... Let X be a hyperbolic space, and let X be mapped to each hyperbolic space to obtain a mapping matrix, where the curvature of the θ-th hyperbolic space is c. θ The mapping matrix is ; 1≤θ≤ ;

[0063] The improved hyperbolic graph convolutional neural network includes K layers of improved hyperbolic graph convolutional layers, where the k-th layer is used to generate the hyperbolic graph convolutional features of the k-th layer according to the following formula. , 1≤k≤K;

[0064] ,

[0065] In the formula, It is a mapping from tangent space to hyperbolic space. It is a mapping from hyperbolic space to tangent space, where N(i) is the set of neighboring nodes of node i, and j is the j-th neighboring node in N(i). It is the scaling aggregation factor of the k-th layer. , These are the weight matrix and bias term of the k-th layer, respectively. The output features of the (k-1)th hyperbolic graph convolutional layer are... and These are addition and multiplication in hyperbolic space, respectively, and × represents the Hadamard product;

[0066] The dynamic curvature fusion layer is used to map the Kth layer hyperbolic graph convolutional features of each hyperbolic space to the tangent space, and then adaptively weighted and fused them to obtain the fused feature matrix Z.

[0067] The classifier takes Z as input and outputs a probability vector of each node in Z belonging to each category, and constructs a predicted category set from the probability vectors of all nodes. ;

[0068] S3, train the category prediction network with G=(V,E) and minimize The parameters of the category prediction network are adjusted to convergence using the Bayesian personalized ranking loss function Loss between Y, resulting in the category prediction model.

[0069] S4, obtain the social network graph Gr to be identified, where the relationships between edges are incomplete, input the category prediction model, and obtain the corresponding fusion feature matrix Zr;

[0070] S5. For each row vector of Zr, calculate the inner product of each pair to generate an inner product matrix, and then convert it to a binary matrix using 0-1. For each element in the binary matrix, if the value is 1, then there is an edge relationship between the corresponding two nodes; otherwise, there is no edge relationship.

[0071] In this embodiment, Calculate according to the following formula:

[0072] ,

[0073] ,

[0074] In the formula, Let be the scaling aggregation factor for node i at layer k, e be the natural constant, and d be the scaling aggregation factor. j It is the number of neighboring nodes of node j. This is for splicing operations.

[0075] In S2, the dynamic curvature fusion layer generates the fusion feature matrix Z specifically through the following steps:

[0076] S21, map the Kth layer hyperbolic graph convolutional features of each hyperbolic space to the tangent space to obtain the tangent space features, where the Kth layer hyperbolic graph convolutional features of the θth hyperbolic space are... The tangent space features are ;

[0077] S22, generate the fusion feature vector for each node, where the fusion feature vector Z of node i is... i We obtain the following formula:

[0078] ,

[0079] ,

[0080] In the formula, Z i Let be the fused feature vector of node i. for The vector in the i-th row, for The weights, where exp(∙) is the exp function. For temperature parameters, Let λ be the i-th row vector of the tangent space feature;

[0081] S23, Z i As the i-th row vector of Z, the fused feature matrix Z is generated.

[0082] In S3, a training session includes steps S31 to S33:

[0083] S31. Input G=(V,E) into the category prediction network, and randomly delete some edges through the random deletion layer to obtain the preprocessed network graph G' and the preprocessed adjacency matrix A'.

[0084] S32, mapping the node feature matrix X to A hyperbolic space is obtained. Each mapping matrix is ​​processed by an improved hyperbolic graph convolutional neural network to obtain... Each output is then passed through a dynamic curvature fusion layer to obtain a fused feature matrix Z. A classifier then outputs the predicted category for each node in Z, forming a set of predicted categories. ;

[0085] S33, Calculation The Bayesian personalized ranking loss function Loss for Y is used, and the parameters of the category prediction network are adjusted by minimizing the Loss.

[0086] Example 2: See Figure 1 and Figure 2 Based on Example 1, in order to more intuitively express the predictive effect of the present invention, this example uses the method of the present invention and the GCN method for comparative experiments.

[0087] A research social network graph G=(V,E) is constructed based on the ACM (Association for Computing Machinery) dataset. Nodes in the ACM dataset include publications, authors, conferences, journals, and research fields. In this embodiment, nodes are predefined as "publications," and edges are "citation links." The dimension of the node feature vector is 1870, obtained based on node feature encoding. The node feature is the textual representation of "publication." Node types are categorized into three types according to research fields: databases, wireless communication, and data mining. The resulting social network graph contains 3025 nodes and 13128 edges, with node types categorized into three classes.

[0088] Experimental grouping: The experimental group was the present invention, and the control group was the GCN method.

[0089] Evaluation metrics: AUC, AP. AUC stands for Area Under Curve, referring to the area under the ROC curve. AP stands for Average Precision, referring to the average precision.

[0090] Hyperparameter settings: =3, the curvatures of the three hyperbolic spaces are c1=-5.0, c2=-1.0, c3=-0.05, the dimension of the fused feature matrix Z is 16, and the learning rate is 0.001.

[0091] Experimental environment: The hardware environment is a 2.39 GHz E5-2680 CPU with 8 cores, 64 GB RAM, and an 11G GTX1080ti GPU. The software environment is PyTorch Geometric.

[0092] The proposed method and the GCN method were trained using a scientific research social network graph. Edge prediction was performed, and AUC and AP were used for evaluation, resulting in Table 1 below:

[0093] Table 1. Comparison of Performance Indicators of Different Methods

[0094] method AUC(%) AP(%) GCN method 84.27 83.67 This invention 87.06 86.91

[0095] As shown in Table 1, the AUC and AP scores of the present invention are significantly higher than those of the GCN method, verifying the effectiveness of the model and method of the present invention.

[0096] Furthermore, the social network is not limited to one constructed based on the ACM dataset; it can be constructed as long as the nodes, the edge relationships between nodes, and the node features are determined.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A social network relationship prediction method based on a multi-curvature hyperbolic graph neural network, characterized in that, Comprising the following steps: S1, constructing a social network graph G=(V, E); Predefining nodes, node categories, edge relationships, node feature vectors, if there is an edge relationship between two nodes, an edge is generated between the two nodes; All nodes and edges form a node set V and an edge set E, respectively, to construct G=(V, E), and a node feature matrix X, an adjacency matrix A, and a category label set Y of G are constructed, wherein the i-th row vector of X is the node feature vector of node i, 1≤i≤|V|, and |V| is the total number of nodes; S2, constructing a category prediction network, comprising a random deletion layer, a multi-curvature hyperbolic space mapping layer, an improved hyperbolic graph convolutional neural network, a dynamic curvature fusion layer, and a classifier; The random deletion layer is used to randomly delete edges of part of the nodes in G to generate a preprocessed network graph G', and a preprocessed adjacency matrix A' is regenerated according to G'; The multi-curvature hyperbolic space mapping layer is used for learning generation The X is mapped to each hyperbolic space to obtain a mapping matrix, wherein the curvature of the theta hyperbolic space is c θ , and the mapping matrix is ; 1≤θ≤ ; The improved hyperbolic graph convolutional neural network comprises K layers of improved hyperbolic graph convolutional layers, wherein the kth layer is used to generate the hyperbolic graph convolutional features of the kth layer according to the following formula , 1≤k≤K; ， wherein, is a mapping from the tangent space to the hyperbolic space, is a mapping from the hyperbolic space to the tangent space, N(i) is a set of neighbor nodes of node i, j is the jth neighbor node in N(i), is a scaling aggregation factor of the kth layer, , are a weight matrix and a bias term of the kth layer, respectively, is an output feature of the k-1th hyperbolic graph convolution layer, and are addition and multiplication in the hyperbolic space, respectively, and x is the Hadamard product. The dynamic curvature fusion layer is used to map the K-th layer hyperbolic graph convolutional features of each hyperbolic space to the tangent space respectively, and then adaptively weighted fusion is performed to obtain a fusion feature matrix Z; The classifier is used to input Z, output the probability vector of each node in Z belonging to each category, and form a set of predicted categories by combining the probability vectors of all nodes ; S3, training the category prediction network with G=(V, E) and minimizing and Y Bayesian personalized ranking loss function Loss adjusts the parameters of the category prediction network to convergence, and obtains a category prediction model; S4, obtaining a to-be-identified social network graph Gr, wherein the edge relationship is incomplete, inputting the category prediction model, and obtaining a corresponding fusion feature matrix Zr; S5, for each row vector of Zr, the inner product of each pair is calculated to generate an inner product matrix, which is then 0-1 converted into a binary matrix, and for the elements in the binary matrix, if the value is 1, there is an edge relationship between the corresponding two nodes, otherwise there is no edge relationship. 2.The social network relationship prediction method based on the multi-curvature hyperbolic graph neural network according to claim 1, characterized in that, The calculation is made according to the following formula: ， ， wherein, is the scaling aggregation factor for the kth layer node i, e is the natural constant, d j is the number of neighbor nodes of node j, is the concatenation operation. 3.The social network relationship prediction method based on the multi-curvature hyperbolic graph neural network according to claim 1, characterized in that, In S2, the dynamic curvature fusion layer generates the fusion feature matrix Z, which comprises the following steps: S21, respectively map the Kth layer hyperbolic graph convolution feature of each hyperbolic space to the tangent space to obtain a tangent space feature, wherein the Kth layer hyperbolic graph convolution feature of the θth hyperbolic space is , and the tangent space feature is ; S22, generating a fusion feature vector of each node, wherein the fusion feature vector Z i is obtained according to the following formula: ， ， where Z i is the fused feature vector for node i, is is the i-th row vector of the matrix is is the weight of the i-th row vector of the matrix is the temperature parameter, is the i-th row vector of the λ-th tangent space feature S23, generating a fusion feature matrix Z from the ith row vector of Z i The fusion feature matrix Z is generated as the ith row vector of Z. 4.The social network relationship prediction method based on the multi-curvature hyperbolic graph neural network according to claim 1, characterized in that, In S3, one training comprises steps S31-S33: S31, inputting G=(V, E) into the category prediction network, randomly deleting part of the edges through the random deletion layer to obtain a preprocessed network graph G' and a preprocessed adjacency matrix A'; S32, mapping the node feature matrix X to hyperbolic space, obtaining a mapping matrix, obtaining an output from each mapping matrix through an improved hyperbolic graph convolutional neural network, and obtaining a fusion feature matrix Z through a dynamic curvature fusion layer, outputting the predicted category of each node in Z through a classifier to form a predicted category set ; S33, calculating and Y's Bayesian personalized ranking loss function Loss, and minimize Loss to adjust the parameters of the category prediction network.

Images