A method, system, device and medium for node classification across hypergraphs

Abstract

This application discloses a method, system, device, and medium for cross-hypergraph node classification, relating to the fields of hypergraph learning and transfer learning. The method includes: constructing a source hypergraph containing labeled nodes and a target hypergraph containing unlabeled nodes; generating unified node representations for both types of hypergraphs using a model incorporating higher-order and local consistency aggregators and attention mechanisms; training a node classifier by combining edge and conditional distribution alignment; and predicting the node categories of the target hypergraph using the trained classifier. This application integrates local and higher-order association features and improves the classifier's transfer capability through distribution alignment, effectively solving the problem of cross-hypergraph knowledge transfer in scenarios with scarce labels, significantly improving target domain classification performance, and is applicable to higher-order relational data such as academic networks, biological interactions, and social groups.

Description

Technical Field

[0001] This application belongs to the field of hypergraph learning and transfer learning technology, and in particular relates to a method, system, device and medium for node classification across hypergraphs. Background Technology

[0002] Hypergraphs (or hypergraph networks) are a generalization of standard graphs, consisting of a set of nodes and a set of hyperedges. Unlike standard graphs, which can only represent pairwise relationships between samples, hypergraphs can represent higher-order relationships between samples, meaning a single hyperedge can connect more than two nodes. This characteristic enhances the ability and flexibility of hypergraphs in modeling complex data relationships. Therefore, hypergraph learning is becoming increasingly popular and attracting more and more attention.

[0003] In hypergraph learning, semi-supervised node classification is one of the mainstream research paradigms. This paradigm trains a classifier based on a small number of labeled nodes and then predicts the category of the remaining unlabeled nodes. To adapt to this paradigm, researchers have proposed various technical solutions: some solutions encode higher-order data relationships through the hypergraph structure and extend convolution operations to the hypergraph learning process based on tools such as the hypergraph Laplacian operator and truncated Chebyshev polynomials; others first transform the hypergraph into a standard graph and then optimize the classification performance by comparing the learning representations of each network layer of the standard graph with those of the hypergraph. These existing studies have achieved certain results in semi-supervised hypergraph node classification tasks, but they still face significant challenges in practical applications.

[0004] The core issue lies in the high cost of label acquisition and the lack of adaptation in cross-hypergraph scenarios. On the one hand, collecting label information of nodes in a hypergraph often requires a large investment of manpower and resources, resulting in high costs. For newly formed hypergraphs, it is even more difficult to fully label nodes. On the other hand, in practice, there are many auxiliary hypergraphs (source hypergraphs) that are associated with the new hypergraph (target hypergraph) and have complete labels. For example, in citation networks, the ACM hypergraph with labeled research topics and the unlabeled DBLP hypergraph may have similar academic topic associations. In such scenarios, it is urgent to use the knowledge of the source hypergraph to assist in the node classification of the target hypergraph, i.e., cross-hypergraph node classification tasks.

[0005] However, existing technologies are insufficient to meet the practical needs of cross-hypergraph node classification: First, traditional unsupervised domain adaptation models are mainly designed for independent and identically distributed data, and cannot effectively handle structured data such as hypergraphs that have non-independent and identically distributed characteristics; Second, cross-network (cross-graph) node classification models that have emerged in recent years are essentially designed for standard graphs that can only represent pairwise relationships, and have failed to adapt to the high-order relationship representation needs of hypergraphs, and cannot fully explore the complex relationship information between nodes in hypergraphs.

[0006] In summary, cross-hypergraph node classification is a technical problem with significant practical value but insufficient research. Existing methods have obvious shortcomings in adapting to the structural characteristics of hypergraphs and utilizing source hypergraph knowledge to assist in target hypergraph classification. There is an urgent need to design a node classification method specifically for cross-hypergraph scenarios to solve the problem of efficient classification when the target hypergraph label is missing. Summary of the Invention

[0007] The technical problem to be solved by this application is to provide a method, system, device and medium for classifying nodes across a hypergraph, so as to solve the problems mentioned in the background art.

[0008] To address the aforementioned technical problems, this application provides the following technical solution: Firstly, this application provides a method for classifying nodes across a hypergraph, including: Construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; A preset local and higher-order consistency encoding and adaptation model is applied to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency aggregator, a local consistency aggregator and an attention mechanism module. The higher-order consistency aggregator includes: a higher-order consistency generator. Based on the unified node representations corresponding to the source hypergraph and the target hypergraph, a node classifier on the source hypergraph is trained by using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier. Based on the node classifier on the source hypergraph, the nodes of the target hypergraph are classified and predicted to obtain the node classification result of the target hypergraph.

[0009] Further, the step of applying a preset local and higher-order consistency encoding and adaptation model to the source hypergraph and the target hypergraph respectively to generate unified node representations corresponding to the source hypergraph and the target hypergraph includes: The source hypergraph and the target hypergraph are respectively constructed using the higher-order consistency generator to form a second source hypergraph and a second target hypergraph; For the source hypergraph and the target hypergraph, the local consistency aggregator is used to aggregate the features of each node in the local hyperedge neighborhood to generate the local consistency node representation of the source hypergraph and the local consistency node representation of the target hypergraph. For the second source hypergraph and the second target hypergraph, the higher-order consistency aggregator is used to aggregate the higher-order related information of each node in the global structural context, generating the higher-order consistent node representation of the source hypergraph and the higher-order consistent node representation of the target hypergraph. The attention mechanism module is used to perform a weighted fusion of the local consistency nodes and the higher-order consistency nodes of the source hypergraph to obtain a unified node representation corresponding to the source hypergraph; and to perform a weighted fusion of the local consistency nodes and the higher-order consistency nodes of the target hypergraph to obtain a unified node representation corresponding to the target hypergraph.

[0010] Furthermore, the training method for edge distribution alignment includes: constructing a domain discriminator and forcing the encoder to generate domain-confusing embeddings through gradient inversion training, so that the node classifier cannot rely on domain-specific features; the training method for conditional distribution alignment includes: constructing a source prototype based on the source domain and guiding the target domain nodes to approach their corresponding source prototypes, thereby explicitly transferring the semantic decision structure of the node classifier to the target domain.

[0011] Furthermore, the training method for edge distribution alignment employs the following loss function: ; in, This represents the number of nodes in the source hypergraph; Indicates the number of nodes in the target hypergraph; Indicates the first i Is each node from the source hypergraph or the target hypergraph? Indicates the first i The domain prediction results for each node.

[0012] Furthermore, the training method for conditional distribution alignment employs the following loss function: ; in, Indicates the number of categories, This indicates the number of target nodes marked as c; This represents the number of target nodes marked as k; It is expressed as a temperature parameter.

[0013] Furthermore, the step of classifying and predicting the enhanced node embedding representation corresponding to the target hypergraph based on the node classifier on the source hypergraph to obtain the node classification result of the target hypergraph uses the following loss function: ; ; in, Represents a node Category The predicted probability; This represents the softmax function. This represents the prediction result for the source hypergraph nodes. and This represents the trainable parameters.

[0014] Furthermore, the method also includes: using accuracy as an evaluation metric to evaluate the node classification results.

[0015] Secondly, this application also provides a node classification system across a hypergraph, comprising: A construction module is used to construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; The feature extraction module is used to apply a preset local and higher-order consistency encoding and adaptation model to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency generator, a local consistency aggregator, a higher-order consistency aggregator and an attention mechanism module. The training module is used to train a node classifier on the source hypergraph by using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier based on the unified node representations corresponding to the source hypergraph and the target hypergraph, respectively. The classification module is used to classify and predict the nodes of the target hypergraph based on the node classifier on the source hypergraph, and obtain the node classification result of the target hypergraph.

[0016] Thirdly, this application also provides a computer electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the node classification method for a hypergraph as described in any of the above.

[0017] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the node classification method for a hypergraph as described in any of the preceding claims.

[0018] This application provides a node classification method, system, device, and medium for hypergraphs, the advantages of which are: 1. Breaking through the bottleneck of cross-hypergraph classification technology and filling the gap in the field: In view of the problem that existing technologies are difficult to adapt to high-order relationships in hypergraphs and cannot effectively handle non-independent and identically distributed data in cross-hypergraphs, this application has specially designed a local and high-order consistency coding and adaptation mechanism, which for the first time realizes efficient node classification in cross-hypergraph scenarios, provides a reliable classification solution for target hypergraphs with missing labels, and fills the research gap in this technical field.

[0019] 2. Enhancing the discriminative and transferable nature of node feature representations: A local consistency aggregator captures direct node associations, while a higher-order consistency generator based on clique expansion and random walks mines deep indirect relationships. An attention mechanism then intelligently fuses these two types of features, ensuring that node representations simultaneously contain both local and higher-order association information, resulting in stronger class discrimination capabilities. Furthermore, adversarial domain adaptation and source-prototype contrastive learning achieve joint alignment of edge and conditional distributions, significantly improving the cross-hypergraph transferability of features and ensuring the classifier's adaptability to the target hypergraph.

[0020] 3. Optimize classification performance and reduce interference from anomalous nodes: This invention optimizes target node features using source prototype comparison learning and the core features of the source hypergraph as templates. This reduces the computational cost of pairwise node comparisons and effectively minimizes the impact of individual anomalous nodes in the source hypergraph on the classification results, making the classifier more robust. Experiments on multiple real-world datasets, such as citation networks and social networks, show that the classification accuracy of this method significantly outperforms mainstream baseline models such as HGNN, GCN, and GAT, demonstrating outstanding performance.

[0021] 4. Balancing model training efficiency and stability: The local consistency aggregator uses symmetric normalized hypergraph convolution operations to avoid gradient vanishing and numerical instability caused by multiple convolutions; the overall objective function coordinates classification loss and distribution alignment loss by balancing parameters, making the model training process more stable and convergent faster, thus improving training efficiency while ensuring classification performance. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a node classification method across a hypergraph according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a node classification system across a hypergraph according to an embodiment of this application; Figure 3This is a schematic diagram of the structure of a computer electronic device according to an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Please see Figure 1 The node classification method across a hypergraph provided in this application includes at least the following steps: S10. Construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes.

[0026] Specifically, in this embodiment, two hypergraphs are constructed: a source hypergraph containing labeled nodes and a target hypergraph containing unlabeled nodes. The source and target hypergraphs have the same dimension for their node attributes, denoted as . d Furthermore, the source hypergraph and the target hypergraph share the same label space, where the number of categories is denoted as . C .

[0027] Set subscript " s "and" t ", " represent the source hypergraph and the target hypergraph, respectively. The constructed source hypergraph takes the form of: ,in and These represent the source hypergraph nodes and the hyperedge set, respectively. and Let be the number of nodes and the number of hyperedges in the source hypergraph, respectively, and each hyperedge is a non-empty subset of the nodes. The incidence matrix of the source hypergraph, if ,but ;otherwise . It is a diagonal matrix that stores the positive weights of the hyperedges. It is the feature representation of the source hypergraph node, and its label is a one-hot matrix. This means that if the node Category Mark, then ;otherwise .

[0028] Similarly, and These represent the number of nodes and hyperedges in the target hypergraph, respectively. The target hypergraph is represented as... ,in and These are the sets of target nodes and hyperedges, respectively. This is the incidence matrix of the target hypergraph. The positive weights of the target hyperedges are stored in a diagonal matrix. middle. This represents the feature representation of the target node. It's worth noting that if there is no information about the importance of hyperedges in the source and target hypergraphs, it can be simply... and The diagonal elements are set to 1. Overall, the goal of cross-hypergraph node classification is to utilize... and The information in the dataset is used to train a classifier that performs well when classifying unlabeled target nodes.

[0029] S20. Apply a preset local and higher-order consistency encoding and adaptation model to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency aggregator, a local consistency aggregator and an attention mechanism module. The higher-order consistency aggregator includes: a higher-order consistency generator.

[0030] In one embodiment of this application, step S2O includes: S201. For the source hypergraph and the target hypergraph, respectively, a second source hypergraph and a second target hypergraph are constructed using the higher-order consistency generator.

[0031] Specifically, the higher-order consistency generator is a method for studying higher-order consistency relationships in hypergraphs based on clique expansion and random walk strategies. Its main idea is to measure the higher-order approximation between nodes and hyperedges in a hypergraph network. First, given a hypergraph containing n nodes and m hyperedges with an incidence matrix of... The hypergraph is converted into a standard graph using clique expansion. This establishes relationships between nodes and obtains pairwise connections between them. Let To indicate The adjacency matrix of node relationships. The nodes are calculated as follows: i and j The degree of correlation between them:

[0032] in, It is an adjustable parameter. It is a connection node i and j The number of hyperedges. Therefore, the more hyperedges connecting two nodes, the stronger their association.

[0033] Subsequently, based on the adjacency matrix A random walk strategy is used to calculate the transition probability matrix. ,in Indicates passage Step from node Jump to node Possibility: ; Where A is of Norm.

[0034] After obtaining the transition probability matrix Then, calculate the point mutual information matrix (denoted as ). Its formula can be expressed as: ; in Represents a node i and j The correlation between them (i.e., higher-order consistency) is as follows:

[0035]

[0036]

[0037] in, Indicates from node go through Step random walk to node The transition probability is the proportion of the total transition probabilities of all node pairs. Indicates from node Departure, Passing The sum of probabilities of reaching any node in one step is the proportion of the total sum of all transition probabilities. This indicates starting from any node and passing through... Step to Node The sum of probabilities of transitions is the proportion of the total sum of all transition probabilities.

[0038] Finally, based on the mutual information matrix between points G It can be done through aggregation nodes i With super-edge j The higher-order proximity between a node and its hyperedge is measured by the topological proximity between each node. Specifically, this is achieved by storing the higher-order proximity in a matrix. From this, we obtain the formula: ; in, Represents a node With super-edge Higher-order proximity; Represents a node and The correlation between them; Represents a node With super-edge The relationship, if belong Then there is ,otherwise .

[0039] Similarly, the same operation is applied to the target hypergraph, thereby forming a second target hypergraph network.

[0040] S202. For the source hypergraph and the target hypergraph, the local consistency aggregator is used to aggregate the features of each node in the local hyperedge neighborhood to generate the local consistency node representation of the source hypergraph and the local consistency node representation of the target hypergraph.

[0041] Specifically, for each node in the hypergraph, local consistency relationships are captured using direct neighbors connected by hyperedges. More specifically, each hypergraph network is encoded using a local consistency aggregator. Given a source hypergraph network... The local consistency aggregator performs convolution operations on nodes. To avoid gradient vanishing and numerical instability caused by stacking multiple hypergraph convolutional layers, symmetric normalization is used, resulting in the following formula: ; in, For activation function, It is the first The trainable parameters of the layer, They are the first Layer and first The node representation of the layer, and ; It is a diagonal matrix used to store node degrees, and its elements are... ,and It is a hyperdiagonal matrix, whose elements .

[0042] Consider using a hypergraph convolution with two layers in the local consistency aggregator, denoted as . To learn the local consistency node representation of the potential source hypergraph, denoted as , can be represented as: ; Similarly, for target hypergraph networks Applying the normalized hypergraph convolution operation to the target node, its formula can be expressed as: ; in , It is a diagonal matrix of node degree. , and have It is a hyperdiagonal matrix whose elements satisfy... Therefore, by constructing a two-layer local consistency aggregator, the locally consistent node representation of the target hypergraph (denoted as ) is obtained. It can be calculated as follows: ; S203. For the second source hypergraph and the second target hypergraph, respectively, the higher-order consistency aggregator is used to aggregate the higher-order related information of each node in the global structural context, and generate the higher-order consistent node representation of the source hypergraph and the higher-order consistent node representation of the target hypergraph.

[0043] Specifically, firstly, for a given source hypergraph... With target hypergraph The matrices are calculated using the formulas in step S201. and To study their higher-order proximity (i.e., higher-order consistency relations).

[0044] Then, this application designs a high-order consistency aggregator to encode each hypergraph. Specifically, for the high-order consistency node representation information in the source hypergraph network, a normalized hypergraph convolution operation is applied, as shown below: ; in, It is the first The trainable parameters of the layer, For a diagonal matrix representing node degree, its elements and This is a hyperdiagonal matrix with elements .

[0045] Similarly, the normalized hypergraph convolution operation used to encode the representation information of higher-order consistent nodes in the target hypergraph network can be described as follows: ; in , It is a diagonal degree matrix that satisfies ,and It is a diagonal hyperedge degree matrix, whose elements .

[0046] set up It is a two-level higher-order consistent aggregator. It will be... The learned latent representations of the source node and the target node are denoted as follows: and Then we have: ; ; Among them, the trainable parameters of the higher-order consistency aggregator on the source hypergraph and the target hypergraph are... Set to the same.

[0047] S204. Using the attention mechanism module, the local consistency nodes of the source hypergraph and the higher-order consistency node representations of the second source hypergraph are weighted and fused to obtain the unified node representation corresponding to the source hypergraph; and the local consistency nodes of the target hypergraph and the higher-order consistency node representations of the second target hypergraph are weighted and fused to obtain the unified node representation corresponding to the target hypergraph.

[0048] Specifically, by utilizing the relationship between the local consistency node representations and higher-order consistency node representations in the source and target hypergraphs, two different node representations of the source hypergraph network were obtained, namely... and And two different node representations of the target hypergraph network, namely and Subsequently, considering the different contributions of local and higher-order consistency information, an attention mechanism is used to encode them. Let... Indicates origin from source hypergraph network or target hypergraph network The elements are then processed using a linear transformation layer. Calculate the weight coefficients for the local view and the higher-order view for each node separately. and The details are as follows: ; in, and They are and The vector, and express and The splicing. Then, the softmax function is further applied to... Normalization is performed as follows: ; Therefore, for those from (or For a node, the coefficients of its local view and higher-order view are obtained respectively. and (or and By calculating the weight coefficients of each node in both hypergraphs and filling them into the diagonal matrix (denoted as ) as the values ​​of the diagonal elements, we can achieve this. , , and The local and higher-order consistency information is encoded to obtain a unified node representation for the source and target hypergraph networks, denoted as […]. and The details are as follows: ; in, Dimensions representing latent features.

[0049] S30. Based on the unified node representations corresponding to the source hypergraph and the target hypergraph, a node classifier on the source hypergraph is trained using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier.

[0050] In one embodiment of this application, the edge distribution alignment training method includes: constructing a domain discriminator, and forcing the encoder to generate domain-confusing embeddings through gradient inversion training, so that the node classifier cannot rely on domain-specific features; the conditional distribution alignment training method includes: constructing a source prototype based on the source domain, and guiding target domain nodes to move closer to their corresponding source prototypes, thereby explicitly transferring the semantic decision structure of the node classifier to the target domain. Specifically, the edge distribution alignment training method in the steps aims to align the edge distributions between the source hypergraph network and the target hypergraph network. This is achieved by constructing a domain discriminator to force the alignment. and Make them as similar as possible.

[0051] Specifically, an adversarial training scheme is employed to enable the discriminator to determine whether input features originate from the source hypergraph network or the target hypergraph network. Its main objective is to... and They are similar enough that the domain discriminator cannot distinguish them, allowing the source hypergraph node classifier to be used to predict the label of the target node. Then, a gradient reversal layer (GRL) is introduced to implement adversarial learning. In the forward propagation phase, the GRL acts as an identity mapping, directly passing the input data to the next layer; while in the backpropagation phase, the GRL reverses the gradient direction, adjusting parameters to maximize the discriminator's output. The domain adversarial loss is defined as: ; in, This represents the number of nodes in the source hypergraph; Indicates the number of nodes in the target hypergraph; Indicates the first i Is each node from the source hypergraph or the target hypergraph? Indicates the first i The domain prediction results for each node.

[0052] It should be noted that simply aligning edge distributions may not be sufficient to achieve cross-domain invariant feature representation learning. Although some cross-network models use conditional MMD to align conditional distributions, they fail to fully utilize label information. Therefore, in this embodiment, a training method based on conditional distribution alignment is also adopted to further enhance the transferability of node representations learned in the source and target hypergraphs and improve feature discrimination capabilities.

[0053] Specifically, in this embodiment, a source-prototype contrast function is designed to measure the similarity of the learned representations. Here, the source prototype is considered as the center point of different categories in the source hypergraph. The comparison between the source prototype and the target node is performed for two reasons: first, it is computationally less costly than comparing each source hypergraph node with the target node; second, it reduces the influence of individual outlier nodes in the source hypergraph (i.e., nodes whose features are significantly different from most nodes). To address the issue of unlabeled target nodes, a source classifier is used... Pseudo-tags are generated for it. Therefore, the target pseudo-tags will be gradually refined.

[0054] In the source hypergraph, based on Calculated Defined as labeled as category The center of the node. Let Indicates pseudo-tags as The The representation of each target node, and let (in ) indicates that pseudo-tags are different The The representation of each target node. Here, and All from The vector. Next, ( ) are considered positive sample pairs, and ( These are considered negative sample pairs. This is achieved by using a cosine similarity function. As a scoring metric, the source-prototype comparison function is defined as follows: ; in, Indicates the number of categories, This indicates the number of target nodes marked as c; This represents the number of target nodes marked as k; It is expressed as a temperature parameter.

[0055] S40. Based on the node classifier on the source hypergraph, classify and predict the nodes of the target hypergraph to obtain the node classification result of the target hypergraph.

[0056] Specifically, in step S30, the nodes on the labeled source hypergraph are obtained, and a node classifier is trained, the formula of which is expressed as follows: ; in, It is the softmax function. This represents the prediction result for the source hypergraph nodes. and These are trainable parameters. The cross-entropy function is used to calculate the classification loss, and the specific formula is as follows: ; in, Represents a node Category The predicted probability.

[0057] In this embodiment, given a target hypergraph The model is used to feed the data to obtain the target node representation. It is used to make the following predictions: ; in, Let represent the prediction probability matrix. Then, the th... The final prediction result for each node is determined by selecting the category with the highest probability, as follows: ; In one embodiment of this application, the method further includes: S50. The accuracy rate is used as the evaluation index to evaluate the node classification results.

[0058] This can be understood as follows: for node classification, accuracy is the proportion of correctly predicted results out of all predicted samples. In hyperedge classification, accuracy refers to the proportion of correctly predicted labels out of the total number of labels for that sample. Accuracy refers to the proportion of correctly classified nodes in the target hypergraph out of the total number of nodes in the target hypergraph, reflecting the model's ability to classify unlabeled target nodes. Let... Indicates the first The accuracy metric is calculated based on the true labels of each target node as follows: ; in, It is an indicator function.

[0059] It should be noted that the node classification method proposed in this application includes three loss terms: cross-entropy loss based on labeled source hypergraph nodes, domain adversarial loss to reduce edge distribution differences, and source-prototype contrast loss to minimize conditional distribution differences, denoted as follows: , and Its overall objective function is expressed as: ; in, and These are the equilibrium parameters of the model.

[0060] This application provides a node classification method for hypergraphs, the advantages of which are: 1. Breaking through the bottleneck of cross-hypergraph classification technology and filling the gap in the field: In view of the problem that existing technologies are difficult to adapt to high-order relationships in hypergraphs and cannot effectively handle non-independent and identically distributed data in cross-hypergraphs, this application has specially designed a local and high-order consistency coding and adaptation mechanism, which for the first time realizes efficient node classification in cross-hypergraph scenarios, provides a reliable classification solution for target hypergraphs with missing labels, and fills the research gap in this technical field.

[0061] 2. Enhancing the discriminative and transferable nature of node feature representations: A local consistency aggregator captures direct node associations, while a higher-order consistency generator based on clique expansion and random walks mines deep indirect relationships. An attention mechanism then intelligently fuses these two types of features, ensuring that node representations simultaneously contain both local and higher-order association information, resulting in stronger class discrimination capabilities. Furthermore, adversarial domain adaptation and source-prototype contrastive learning achieve joint alignment of edge and conditional distributions, significantly improving the cross-hypergraph transferability of features and ensuring the classifier's adaptability to the target hypergraph.

[0062] 3. Optimize classification performance and reduce interference from anomalous nodes: This invention optimizes target node features using source prototype comparison learning and the core features of the source hypergraph as templates. This reduces the computational cost of pairwise node comparisons and effectively minimizes the impact of individual anomalous nodes in the source hypergraph on the classification results, making the classifier more robust. Experiments on multiple real-world datasets, such as citation networks and social networks, show that the classification accuracy of this method significantly outperforms mainstream baseline models such as HGNN, GCN, and GAT, demonstrating outstanding performance.

[0063] 4. Balancing model training efficiency and stability: The local consistency aggregator uses symmetric normalized hypergraph convolution operations to avoid gradient vanishing and numerical instability caused by multiple convolutions; the overall objective function coordinates classification loss and distribution alignment loss by balancing parameters, making the model training process more stable and convergent faster, thus improving training efficiency while ensuring classification performance.

[0064] Please see Figure 2 This application also provides a node classification system 200 across a hypergraph, comprising: Construction module 201 is used to construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; The feature extraction module 202 is used to apply a preset local and higher-order consistency encoding and adaptation model to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency generator, a local consistency aggregator, a higher-order consistency aggregator and an attention mechanism module. The training module 203 is used to train a node classifier on the source hypergraph by using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier based on the unified node representations corresponding to the source hypergraph and the target hypergraph, respectively. The classification module 204 is used to perform classification prediction on the nodes of the target hypergraph based on the node classifier on the source hypergraph, and obtain the node classification result of the target hypergraph.

[0065] Please see Figure 3 This application also provides a computer electronic device 300, including a memory 303 and a processor 302. The memory 303 stores a computer program, and the processor 302 executes the computer program to implement the steps of the node classification method for the hypergraph described above.

[0066] Specifically, the computer electronic device 300 includes: a transceiver 301, a bus interface, and a processor 302. The processor 302 is used to construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; a preset local and higher-order consistency coding and adaptation model is applied to the source hypergraph and the target hypergraph respectively to generate unified node representations corresponding to the source hypergraph and the target hypergraph, wherein the local and higher-order consistency coding and adaptation model includes: a higher-order consistency aggregator, a local consistency aggregator, and an attention mechanism module, and the higher-order consistency aggregator includes: a higher-order consistency generator; based on the unified node representations corresponding to the source hypergraph and the target hypergraph respectively, a preset node classifier is trained using edge distribution alignment and conditional distribution alignment training methods to obtain a node classifier on the source hypergraph; based on the node classifier on the source hypergraph, the nodes of the target hypergraph are classified and predicted to obtain the node classification result of the target hypergraph.

[0067] In this embodiment of the application, the computer electronic device 300 further includes a memory 303. Figure 3 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors 302 (represented by processor 302) and various circuits of memory 303 (represented by memory 303). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 301 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. The processor 302 is responsible for managing the bus architecture and general processing, and the memory 303 can store data used by the processor 302 during operation.

[0068] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the node classification method for the hypergraph described above.

[0069] In this embodiment, the computer-readable storage medium can be a non-volatile storage medium or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0070] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0071] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0072] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0073] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0074] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a terminal device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A node classification method for a hypergraph, characterized in that, include: Construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; A preset local and higher-order consistency encoding and adaptation model is applied to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency aggregator, a local consistency aggregator and an attention mechanism module. The higher-order consistency aggregator includes: a higher-order consistency generator. Based on the unified node representations corresponding to the source hypergraph and the target hypergraph, a node classifier on the source hypergraph is trained by using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier. Based on the node classifier on the source hypergraph, the nodes of the target hypergraph are classified and predicted to obtain the node classification result of the target hypergraph.

2. The node classification method according to claim 1, characterized in that, The step of applying a preset local and higher-order consistency coding and adaptation model to the source hypergraph and the target hypergraph respectively to generate unified node representations corresponding to the source hypergraph and the target hypergraph includes: The source hypergraph and the target hypergraph are respectively constructed using the higher-order consistency generator to form a second source hypergraph and a second target hypergraph; For the source hypergraph and the target hypergraph, the local consistency aggregator is used to aggregate the features of each node in the local hyperedge neighborhood to generate the local consistency node representation of the source hypergraph and the local consistency node representation of the target hypergraph. For the second source hypergraph and the second target hypergraph, the higher-order consistency aggregator is used to aggregate the higher-order related information of each node in the global structural context, generating a higher-order consistent node representation of the second source hypergraph and a higher-order consistent node representation of the second target hypergraph. Using the attention mechanism module, the local consistency nodes of the source hypergraph and the higher-order consistency node representations of the second source hypergraph are weighted and fused to obtain the unified node representation corresponding to the source hypergraph; and the local consistency nodes of the target hypergraph and the higher-order consistency node representations of the second target hypergraph are weighted and fused to obtain the unified node representation corresponding to the target hypergraph.

3. The node classification method according to claim 1, characterized in that, The training method for edge distribution alignment includes: constructing a domain discriminator and forcing the encoder to generate domain-confusing embeddings through gradient inversion training, so that the node classifier cannot rely on domain-specific features; The training method for conditional distribution alignment includes: constructing a source prototype based on the source domain and guiding the target domain nodes to approach their corresponding source prototypes, thereby explicitly transferring the semantic decision structure of the node classifier to the target domain.

4. The node classification method according to claim 3, characterized in that, The training method for edge distribution alignment uses the following loss function: ； in, This represents the number of nodes in the source hypergraph; Indicates the number of nodes in the target hypergraph; Indicates the first i Is each node from the source hypergraph or the target hypergraph? Indicates the first i The domain prediction results for each node.

5. The node classification method according to claim 3, characterized in that, The training method for conditional distribution alignment uses the following loss function: ； in, Indicates the number of categories, This indicates the number of target nodes marked as c; This represents the number of target nodes marked as k; It is expressed as a temperature parameter.

6. The node classification method according to claim 1, characterized in that, The node classification prediction of the target hypergraph is performed based on the node classifier on the source hypergraph to obtain the node classification result of the target hypergraph, using the following loss function: ； ； in, Represents a node Category The predicted probability; This represents the softmax function. This represents the prediction result for the source hypergraph nodes. and This represents the trainable parameters.

7. The node classification method according to claim 1, characterized in that, The method further includes: Accuracy is used as the evaluation metric to evaluate the node classification results.

8. A node classification system for a cross-hypergraph, characterized in that, include: A construction module is used to construct a source hypergraph and a target hypergraph, wherein the source hypergraph contains labeled nodes and the target hypergraph contains unlabeled nodes; The feature extraction module is used to apply a preset local and higher-order consistency encoding and adaptation model to the source hypergraph and the target hypergraph respectively to generate a unified node representation corresponding to the source hypergraph and the target hypergraph. The local and higher-order consistency encoding and adaptation model includes: a higher-order consistency aggregator, a local consistency aggregator and an attention mechanism module. The higher-order consistency aggregator includes: a higher-order consistency generator. The training module is used to train a node classifier on the source hypergraph by using edge distribution alignment and conditional distribution alignment training methods on a preset node classifier based on the unified node representations corresponding to the source hypergraph and the target hypergraph, respectively. The classification module is used to classify and predict the nodes of the target hypergraph based on the node classifier on the source hypergraph, and obtain the node classification result of the target hypergraph.

9. A computer electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the node classification method for a hypergraph according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the node classification method for a hypergraph according to any one of claims 1-7.

Images