Semi-supervised node classification method and system based on tensor hypergraph attention network

Abstract

The application discloses a semi-supervised node classification method and system based on a tensor hypergraph attention network, and belongs to the technical field of hypergraph neural networks.The application constructs a hypergraph sparse adjacency tensor to represent high-order topology, adopts a two-stage space message passing mechanism, models high-order feature interaction in a hyperedge through tensor operation, obtains node neighborhood embedding through attention weighted aggregation, and outputs a classification result after iteratively updating node features.The application models the high-order structure of a hypergraph based on tensor theory, balances the performance and efficiency of the model by using tensor decomposition, and effectively improves the semi-supervised node classification precision and model scalability.

Description

Technical Field

[0001] This invention relates to the fields of hypergraph representation learning and hypergraph neural network technology, and discloses a semi-supervised node classification method and system based on tensor hypergraph attention network. Background Technology

[0002] Graph-structured data is a core framework for modeling complex relationships and interactions between entities in the real world, and it is widely used in numerous practical scenarios. In recent years, graph neural networks have become the core paradigm for graph representation learning, achieving significant application results in fields such as computer vision, data mining, and bioinformatics. However, traditional graph neural networks model based on the assumption of pairwise interactions, meaning that edges in the graph can only connect two entities. This simplification cannot capture the diverse and high-order interactions that are prevalent in real-world scenarios—such as the synergistic effects of multiple metabolites in biochemical reactions, the joint influence of multiple regions in traffic dynamics, and the linkage relationships of multiple joints in human skeletal movements. These high-order interactions involving multiple participants require modeling frameworks with stronger expressive capabilities.

[0003] As an extension of ordinary graphs, hypergraphs allow hyperedges to flexibly connect two or more nodes, fully preserving the high-order dependencies lost in traditional graph structures and providing a natural structural foundation for high-order interaction modeling. Hypergraph neural networks, by extending the principles of graph neural networks to hypergraph structures, enable representation learning of hypergraph data and have become the mainstream technology for high-order graph data processing.

[0004] Most existing implementations of hypergraph neural networks rely on unfolding techniques to transform the hypergraph into a regular graph, with clique unfolding and star unfolding being the most common strategies. Clique unfolding transforms hyperedges into fully connected subgraphs, allowing traditional graph neural networks to be directly applied to hypergraph scenarios. While this approach simplifies the computational process, the mapping between the hypergraph and its unfolded regular graph is non-injective, leading to changes and loss of the hypergraph's inherent topological properties and failing to fully preserve high-order structural information.

[0005] To address the issue of topological information loss during clique unfolding, existing technologies increasingly employ star-shaped unfolding schemes. This approach encodes the hypergraph topology through hyperedges and uses a two-stage spatial message passing mechanism to iteratively propagate information from nodes to hyperedges and back again, thus preserving the structural integrity of the hyperedges to some extent. However, these unfolding-based methods fail to consider the potential high-order feature interactions between multiple nodes within the hyperedge, achieving only basic feature aggregation. They cannot fully exploit and utilize the inherent high-order information and complex dependencies within the hypergraph, resulting in insufficient feature representation capabilities and limited classification accuracy in semi-supervised node classification tasks.

[0006] In hypergraph signal processing research, tensor operations can fully preserve the inherent topological structure of a hypergraph while performing complex mathematical analysis. This paradigm shift from matrix representation to tensor representation provides a new path for capturing higher-order relationships in hypergraphs. Currently, some studies have attempted to model hypergraphs as tensors to achieve higher-order interaction modeling; however, such approaches generally suffer from the curse of dimensionality: as the order of the hypergraph and the length of the hyperedges increase, the number of model parameters grows exponentially, leading to extremely high computational complexity and memory consumption. This not only makes deployment difficult but also easily causes model overfitting, making it impossible to achieve a balance between performance and efficiency in semi-supervised node classification tasks.

[0007] In summary, existing hypergraph neural network technologies have the following technical problems: clique-based expansion schemes lose the inherent topological properties of the hypergraph and cannot fully preserve high-order structural information; star-based expansion schemes do not model the high-order feature interactions of nodes within the hyperedge, failing to fully utilize the high-order information of the hypergraph, resulting in insufficient feature representation capabilities and limited classification accuracy of semi-supervised nodes; existing tensor hypergraph modeling schemes suffer from the curse of dimensionality, with the number of parameters increasing exponentially with the order of the hypergraph, leading to huge computational and memory overhead, making it difficult to balance model expressiveness and operational efficiency. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a semi-supervised node classification method and system based on tensor hypergraph attention networks. This method solves the problems of topological information loss, insufficient utilization of higher-order information, and the curse of dimensionality in existing technologies, while achieving high-precision classification of unknown nodes in hypergraphs under semi-supervised scenarios.

[0009] To achieve the above-mentioned technical effects, the technical solution adopted by this invention is: a semi-supervised node classification method based on tensor hypergraph attention networks, comprising: Obtain the hypergraph to be classified, wherein the hypergraph includes a set of nodes and a set of hyperedges; Based on the set of nodes and the set of hyperedges, construct the sparse adjacency tensor of the hypergraph; Based on the sparse adjacency tensor, a two-stage spatial message passing is performed on the input features of each node in the node set to obtain the neighborhood embedding of the central node. Based on the neighborhood embedding of the central node, the features of the central node are iteratively updated to obtain the updated node features; Based on the updated node features, output the classification results of the unknown nodes in the hypergraph.

[0010] As a preferred implementation, constructing the sparse adjacency tensor of the hypergraph includes: For each node in the hypergraph, determine the set of associated hyperedges that contain that node; Each hyperedge in the set of associated hyperedges is processed to generate an M-order set of associated hyperedges for that node, where M is a preset hypergraph order. Remove the node from each of the M-order associated hyperedges in the set of M-order associated hyperedges to obtain the M-order neighborhood of the node; Construct the sparse adjacency tensor based on the M-order neighborhood of all nodes; The expression for the M-order neighborhood of the node: ; in, For the M-order neighborhood of the central node v, It is an M-order interconnected hyperedge. From The edges obtained after removing the center node in the middle, Let v be the set of associated hyperedges of the central node v.

[0011] As a preferred implementation, the input features of each node in the node set are processed through a two-stage spatial message passing process, including: a node-to-hyperedge aggregation stage and a hyperedge-to-node aggregation stage. In the node-to-hyperedge aggregation stage, the features of multiple nodes belonging to the same hyperedge are fused through tensor operations to obtain the hyperedge embedding of the hyperedge. In the hyperedge-to-node aggregation stage, the hyperedge embeddings of all hyperedges associated with the central node are aggregated to obtain the neighborhood embedding of the central node.

[0012] As a preferred embodiment, obtaining the hyperedge embedding includes: The features of the multiple nodes are each multiplied by an m-modal product with a learnable weight tensor to obtain the hyperedge embedding. The hyperedge embedding expression: ; The slice tensors of the core tensor are obtained through CP decomposition, and the expression of the slice tensors is as follows: ; in, For hyperedge embedding, For factor matrix, For the core tensor, The symbol for the m-modal product operation. Let be the transformed eigenvector of the m-th node of the hyperedge. , ; Let R be the slice tensor, and R be the rank of CP. For learnable vectors, Indicates the outer product operation; The expression for calculating the k-th dimension of the intermediate vector of the hyperedge embedding is: ; in, Let k be the kth dimension of the intermediate vector.

[0013] As a preferred embodiment, obtaining the neighborhood embedding of the central node includes: Based on the input features of the central node and the embedding of each hyperedge, the learnable attention weights corresponding to each hyperedge are calculated. Based on the learnable attention weights corresponding to each hyperedge, the corresponding hyperedge embeddings are weighted and summed to obtain the neighborhood embedding of the center node. The expression for the learnable attention weights: ; The expression for the neighborhood embedding is: ; in, For learnable attention weights, The input features are those of the central node v. For neighborhood embedding.

[0014] As a preferred implementation, the updated node features are obtained, represented as follows: ; in, For the first The updated features of the central node in the layer The initial input features for the central node are... For the first The neighborhood embedding of the central node obtained from the layer.

[0015] Based on the above method, this invention also proposes a semi-supervised node classification system based on tensor hypergraph attention networks, including: The data acquisition module is used to acquire the hypergraph to be classified, wherein the hypergraph includes a set of nodes and a set of hyperedges; A sparse adjacency tensor construction module is used to construct the sparse adjacency tensor of the hypergraph based on the set of nodes and the set of hyperedges. The message passing module is used to perform two-stage spatial message passing on the input features of each node in the node set based on the sparse adjacency tensor to obtain the neighborhood embedding of the central node. The node feature update module is used to iteratively update the features of the central node based on the neighborhood embedding of the central node, so as to obtain the updated node features. The classification result output module is used to output the classification result of the unknown nodes in the hypergraph based on the updated node features.

[0016] As a preferred embodiment, the sparse adjacency tensor construction module is specifically used for: For each node in the hypergraph, determine the set of associated hyperedges that contain that node; Each hyperedge in the set of associated hyperedges is processed to generate an M-order set of associated hyperedges for that node, where M is a preset hypergraph order. Remove the node from each of the M-order associated hyperedges in the set of M-order associated hyperedges to obtain the M-order neighborhood of the node; The sparse adjacency tensor is constructed based on the M-order neighborhood of all nodes.

[0017] As a preferred embodiment, the message passing module includes: The node-to-hyperedge aggregation unit is used in the node-to-hyperedge aggregation stage to fuse the features of multiple nodes belonging to the same hyperedge through tensor operations to obtain the hyperedge embedding of the hyperedge. The hyperedge-to-node aggregation unit is used to aggregate the hyperedge embeddings of all hyperedges associated with the central node during the hyperedge-to-node aggregation stage to obtain the neighborhood embedding of the central node. The node-to-hyperedge aggregation unit is specifically used to: perform m-modal product operation on the features of the multiple nodes with a learnable weight tensor to obtain the hyperedge embedding.

[0018] The hyperedge-to-node aggregation unit is specifically used to: calculate the learnable attention weights corresponding to each hyperedge based on the input features of the central node and the embedding of each hyperedge; Based on the learnable attention weights corresponding to each hyperedge, the corresponding hyperedge embeddings are weighted and summed to obtain the neighborhood embedding of the center node.

[0019] As a preferred embodiment, the node feature update module obtains the updated node features in the following way: ; in, For the first The updated features of the central node in the layer The initial input features for the central node are... For the first The neighborhood embedding of the central node obtained from the layer.

[0020] Compared with the prior art, the present invention achieves efficient representation of the high-order topology of the hypergraph by constructing a sparse adjacency tensor, avoiding the huge memory occupation caused by explicitly constructing the adjacency tensor, while retaining the inherent high-order topology properties of the hypergraph, and solving the problem of topology information loss in existing clique expansion schemes. A two-stage spatial message passing mechanism based on tensor operations was designed. In the aggregation stage from node to hyperedge, the features of multiple nodes within the hyperedge are fused through m-modal product operations to capture high-order feature interactions between nodes. This fully explores the inherent high-order information and complex dependencies in the hypergraph, solving the problems of insufficient utilization of high-order information and limited feature expression ability in existing star-shaped unfolding schemes, and significantly improving the feature discrimination of semi-supervised node classification. A learnable attention mechanism is introduced in the hyperedge-to-node aggregation stage, which can adaptively assign weights to different hyperedges, quantify the contribution of different hyperedges to the feature update of the central node, and enable the model to focus on the interaction between key hyperedges and nodes, further improving the accuracy of feature aggregation and classification performance. By cascading Tucker decomposition and CP decomposition, two-level parameter compression was performed on the weight tensor and the core tensor, which alleviated the curse of dimensionality problem in tensor hypergraph modeling, significantly reduced the computational complexity and memory overhead of the model, and at the same time preserved the high-order interactive modeling capability of tensor operations, thus achieving a balance between model performance and running efficiency. In semi-supervised scenarios, model training can be completed using only nodes with known labels, achieving high-precision classification of unknown nodes in the hypergraph. Experiments on multiple datasets have verified that the classification accuracy of this invention is superior to existing mainstream hypergraph neural network models in multiple scenarios such as academic networks, political networks, 3D point cloud classification in computer vision, and human skeleton action recognition, demonstrating excellent generalization ability and practical application value. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the semi-supervised node classification method based on tensor hypergraph attention network of the present invention. Figure 2 This is a schematic diagram of the modules of the semi-supervised node classification system based on tensor hypergraph attention network of the present invention; Figure 3 This is a schematic diagram of the message passing module structure of the semi-supervised node classification system based on tensor hypergraph attention network of the present invention; Figure 4 This invention is based on a hypergraph neural network architecture using tensor attention mechanism; Figure 5 The present invention presents experimental results of a semi-supervised node classification method based on tensor hypergraph attention networks. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0023] refer to Figure 1 , Figure 4 Example 1: This example provides a semi-supervised node classification method based on tensor hypergraph attention networks. The specific steps are as follows: Step 1: Hypergraph Data Acquisition Get the hypergraph to be classified ,in For a set of nodes, , The total number of nodes; For a set of superedges, Each super edge For a set of nodes A non-empty subset may contain two or more nodes. Each node in the node set corresponds to an entity object in the business scenario, such as biological metabolites, traffic areas, or human key points. Each node has an input feature vector, which is the attribute feature encoding of the node. Each hyperedge in the hyperedge set corresponds to a multivariate high-order interaction relationship between entity objects. Simultaneously, known category labels of a small number of nodes in the node set are obtained for model training in a semi-supervised scenario. Step 2: Sparse Adjacency Tensor Construction Based on the set of nodes and the set of hyperedges, construct the sparse adjacency tensor of the hypergraph, specifically including: S21. For each node in the hypergraph Determine the set of associated superedges that contain the node. ,node It is the central node of this associated superedge set; S22. Process each hyperedge in the set of associated hyperedges to generate the set of associated hyperedges of order M for that node. Where M is the preset hypergraph order, and the value of M ranges from 1 to 2. , The longest hyperedge in the hypergraph contains the number of nodes. The generation rule for the M-order correlated hyperedge set is as follows: If it exceeds the edge If the number of nodes is less than M, the central node can be added repeatedly. In this way, the hyperedge is extended to order M to obtain an M-order correlated hyperedge; If it exceeds the edge If the number of nodes contained is equal to M, then this hyperedge is directly treated as an M-order associative hyperedge. If it exceeds the edge If the number of nodes included is greater than M, downsample the hyperedge and extract the nodes containing the center node. Given M nodes, generate an M-order interconnected hyperedge; Through the above processing, the original hypergraph is transformed into an M-uniform hypergraph, and the number of nodes of all hyperedges is unified to M, providing a unified structural basis for subsequent tensor operations; S23. From each M-order interconnected hyperedge in the set of M-order interconnected hyperedges Remove the central node The M-order neighborhood of this node is obtained. Its expression is: ; in, From The set of nodes obtained after removing the central node has M-1 nodes; S24. Based on the M-order neighborhoods of all nodes, construct a sparse adjacency tensor of the hypergraph, storing only non-zero adjacency relationships to avoid the memory consumption caused by explicitly constructing the full adjacency tensor, thus achieving efficient representation of high-order topology of the hypergraph.

[0024] Step 3: Two-stage spatial message passing to obtain the neighborhood embedding of the central node. Based on the constructed sparse adjacency tensor, a two-stage spatial message passing is performed on the input features of each node, including a node-to-hyperedge aggregation stage and a hyperedge-to-node aggregation stage, as detailed below: Phase 1: Aggregation from nodes to hyperedges The features of M-1 nodes belonging to the same M-order associated hyperedge are fused through tensor operations to obtain the hyperedge embedding of that hyperedge.

[0025] In this embodiment, the weight tensor is compressed in one stage based on Tucker decomposition, and then the core tensor is compressed in two stages through CP decomposition. This reduces the number of parameters while retaining the ability to model high-order interactions. The specific process is as follows: First, the input features of each node inside the hyperedge are projected to obtain the projected feature vector: ; in, Let be the input feature vector of the m-th node inside the hyperedge. Let m be the learnable factor matrix of the m-th mode. For the input feature dimension, For Tucker rank; Then, the projected node features are multiplied by the core tensor using the m-modal product to obtain the intermediate vector. The slice tensors of the core tensor are obtained through CP decomposition, and the expression for the slice tensors is as follows: ; in, For the core tensor, , , For CP rank, For learnable vectors, Indicates the outer product operation; The expression for calculating the k-th dimension of the intermediate vector is: ; Finally, through the factor matrix Projecting the intermediate vector yields the final hyperedge embedding: ; in, To output the factor matrix of the modes, The output dimension of the hyperedge embedding.

[0026] Phase 2: Aggregation from hyperedges to nodes Aggregate the hyperedge embeddings of all hyperedges associated with the central node to obtain the neighborhood embedding of the central node, specifically including: First, based on the input features of the central node and the embedding of each hyperedge, the learnable attention weights corresponding to each hyperedge are calculated, as expressed by: ; in, The input features of the center node v are LeakyReLU, a non-linear activation function, and its negative slope is set to 0.2 in this embodiment. Then, based on the calculated attention weights, a weighted sum is performed on the embeddings of all hyperedges associated with the central node to obtain the neighborhood embedding of the central node, expressed as: ; in, For learnable attention weights, For neighborhood embedding.

[0027] Step 4: Iterative update of node features Based on the neighborhood embedding of the central node, the features of the central node are iteratively updated to obtain the updated node features. The update expression is: ; in, The updated features of the central node in the l-th layer network. The initial input features for the central node are... In this embodiment, the neighborhood embedding is obtained from the l-th layer network. A splicing operation is used, and the number of network layers is set to 2 to avoid oversmoothing caused by stacking too many convolutional layers.

[0028] Step 5: Output classification results The updated node features are input into the fully connected layer and the softmax classification layer, and the classification results of unknown nodes in the hypergraph are output.

[0029] refer to Figure 5 This invention proposes a tensor-based hypergraph attention network, T-HAN, which utilizes high-order relation modeling for semi-supervised node classification in fields such as computer vision.

[0030] This invention was evaluated on six datasets across three categories: computer vision tasks (NTU2012 and ModelNet40), academic networks (Citeseer and DBLP-CA), and political networks (Senate and House). Compared with state-of-the-art hypergraph neural networks, the classification accuracy is as follows: Figure 5 As shown in the figure. This comprehensive evaluation verifies that the proposed method has good model generalization ability and classification performance in the hypergraph node classification task.

[0031] refer to Figure 2 , Figure 3 Example 2: This example provides a semi-supervised node classification system based on tensor hypergraph attention network to implement the method described in Example 1. The system includes a data acquisition module, a sparse adjacency tensor construction module, a message passing module, a node feature update module, and a classification result output module connected in sequence.

[0032] The data acquisition module is used to acquire the hypergraph to be classified, wherein the hypergraph includes a set of nodes and a set of hyperedges; A sparse adjacency tensor construction module is used to construct the sparse adjacency tensor of the hypergraph based on the set of nodes and the set of hyperedges. The message passing module has built-in node-to-hyperedge aggregation unit and hyperedge-to-node aggregation unit. The node-to-hyperedge aggregation unit is used to fuse the features of multiple nodes belonging to the same hyperedge through tensor operations to generate hyperedge embeddings. The hyperedge-to-node aggregation unit is used to perform weighted aggregation of hyperedge embeddings based on an attention mechanism to generate neighborhood embeddings of the central node. The node feature update module is used to iteratively update the features of the central node based on the neighborhood embedding of the central node, so as to obtain the updated node features. The classification result output module is used to output the classification result of the unknown nodes in the hypergraph based on the updated node features.

[0033] In this embodiment, the specific functional implementation of each module is completely consistent with the corresponding method steps in Embodiment 1, and will not be repeated here.

[0034] The above description is merely a specific embodiment of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0035] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A semi-supervised node classification method based on tensor hypergraph attention networks, characterized in that, include: Obtain the hypergraph to be classified, wherein the hypergraph includes a set of nodes and a set of hyperedges; Based on the set of nodes and the set of hyperedges, construct the sparse adjacency tensor of the hypergraph; Based on the sparse adjacency tensor, a two-stage spatial message passing is performed on the input features of each node in the node set to obtain the neighborhood embedding of the central node. Based on the neighborhood embedding of the central node, the features of the central node are iteratively updated to obtain the updated node features; Based on the updated node features, output the classification results of the unknown nodes in the hypergraph.

2. The semi-supervised node classification method based on tensor hypergraph attention network according to claim 1, characterized in that, Constructing the sparse adjacency tensor of the hypergraph includes: For each node in the hypergraph, determine the set of associated hyperedges that contain that node; Each hyperedge in the set of associated hyperedges is processed to generate an M-order set of associated hyperedges for that node, where M is a preset hypergraph order. Remove the node from each of the M-order associated hyperedges in the set of M-order associated hyperedges to obtain the M-order neighborhood of the node; Construct the sparse adjacency tensor based on the M-order neighborhood of all nodes; The expression for the M-order neighborhood of the node: ； in, For the M-order neighborhood of the central node v, It is an M-order interconnected hyperedge. From The edges obtained after removing the center node in the middle, Let v be the set of associated hyperedges of the central node v.

3. The semi-supervised node classification method based on tensor hypergraph attention network according to claim 2, characterized in that, Two-stage spatial message passing is performed on the input features of each node in the node set, including: a node-to-hyperedge aggregation stage and a hyperedge-to-node aggregation stage. In the node-to-hyperedge aggregation stage, the features of multiple nodes belonging to the same hyperedge are fused through tensor operations to obtain the hyperedge embedding of the hyperedge. In the hyperedge-to-node aggregation stage, the hyperedge embeddings of all hyperedges associated with the central node are aggregated to obtain the neighborhood embedding of the central node.

4. The semi-supervised node classification method based on tensor hypergraph attention network according to claim 3, characterized in that, The process of obtaining the hyperedge embedding includes: The features of the multiple nodes are each multiplied by an m-modal product with a learnable weight tensor to obtain the hyperedge embedding. The hyperedge embedding expression: ； The slice tensors of the core tensor are obtained through CP decomposition, and the expression of the slice tensors is as follows: ； in, For hyperedge embedding, For the factor matrix, For the core tensor, The symbol for the m-modal product operation. Let be the transformed eigenvector of the m-th node of the hyperedge. , ; Let R be the slice tensor, and R be the CP rank. For learnable vectors, Indicates the outer product operation; The expression for calculating the k-th dimension of the intermediate vector of the hyperedge embedding is: ； in, Let k be the kth dimension of the intermediate vector.

5. The semi-supervised node classification method based on tensor hypergraph attention network according to claim 4, characterized in that, Obtaining the neighborhood embedding of the central node includes: Based on the input features of the central node and the embedding of each hyperedge, the learnable attention weights corresponding to each hyperedge are calculated. Based on the learnable attention weights corresponding to each hyperedge, the corresponding hyperedge embeddings are weighted and summed to obtain the neighborhood embedding of the center node. The expression for the learnable attention weights: ； The expression for the neighborhood embedding is: ； in, For learnable attention weights, The input features are those of the central node v. For neighborhood embedding.

6. The semi-supervised node classification method based on tensor hypergraph attention network according to claim 5, characterized in that, The updated node features are represented as follows: ； in, For the first The updated features of the central node in the layer. The initial input features for the central node are... For the first The neighborhood embedding of the central node obtained from the layer.

7. A semi-supervised node classification system based on tensor hypergraph attention networks, characterized in that, include: The data acquisition module is used to acquire the hypergraph to be classified, wherein the hypergraph includes a set of nodes and a set of hyperedges; A sparse adjacency tensor construction module is used to construct the sparse adjacency tensor of the hypergraph based on the set of nodes and the set of hyperedges. The message passing module is used to perform two-stage spatial message passing on the input features of each node in the node set based on the sparse adjacency tensor to obtain the neighborhood embedding of the central node. The node feature update module is used to iteratively update the features of the central node based on the neighborhood embedding of the central node, so as to obtain the updated node features. The classification result output module is used to output the classification result of the unknown nodes in the hypergraph based on the updated node features.

8. The semi-supervised node classification system based on tensor hypergraph attention network according to claim 7, characterized in that, The sparse adjacency tensor construction module is specifically used for: For each node in the hypergraph, determine the set of associated hyperedges that contain that node; Each hyperedge in the set of associated hyperedges is processed to generate an M-order set of associated hyperedges for that node, where M is a preset hypergraph order. Remove the node from each of the M-order associated hyperedges in the set of M-order associated hyperedges to obtain the M-order neighborhood of the node; The sparse adjacency tensor is constructed based on the M-order neighborhood of all nodes.

9. The semi-supervised node classification system based on tensor hypergraph attention network according to claim 8, characterized in that, The message passing module includes: The node-to-hyperedge aggregation unit is used in the node-to-hyperedge aggregation stage to fuse the features of multiple nodes belonging to the same hyperedge through tensor operations to obtain the hyperedge embedding of the hyperedge. The hyperedge-to-node aggregation unit is used to aggregate the hyperedge embeddings of all hyperedges associated with the central node during the hyperedge-to-node aggregation stage to obtain the neighborhood embedding of the central node. The node-to-hyperedge aggregation unit is specifically used to: perform m-modal product operations on the features of the multiple nodes with a learnable weight tensor to obtain the hyperedge embedding. The hyperedge-to-node aggregation unit is specifically used to: calculate the learnable attention weights corresponding to each hyperedge based on the input features of the central node and the embedding of each hyperedge; Based on the learnable attention weights corresponding to each hyperedge, the corresponding hyperedge embeddings are weighted and summed to obtain the neighborhood embedding of the central node.

10. The semi-supervised node classification system based on tensor hypergraph attention network according to claim 9, characterized in that, The node feature update module obtains the updated node features in the following way: ； in, For the first The updated features of the central node in the layer. The initial input features for the central node are... For the first The neighborhood embedding of the central node obtained from the layer.

Images