Hypergraph link prediction method based on structure perception and edge feature fusion

Abstract

The invention discloses a hypergraph link prediction method based on structure perception and edge feature fusion, and the method comprises the steps: introducing edge information explicit modeling under a hypergraph frame, combining a self-adaptive sub-graph sampling mechanism, taking the difference of a local structure into consideration while modeling a high-order relation, and introducing a hypergraph modeling method with edge information as the center. According to the method, nodes in an original graph are converted into edges in a hypergraph, the edges of the original graph are converted into nodes in the hypergraph, by means of the natural high-order relation modeling capacity of a hypergraph convolutional network, the over-smoothing problem is relieved while the expression capacity of a model for a complex structure is effectively improved, and by dynamically adjusting the hop count, the over-smoothing performance of the model is improved. And the most relevant neighbor node is selected to construct a local context so as to improve the structural information capture capability of link prediction.

Description

Technical Field

[0001] This invention relates to the technical field of link prediction, and in particular to a hypergraph link prediction method based on structure awareness and edge feature fusion. Background Technology

[0002] Link prediction, a core problem in complex network analysis, aims to predict potential edges in a graph that are not yet explicitly observed. This task has wide applications in various fields. For example, in social networks, it helps discover potential social relationships by recommending acquaintances to users. In recommender systems, link prediction uses users' historical search and purchasing behavior preferences to predict products or content that users might be interested in, thereby improving user satisfaction and boosting sales. In bioinformatics, it can be used to discover unproven protein-protein interactions, contributing to medical research and pharmaceutical applications. To support the development of the graph, early researchers mostly adopted intuitive heuristic methods for specific application scenarios of the aforementioned link prediction task. These methods typically extract features from the local or global structure of the graph and assign a similarity score to each pair of candidate nodes. The higher the score, the greater the likelihood of establishing an edge between the nodes. Common heuristic methods include common neighbors, Adamic-Adar index, resource allocation, shortest path, and Katz index. However, while these heuristic methods perform well in certain specific scenarios and have high computational efficiency, they often rely on manually designed rules and have certain limitations in characterizing complex structural features and improving generalization ability.

[0003] Chinese patent application number 201810999492.8 discloses a social network link prediction method based on multi-source heterogeneous data fusion. The method includes the following steps: Step 1, extracting training and test sets from the user relationship topology graph; Step 2, using network representation learning to learn and obtain the node set social network user vector from the positive samples of the training set; Step 3, constructing a user-location check-in frequency matrix based on user check-in records, and then using Poisson matrix decomposition to obtain the user access preference vector in the low-dimensional vector space; Step 4, using a deep learning model to capture the association between the two data sources, the user relationship topology graph and the user check-in records, to obtain the user access preference vector; Step 5, fusing the user access preference vector and the social user vector, and inputting them into a convolutional neural network for training to obtain link prediction. The shortcomings of this method are that if the input social network is not sparse, the use of Poisson matrix decomposition will lose useful information, and the use of multi-source data increases the amount of data processing and computational complexity.

[0004] Chinese Patent Application No. 2020107524253.3 discloses a method for predicting social network links using knowledge graph embedding and temporal convolutional networks. The method includes the following steps: Step 1, processing raw social sample data and extracting tuples of phrases and related independent variables, constructing structured temporal tuples and linking them to a knowledge graph, constructing a subgraph from the knowledge graph and extracting event embedding vectors; Step 2, constructing a social network, transforming real-world non-Euclidean data (i.e., first-order network topology) into an adjacency matrix, and converting the first-order adjacency matrix into a Laplace matrix in vector form. Step 3: Integrate event embedding vectors and network adjacency matrices; Establish a link prediction model based on an improved temporal convolutional network, and predict network links by combining event embedding vectors with network links. When the phrase-related independent variable tuples extracted by this method are input into the network model for model training, unnecessary information is included in the training due to the lack of information filtering of the phrase-related independent variable tuples, resulting in noise in the network. This method can only extract the first-order network topology in social networks, and its feature vectors cannot express the high-order network topology. The features are not rich enough, resulting in poor performance in social network connection prediction.

[0005] Chinese patent application number 202110280461.9 discloses a link prediction method based on dynamic network representation learning. This method constructs a similarity matrix of a snapshot network by calculating the similarity values ​​between nodes in a dynamic network through a similarity-based aggregation strategy. It then applies a graph convolutional neural network to a single snapshot network for feature aggregation, and uses the adjacency matrix and similarity matrix to guide the feature aggregation process to determine the low-dimensional feature representation of the nodes. The low-dimensional feature representation of the nodes is then input into a logistic regression classifier to obtain the link prediction results of the dynamic network. This method focuses on extracting the structural and dynamic features of the dynamic graph on the whole graph, but ignores the fact that the graph contains different structures and change rules at different levels. The feature extraction is incomplete, resulting in insufficient accuracy in dynamic link prediction.

[0006] Most current mainstream link prediction methods are based on subgraph extraction and graph neural network modeling. However, subgraph extraction strategies with a fixed number of hops are difficult to adapt to different structural densities of graphs, resulting in information redundancy or omission of key structures. Moreover, multi-layer stacked graph neural networks are prone to oversmoothing, which limits their ability to express high-order structural information. Summary of the Invention

[0007] The purpose of this invention is to solve the problems in the prior art and propose a hypergraph link prediction method based on structure awareness and edge feature fusion. This method can dynamically select the neighborhood range according to the importance of the node, improve the model's ability to express complex structures, and effectively improve the prediction accuracy.

[0008] To achieve the above objectives, this invention proposes a hypergraph link prediction method based on structure awareness and edge feature fusion, comprising the following steps: S1. Hypergraph Convolutional Network Based on Structure-Aware Subgraph Sampling: For the neighborhood information of the two end nodes of the target link, the centrality index of the node is calculated to dynamically optimize the hop count of the subgraph extraction, and the complete neighborhood subgraph is extracted in this way. The most relevant neighbor nodes within the range are selected to construct local context information. Then, a dual-radius node labeling algorithm is used to label each node in each subgraph. After obtaining the end node labeling information of the edge, it is concatenated in the form of one-hot encoding to obtain the edge features. Finally, the subgraph is transformed into a dual hypergraph and input into the hypergraph convolutional network to obtain the hypergraph representation. S2. Attribute and Topological Feature Generation: Calculate the similarity matrix between any two nodes in the attribute space, select the most similar nodes as neighbors for each node to form a nearest neighbor graph, encode the nearest neighbor graph using GAT, obtain the node feature representation of each layer, and use an average aggregator to represent the node features. S3. Feature Fusion and Link Prediction: Feature representations generated from different perspectives are fused using an adaptive attention mechanism to enhance the discriminative ability in the link prediction task.

[0009] Preferably, the hypergraph convolutional network based on structure-aware subgraph sampling in step S1 is constructed through the following three-stage process: S1-1, Subgraph Extraction: Adaptive neighborhood-aware subgraph extraction through a fusion strategy of hop selection driven by structural centrality and differential sampling; S1-2, Feature Generation: After obtaining the set of subgraphs corresponding to all candidate links in the original graph, the rich contextual information in the subgraphs is further modeled. In the feature generation module, structural side information is introduced to assist training and enhance the model's ability to model the structural information of nodes and edges. S1-3, Hypergraph Convolution: The output of the node labeling function is used as the side information of the node structure in the graph, and the edge information is further expressed. Using the idea of ​​dual hypergraph transformation, the nodes in the original graph are regarded as hyperedges in the hypergraph, and the edges in the original graph are regarded as hypernodes in the hypergraph, thus modeling higher-order information.

[0010] Preferably, in step S2, the adjacency matrix of the original input graph is directly used as the structural representation of the topology graph to obtain link topology information, and the same graph neural network is used for encoding to obtain the node feature representation from the topology graph. The encoding process shares network parameters to ensure the consistency of the representation space under different views.

[0011] Preferably, in step S3, the hypergraph modeling representation, the nearest neighbor graph modeling representation, and the node feature representation under the topological graph modeling are jointly input into the attention fusion module to learn the importance weights of different perspectives to the final representation. The importance weights directly reflect the contribution of features from different perspectives to different node pairs. In the subsequent prediction stage, the cosine similarity method is used to evaluate the potential connection probability between node pairs.

[0012] Preferably, in step S1-1, in the link prediction task, subgraph extraction is performed on the predicted edges, the nodes and their neighborhood information are defined under the hop count, node centrality is calculated, and node centrality is composed of betweenness centrality and normalization centrality index. These two are weighted and fused to model the neighborhood information. Betweenness centrality is used to measure the importance of a specific node as an intermediary in the shortest path of other nodes in the graph. The neighborhood structure perception score under the current hop count is defined and used to calculate the selection factor for subgraph probability sampling. The neighborhood subgraph when the hop count is greater than 1 is reconstructed. First, the differential neighborhood is defined, which represents the node information added based on the previous hop count when the current hop count is 1. A mask matrix is ​​generated for the node set and used as the differential information. The node set after masking is probabilistically sampled according to the selection factor of its corresponding hop count, and a complete subgraph set is further constructed.

[0013] Preferably, in step S1-2, a node position label mapping function is used to label all nodes in each closed subgraph using the DRNL algorithm. By calculating the shortest path length between any node in the subgraph and the target node, the relative position of the node in the subgraph is marked, thus obtaining the structural side information of the node. Furthermore, the information of the two end nodes with edges in the subgraph is concatenated to obtain the edge label information.

[0014] Preferably, in S1-3, the set of nodes and edges of the original graph, the set of supernodes and superedges in the hypergraph, and the superedge features and supernode features in the hypergraph are first defined. The superedge features and supernode features are then subjected to independent convolution operations, and after passing through several hidden layers, they are concatenated to obtain the hypergraph feature representation.

[0015] The beneficial effects of this invention are as follows: This invention introduces explicit edge information modeling within a hypergraph framework and combines it with an adaptive subgraph sampling mechanism to model higher-order relationships while taking into account the differences in local structures. It introduces an edge-information-centric hypergraph modeling method, converting nodes in the original graph into edges in the hypergraph, and vice versa. Leveraging the inherent ability of hypergraph convolutional networks to model higher-order relationships, this effectively enhances the model's ability to express complex structures while mitigating oversmoothing. By dynamically adjusting the hop count, the most relevant neighbor nodes are selected to construct the local context, improving the structural information capture capability of link prediction. Combining a structure centrality-driven hop count selection mechanism with a differential sampling algorithm, it can focus on semantically more relevant neighborhoods while avoiding information redundancy. Based on this, node labeling and side information feature generation steps are performed, using the constructed edge features as input to the hypernode features in the hypergraph, providing necessary feature information for dual hypergraph transformation. Next, a hypergraph convolutional network is used to capture the potential high-order relationships between nodes and edges. Meanwhile, in order to fully integrate information from different levels in the graph, two embedding representations are constructed using a graph attention network, one from the perspective of node attributes and the other from the perspective of topological structure. Finally, the two embedding representations are fused with the hypergraph embedding representation across views to improve the expressive power of link modeling from multiple perspectives. The overall framework aims to balance the semantics of edge relationships with the expressive power of information from different perspectives, thereby achieving better performance in link prediction tasks.

[0016] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of a hypergraph link prediction method based on structure awareness and edge feature fusion according to the present invention. Figure 2 This invention relates to a hypergraph link prediction method based on structure awareness and edge feature fusion, which is a subgraph of different sizes extracted under different hop counts under existing methods. Figure 3 This is a comparison of the prediction performance of various variant models of the SEH-LP hypergraph link prediction method based on structure awareness and edge feature fusion in different datasets according to the present invention. Figure 4 This is a comparison diagram of different GNN methods for hypergraph link prediction based on structure awareness and edge feature fusion, as proposed in this invention. Figure 5 This is a comparison diagram of different prediction methods for a hypergraph link prediction method based on structure awareness and edge feature fusion, as proposed in this invention. Figure 6 This invention presents experimental results for a hypergraph link prediction method based on structure awareness and edge feature fusion under different embedding dimensions. Detailed Implementation

[0018] 1 Model Architecture The overall architecture of the proposed hypergraph link prediction method based on structure awareness and edge feature fusion is shown in Figure 1. As a link prediction model based on subgraph extraction, it first addresses the problem of local structure representation in link modeling by proposing an adaptive closed subgraph extraction strategy. Combining a hop selection mechanism driven by structure centrality and a differential sampling algorithm, it can focus on semantically more relevant neighborhoods while avoiding information redundancy. Based on this, it performs node annotation and side information feature generation steps, using the constructed edge features as input to the hypernode features in the hypergraph, providing necessary feature information for dual hypergraph transformation. Next, it captures the potential higher-order relationships between nodes and edges through a hypergraph convolutional network. Simultaneously, to fully integrate information from different levels in the graph, it constructs two embedding representations—one from the node attribute perspective and one from the topological structure perspective—through a graph attention network. Finally, it performs cross-view fusion of the two embedding representations with the hypergraph embedding representation, improving the expressive power of link modeling from multiple angles. The overall framework aims to balance the semantics of edge relationships with the expressive power of information from different perspectives, thereby achieving better performance in link prediction tasks. The specific design and implementation of each module will be introduced in turn: 1.1 Hypergraph Convolutional Network Based on Structure-Aware Subgraph Sampling This invention first calculates the centrality index of the nodes based on the neighborhood information of the two end nodes of the target link to dynamically optimize the hop count of the extracted subgraph, and then extracts the complete neighborhood subgraph. The most relevant neighboring nodes within the range are selected to construct local context information. Then, a dual-radius node labeling algorithm is used to label each node in each subgraph. After obtaining the end node labeling information of the edge, it is concatenated in the form of one-hot encoding to obtain the edge features. Finally, the subgraph is transformed into a dual hypergraph and input into a hypergraph convolutional network to obtain the hypergraph representation. 1.1.1 Subgraph Extraction In the link prediction task, the core of the model can be summarized into three parts: subgraph extraction strategy, feature extraction method based on subgraph structure, and target edge prediction mechanism based on binary classifier. Among them, the feature extraction module is usually implemented with the help of graph neural network, which is widely regarded as an effective method to extract discriminative representations from graph structure. However, how to efficiently construct the input graph and extract the subgraph structure that can enhance the model's discriminative ability is one of the key factors affecting the final performance of the model. Existing methods typically employ neighborhood sampling strategies to extract subgraphs from the original graph using a fixed number of hops. Figure 2 The text simply illustrates the different sizes of the extracted subgraphs under different hop counts. Figure 2In this diagram, black nodes represent target prediction node pairs, while dark gray, light gray, and white nodes represent the 1-hop, 2-hop, and 3-hop neighbors of the target node, respectively. Firstly, previous work has shown that model performance typically doesn't improve further after the hop count exceeds 3. Although there's only a one-hop difference between 1-hop and 2-hop, on a sizable dataset, the number of nodes and edges from the 2-hop neighborhood increases exponentially. This expansion of redundant information increases storage and computational overhead, negatively impacting model performance. Secondly, a fixed hop count that's too small may fail to cover potentially crucial structural information, resulting in a lack of training samples for the model. Besides the potential for redundancy in the number of subgraphs due to the fixed hop count setting, the redundancy of information contained within the subgraph is also a significant factor affecting model performance. Existing hop count-based subgraph extraction methods often simply include all neighboring nodes within the hop count range. Although some methods introduce random sampling to reduce the size of nodes and edges, the lack of targeted screening still leads to the introduction of redundant information. Based on the above analysis, a hop count selection and differential sampling fusion strategy driven by structural centrality is proposed to adaptively perceive the neighborhood. Specifically, in the task of link prediction, the predicted edges... Subgraph extraction is performed, defined in the hop count. Next node and nodes The neighborhood information is The calculation is shown in equation (1): (1); in, Represents a node and nodes The shortest path length between them, and the node centrality is calculated using equation (2). , The model is composed of betweenness centrality and normalized centrality index. These two are weighted and fused to model neighborhood information. Betweenness centrality measures the importance of a specific node as an intermediary in the shortest path between other nodes in the graph. The specific definition is shown in equation (3). This indicates that from the node To the node Total number of paths to the shortest path: (2); (3); Then, define the current hop count. Neighborhood structure perception score And use it as a selection factor for subgraph probability sampling. The calculation is shown in equations (4) and (5), where, This is the maximum number of hops defined in subgraph extraction: (4); (5); However, when fusing multi-hop neighbor subgraph information, significant structural nesting exists between neighborhoods with different hop counts, i.e. Neighborhood information with a larger hop count often contains neighborhood information with a smaller hop count, leading to information overlap. To address this issue, we reconstruct the neighborhood subgraph for hop counts greater than 1, first defining the differential neighborhood domain. This indicates that the current hop count is The time is based on the newly added node information in the previous hop count, and its definition is shown in equation (6), where : (6);

[0019] Finally, in order to preserve the complete neighborhood structure, the node set... Generate mask matrix And this is used as differential information for each node. If it belongs to the difference set ,but Otherwise, it is 0. Finally, for the masked result... According to the selection factor of its corresponding number of hops Probability sampling is performed, and each subgraph is constructed as shown in equation (7). : (7); in, Indicates the original image edge set in exist The set of edges obtained by the up-induced curve retains the original topological structure of the subgraph, and as shown in equation (8), it further constitutes the complete set of subgraphs. : (8); 1.1.2 Feature Generation After obtaining the set of subgraphs corresponding to all candidate links in the original graph, the rich contextual information in the subgraphs is further modeled. In the feature generation module, structural side information is introduced to assist training and enhance the model's ability to model the structural information of nodes and edges. Specifically, a node position label mapping function is used to apply the DRNL algorithm to all nodes in each closed subgraph for node labeling. The specific calculation method is as follows: (9); in, In addition, nodes with special circumstances will be handled according to the following rules: (10); This method calculates the shortest path length between any node and the target node in the subgraph, thereby marking its relative position in the subgraph and obtaining the structural information of the node. Furthermore, it concatenates the information of the two end nodes with an edge in the subgraph to obtain the edge's label information, as shown below: (11); Among them, nodes and nodes From the same subgraph And there are edges in this subgraph. Based on the above, this invention provides high-quality input information for subsequent feature representation learning, which helps to improve the accuracy and robustness of link prediction tasks. 1.1.3 Hypergraph Convolution When a task requires high-order structural information, it is necessary to introduce lateral information. In graph neural network tasks, single-node features tend to express "individual attributes." For example, in academic networks, a node represents a paper, and its features may be bag-of-words vectors or embedding vectors, reflecting the text content of the paper itself. In social networks, a node represents a user, and its features may be age, interests, or user profile information. These features are usually limited to the expression of the node's own attributes and cannot effectively characterize the structural position of the node in the graph. However, in tasks such as link prediction and community detection, the model often needs to understand the number of common neighbors between nodes in the graph, the degree of local clustering, and the similarity of path structures. This information relies on high-order structural features and can reflect the local or global structural pattern to which the node belongs, rather than just the individual attributes of the node itself. However, as the number of layers in a neural network increases, the model theoretically gains the ability to capture higher-order data information. This, to some extent, introduces the risk of oversmoothing, because excessive message passing can lead to homogenization of node representations, resulting in the loss of some information. Consequently, the model struggles to distinguish nodes at different structural positions. Therefore, in order to simplify the network architecture and capture higher-order structural information while reducing the number of model layers, this model uses the output of the node labeling function as the side information of the node structure in the graph, further expressing the edge information. By utilizing the idea of ​​dual hypergraph transformation, higher-order information is modeled. Compared to ordinary graphs, hypergraphs can naturally express one-to-many relationships and achieve higher-order structural aggregation in single-step information propagation. Dual hypergraph transformation is a line graph-based method that treats nodes in the original graph as hyperedges in the hypergraph, and edges in the original graph as hypernodes in the hypergraph. Specifically, the original graph is first defined as... Super Image , and Let represent the sets of nodes and edges in the original graph, respectively. and Let represent the sets of supernodes and superedges in the hypergraph, respectively, as defined in equation (12): (12); Furthermore, the hyperedge features and hypernode features in a hypergraph are represented as follows: and The definition is shown in equation (13): (13); According to equations (14) and (15), the hyperedge features and hypernode features are independently convolutional, and after passing through several hidden layers, they are concatenated to obtain the hypergraph feature representation. : (14); (15); in, and The weight matrix is ​​a learnable matrix. and , For activation function, Let be the normalized adjacency matrix of the hypergraph. H is the incidence matrix of the hypergraph, I is the identity matrix, D is the node degree matrix, and B is the edge degree matrix; 1.2 Attribute and Topological Feature Generation Given that link generation is influenced by both node attributes and graph structure information, this invention redesigns the network embedding module to learn node representations more effectively. Specifically, it calculates the similarity matrix between any two nodes in the attribute space. Each element Defined as: (16); in, and Representing nodes respectively and nodes The attribute vector, i.e., the attribute matrix of the original graph nodes. The row and number Okay, next, in order to build a node feature-based... Nearest neighbor graph, for each node, select the previous one according to (16). The most similar nodes are selected as neighbors to form a set of neighbors. Nearest neighbor graph Subsequently, GAT was used to analyze the graph. Encode the nodes to obtain the node representations for each layer, and update the rules as shown in (17): (17); The present invention further uses an average aggregator, thereby representing the nodes of each layer. Specifically, as shown in (18): (18); in, As the initial input features, Represents a node The neighborhood group, This represents the activation function. This represents a vector concatenation operation, ultimately forming... , ; Similarly, the adjacency matrix A of the original input graph G can be directly used as the topological graph. The structure is represented to obtain link topology information, and the same graph neural network is used for encoding to obtain node representations from the topology graph. This process is related to The encoding process shares network parameters to ensure consistency of the representation space under different views; 1.3 Feature Fusion and Link Prediction The feature representations generated from the three different perspectives mentioned above are fused using an adaptive attention mechanism to enhance the discriminative ability in link prediction tasks. Specifically, the hypergraph modeling representation is used... , Nearest neighbor graph modeling representation and node feature representation under topology graph modeling The common inputs are fed into the attention fusion module to learn the importance weights of different perspectives on the final representation. The attention fusion formula is expressed in the following form: (19); in, Represented as hypergraph, respectively Attention weights for the nearest neighbor graph and the topological graph in the final fused representation. These weights can intuitively reflect the contribution of features from different perspectives to different node pairs. Therefore, the final representation Z is as follows: (20); In the subsequent prediction stage, the cosine similarity method is used to evaluate the potential connection probability between node pairs, as shown in Equation (21), which assigns nodes... u and nodes v The final representation is input into the cosine similarity calculation formula to obtain the similarity score. : (twenty one).

[0020] 2 Experiments The experimental platform for this invention is a server equipped with a Linux system, an Intel i9-13900K CPU with a clock speed of 2.995 GHz, an NVIDIA GeForce RTX 4090 GPU, 48 GB of memory, and is developed using the PyG graph neural network library based on PyTorch, with Python 3.10 as the runtime environment.

[0021] 2.1 Dataset This invention uses two standard datasets from citation networks, Cora and PubMed, as well as a coauthor network dataset, CoauthorCS. In the citation network datasets, the Cora dataset includes 2708 papers from 7 categories and 5429 links between papers that cite each other. Each paper document is described by a 1433-dimensional vector. Similar to the Cora dataset, the PubMed dataset has 19717 nodes and 44338 edges corresponding to paper documents and undirected citation relationships, respectively. The 500 node features are represented as bag-of-words representations of the documents. The specific information of the datasets is shown in Table 1.

[0022] Table 1. Specific parameters of the dataset used in the experiment.

[0023]

[0024] 2.2 Baseline Model To verify the effectiveness of the model proposed in this invention, the following baseline models are used for comparison with this model: CPAGCN utilizes an adaptive graph convolutional network to fuse link information and attribute information to obtain node representations, and preserves the community structure in the node representations through graph reconstruction loss; LGLP is a graph neural network method based on subgraph extraction, which converts the extracted subgraphs into corresponding line graphs, thereby transforming the link prediction task into a node classification task. HGLP learns the high-order features of hypernodes and hyperedges in the hypergraph through a hypergraph convolutional network. TADW combines the random walk concept from DeepWalk with text features, and uses matrix factorization to combine attribute information and structural information to learn node representations. GAT improves the model's ability to express structural information by introducing an attention mechanism that adaptively assigns weights to neighboring nodes. GraphSAGE is a graph neural network method based on neighbor sampling and aggregation. It learns aggregation functions during training to generate node representations, thereby enabling the learning of inductive representations of unseen nodes in large-scale graphs. The above models and the model proposed in this invention were tested in the same environment. The hyperparameter settings of the baseline model were set with reference to the settings in the original paper. In order to make a fair comparison, the experiment was repeated ten times for each method using the optimal settings, and the average value was obtained.

[0025] 2.3 Evaluation Indicators To evaluate the statistical results across all possible thresholds, two commonly used evaluation metrics are further introduced: Area Under the Curve (AUC) and Average Precision (AP). AUC measures the probability that, among all positive and negative sample pairs, the model's prediction score for positive samples (links that exist but are not observed in the training set) is higher than that for negative samples (links that do not exist at all). AP calculates the integral result under the precision-recall curve, aiming to obtain the average precision across all recall rates, as defined below: (twenty two); 2.4 Experimental Design To enhance the reproducibility of this model, the hyperparameter settings mainly used in this model will be introduced. Throughout the experiments, the Adam optimizer was used with a weight decay of 0.0005 and an embedding dimension of 64. For the Cora dataset, the learning rate was set to 0.015, the number of training iterations was 100, the batch size was 8, and hypergraph sampling was not performed. For the PubMed dataset, the learning rate was set to 0.005, the number of training iterations was 100, the batch size was 4, and hypergraph sampling was 20. For the Coauthor dataset, the number of training iterations was 150, the batch size was 4, the learning rate was set to 0.01, and hypergraph sampling was 50. For fair comparison, the embedding dimension of all baseline models was set to 64, and all parameters were set to their default values. LGLP and HGLP were fixed to extract second-order neighborhood subgraphs, and the convolutional channel sizes for the four graphs were set to 32, 32, 32, and 1, respectively.

[0026] 2.5 Experimental Results and Analysis 2.5.1 Comparative Experiment To evaluate the performance of each method, this invention randomly divides the links in the edge set E into 10%, 20%, and 70% as positive samples in the training set, and the remaining portion as positive samples in the test set. Simultaneously, the set of unobserved links is also included. Equal amounts of negative samples were randomly selected and added to both the training and test sets. The proposed model was compared with several baseline models, and the AUC and AP results are shown in Tables 2 and 3. Under different datasets and training set ratios, the proposed model demonstrated good performance. Even with limited training data (TSS) of 10% and 20%, the model consistently ranked among the top performers across all comparison methods. It also achieved competitive results with a high training set ratio of 70%. Specifically, on the PubMed dataset with small sample training, the proposed model improved the AUC by at least 4.3% and AP by at least 3.5% compared to other methods. On the Cora dataset with small sample training, it improved the AUC by at least 2.6% and AP by at least 1.2%. Particularly noteworthy were the significantly improved AUC and AP values ​​on the Cora dataset. The p-values ​​of this model are superior to other baseline methods. In addition, the PubMed dataset has dense and high-dimensional node attribute information. The adaptive graph convolutional network in the CPAGCN model has an advantage in fusing structural and attribute information, which makes it slightly better than this model in 70% of the training set experiments. Coauthor is an academic collaboration network dataset with a relatively regular structure of edges and nodes. Therefore, it converges faster during training than the single LGLP. Based on the above experiments, this model can still maintain good prediction ability in networks with complex structures and high-dimensional attributes, and has strong cross-scene adaptability, thus proving the effectiveness of this model.

[0027] Table 2 Comparison of AUC values ​​for each model under different training set sizes (TSS%)

[0028] Table 3 Comparison of AP values ​​for each model under different training set sizes (TSS%)

[0029]

[0030] 2.5.2 Ablation Experiment To verify the effectiveness of the model proposed in this invention, five variant models of the algorithm will be designed: w / o H-node: Remove convolutions on supernodes and model only superedges to verify the role of supernode features in high-order relation modeling; w / o H-edge: Remove convolutions on hyperedges and model only hypernodes to verify the role of hyperedge features in high-order relation modeling; w / o k-NN: Omitted based on k The perspective of the nearest neighbor graph was used to verify the complementary role of local similarity in model performance. w / o Topo: Omit the perspective of the original topology graph to verify the role of adjacency structure information in multi-view fusion; w / o SA-Sampling: Replace structure-aware subgraph sampling with fixed-hop-count sampling to verify the effectiveness of structure-centrality-driven subgraph sampling in selecting key neighbors; The experimental results are shown in Table 4. First, the performance of the w / o H-edge model is significantly reduced, with the AUC and AP values ​​being the lowest on all three datasets. This indicates that in hypergraph convolutional networks, relying solely on hypernode information is insufficient to learn effective edge representations, as the important structural information provided by hyperedges is ignored. In contrast, while the experimental performance of the w / o H-node model is generally better than that of the w / o H-edge model, it still shows a performance decrease compared to the complete model. This step skips the node annotation and edge feature construction process to remove the edge structural information brought by node annotations, verifying that the lateral information of node annotations plays an important role in edge representation learning. Furthermore, the w / o k-NN model and the w / o Topo model also caused significant performance degradation, indicating... k The neighbor graph provides effective supplementary information on local adjacency structure, and the synergistic effect of the two still has a positive impact on the overall performance of the model. Finally, the performance of the w / o SA-Sampling model on the three datasets is lower than that of the complete model, which effectively verifies the positive role of the structure centrality-driven subgraph sampling strategy in selecting key neighbors. Compared with fixed hop number sampling, it can capture important structural features in the network more accurately. In order to more intuitively show the differences in prediction performance of the SEH-LP variants, Figure 3 shows the numerical performance of each variant model under the evaluation metrics AUC and AP in the form of a line graph.

[0031] Table 4 Ablation Experiment Results

[0032]

[0033] 2.5.3 Other variant experiments In the backbone network architecture, GAT is used as the graph neural network method for representation learning. Two other mainstream graph neural networks, GCN and GraphSAGE, are introduced as alternatives to GAT to compare their performance in link prediction tasks. Experimental results are as follows: Figure 4As shown, the results indicate that GAT achieved the best experimental performance on all datasets, demonstrating that its adaptive weight allocation strategy is more conducive to improving prediction performance. In contrast, GCN and GraphSAGE lack certain feature discrimination capabilities and are more prone to information confusion compared to GAT, thus they are not suitable for the task applied in this model. In addition, this invention also used different prediction score calculation methods for comparative experiments, employing three methods as prediction modules: Dot Product (DP), MLP, and Cosine Similarity. The experimental results are as follows: Figure 5 As shown, the experimental results indicate that the Cosine method performs best on all datasets, demonstrating that the lightweight nonparametric scoring method is more in line with task requirements and focuses more on utilizing the expressive power of structural information itself, rather than using multi-layer neural networks to learn complex prediction functions, thus reducing the risk of overfitting. 2.5.4 Parameter Analysis This invention performs parameter analysis on the embedding dimension of three typical datasets (Cora, PubMed, and CoauthorCS), and conducts comparative experiments on the model under different dimensions (32, 64, 128, and 256) to evaluate the impact of different embedding dimension parameters on link prediction performance. Experimental results... Figure 6As shown in the figure, setting the embedding dimension from the minimum of 32 to 256 results in an 8-fold difference between the minimum and maximum dimensions. Overall, different embedding dimensions have a significant impact on the model. The datasets show some commonalities in their optimal dimension selection, achieving better prediction results when the embedding dimension is set to 64. However, they also exhibit considerable differences; when the embedding dimension is too high or too low, the model performance often suffers. For the Cora dataset, the performance of 32, 128, and 256 dimensions is significantly lower than that of 64 dimensions. This is because the Cora dataset has a higher number of nodes. Given the characteristics of a relatively small number of data points, a relatively balanced class distribution, and a dense graph structure, smaller embedding dimensions (e.g., 32 dimensions) are insufficient to learn adequate structural information and feature representations. Larger embedding dimensions (e.g., 128 or 256 dimensions) introduce redundant information, increasing the risk of overfitting during network training and thus leading to poor performance. On the PubMed dataset, the model achieved the same best performance with an embedding dimension of 64. Unlike the Cora dataset, the performance of the 32-dimensional and 128-dimensional models is roughly equal, slightly lower than that of the 64-dimensional model. Because the PubMed dataset has a larger node size (19,717 nodes) and only three classes, 64-dimensional embeddings can express information relatively evenly. When the dimension increases to 256, the sparse graph structure makes the training sample density insufficient to support the high-dimensional representation after output, resulting in a significant drop in performance. The CoauthorCS dataset shows performance roughly equivalent to the PubMed dataset. The embedding performance at 128 dimensions is slightly better than that at 64 dimensions. This is because the CoauthorCS dataset is a large-scale cooperative network dataset with high edge density, rich structural information and obvious community partitioning characteristics. More importantly, it has 15 classes, which is significantly higher than Cora (7 classes) and PubMed (3 classes). Based on this, such a semantically complex dataset can help the model capture more fine-grained feature differences and complex inter-class structures at a higher embedding dimension (128 dimensions). At the same time, the larger data scale also alleviates the risk of overfitting that may be caused by high-dimensional representation to some extent. Conversely, the model has difficulty effectively distinguishing node features under high class complexity when the embedding dimension is 32, and its expressive power is limited. 2.5.5 Subgraph Extraction Optimization Resource Consumption Analysis

[0034] Table 5. Comparison of storage usage between SEH-LP-A and SEH-LP-N on different datasets.

[0035]

[0036] To further evaluate the resource scheduling efficiency brought about by the adaptive neighborhood-aware strategy in removing redundant information during subgraph extraction, two comparative models, SEH-LP-A and SEH-LP-N, were designed. SEH-LP-A was set as the model for the adaptive neighborhood-aware subgraph extraction strategy, while SEH-LP-N was set as the model for subgraph extraction using a fixed number of hops. For ease of observation, the maximum number of hops for SEH-LP-A and the fixed number of hops for SEH-LP-N were both set to 2. As shown in Table 5, Files refers to the cache file size generated after subgraph extraction. The total size on disk, in MB, is used to measure the performance of different subgraph extraction strategies in data representation compression. Specifically, SEH-LP-A shows better storage usage on both datasets, especially the PubMed dataset, where the total storage capacity of its subgraph representation is reduced by 294.39 MB compared to SEH-LP-N. This demonstrates the advantage of this strategy in compressing redundant structures while retaining certain core feature information. In summary, from the perspective of space resource usage, the effectiveness of the proposed adaptive neighborhood-aware subgraph extraction strategy in terms of modeling compactness and lightweightness is confirmed.

[0037] This invention addresses the challenges of high-order structure modeling and multi-view information fusion in link prediction tasks. It proposes a hypergraph link prediction method based on structure awareness and edge feature fusion. This method achieves dynamic sampling by introducing an adaptive subgraph extraction mechanism driven by structure centrality, avoiding information redundancy and omission caused by fixed hop count. It utilizes dual hypergraph modeling to transform edges into nodes in the hypergraph, effectively alleviating the oversmoothing problem of graph neural networks under multi-layer stacking. Finally, it combines graph embedding representations from attribute and topology perspectives, and achieves multi-view information fusion through a lightweight cross-view attention mechanism, thereby improving the model's discriminative ability.

[0038] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A hypergraph link prediction method based on structure awareness and edge feature fusion, characterized in that: Includes the following steps: S1. Hypergraph Convolutional Network Based on Structure-Aware Subgraph Sampling: For the neighborhood information of the two end nodes of the target link, the centrality index of the node is calculated to dynamically optimize the hop count of the subgraph extraction, and the complete neighborhood subgraph is extracted in this way. The most relevant neighbor nodes within the range are selected to construct local context information. Then, a dual-radius node labeling algorithm is used to label each node in each subgraph. After obtaining the end node labeling information of the edge, it is concatenated in the form of one-hot encoding to obtain the edge features. Finally, the subgraph is transformed into a dual hypergraph and input into the hypergraph convolutional network to obtain the hypergraph representation. S2. Attribute and Topological Feature Generation: Calculate the similarity matrix between any two nodes in the attribute space, select the most similar nodes as neighbors for each node to form a nearest neighbor graph, encode the nearest neighbor graph using GAT, obtain the node feature representation of each layer, and use an average aggregator to represent the node features. S3. Feature Fusion and Link Prediction: Feature representations generated from different perspectives are fused using an adaptive attention mechanism to enhance the discriminative ability in the link prediction task.

2. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: The hypergraph convolutional network based on structure-aware subgraph sampling in step S1 is constructed through the following three-stage process: S1-1, Subgraph Extraction: Adaptive neighborhood-aware subgraph extraction through a fusion strategy of hop selection driven by structural centrality and differential sampling; S1-2, Feature Generation: After obtaining the set of subgraphs corresponding to all candidate links in the original graph, the rich contextual information in the subgraphs is further modeled. In the feature generation module, structural side information is introduced to assist training and enhance the model's ability to model the structural information of nodes and edges. S1-3, Hypergraph Convolution: The output of the node labeling function is used as the side information of the node structure in the graph, and the edge information is further expressed. Using the idea of ​​dual hypergraph transformation, the nodes in the original graph are regarded as hyperedges in the hypergraph, and the edges in the original graph are regarded as hypernodes in the hypergraph, thus modeling higher-order information.

3. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: In step S2, the adjacency matrix of the original input graph is directly used as the structural representation of the topology graph to obtain link topology information, and the same graph neural network is used for encoding to obtain the node feature representation from the topology graph. The encoding process shares network parameters to ensure the consistency of the representation space under different views.

4. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: In step S3, the hypergraph modeling representation, the nearest neighbor graph modeling representation, and the node feature representation under the topological graph modeling are jointly input into the attention fusion module to learn the importance weights of different perspectives to the final representation. The importance weights directly reflect the contribution of features from different perspectives to different node pairs. In the subsequent prediction stage, the cosine similarity method is used to evaluate the potential connection probability between node pairs.

5. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: In step S1-1, in the link prediction task, subgraph extraction is performed on the predicted edges. Nodes and their neighborhood information are defined under the hop count. Node centrality is calculated, which is composed of betweenness centrality and normalized centrality index. These two are weighted and fused to model the neighborhood information. Betweenness centrality is used to measure the importance of a specific node as an intermediary in the shortest path of other nodes in the graph. The neighborhood structure perception score under the current hop count is defined and used to calculate the selection factor for subgraph probability sampling. The neighborhood subgraph when the hop count is greater than 1 is reconstructed. First, the differential neighborhood is defined, which represents the node information added based on the previous hop count when the current hop count is 1. A mask matrix is ​​generated for the node set and used as the differential information. The node set after masking is probabilistically sampled according to the selection factor of its corresponding hop count, further forming a complete subgraph set.

6. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: In steps S1-2, a node position label mapping function is used to label all nodes in each closed subgraph using the DRNL algorithm. By calculating the shortest path length between any node in the subgraph and the target node, the relative position of the node in the subgraph is marked, thus obtaining the structural information of the node. Furthermore, the information of the two end nodes with edges in the subgraph is concatenated to obtain the edge label information.

7. The hypergraph link prediction method based on structure awareness and edge feature fusion as described in claim 1, characterized in that: In S1-3, the set of nodes and edges of the original graph, the set of super nodes and super edges of the hypergraph, and the super edge features and super node features of the hypergraph are first defined. The super edge features and super node features are then subjected to independent convolution operations and concatenated after passing through several hidden layers to obtain the hypergraph feature representation.

Images