A prediction method combining multi-relation hypergraph modeling and contrastive learning

Abstract

This invention relates to the field of artificial intelligence technology, and particularly to a prediction method combining multi-relation hypergraph modeling and contrastive learning. The method includes: constructing a hypersubgraph containing explicit and implicit regulatory relationships from a multi-relation perspective; the construction of the hypersubgraph can be data-driven, knowledge-driven, or rule-driven; performing feature learning on the constructed hypersubgraph using hypergraph convolution operators to extract high-order feature representations of nodes in different hypersubgraphs; enhancing the consistency and discriminative power of node feature representations through contrastive learning to optimize feature representations across subgraphs; and fusing node feature representations from different hypersubgraphs through an attention mechanism to generate a globally optimized node descriptor. This invention not only improves prediction accuracy but also provides important theoretical support for mining potential cancer biomarkers.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a prediction method that combines multi-relation hypergraph modeling and contrastive learning. Background Technology

[0002] Currently, methods for predicting cancer biomarkers mainly include various techniques based on statistical analysis, machine learning, and deep learning. Statistical analysis methods typically rely on calculating simple correlations between molecules, which is efficient but struggles to capture complex nonlinear relationships. While traditional machine learning methods can effectively process some data through feature engineering, they heavily rely on manual feature selection and are ill-suited to high-dimensional, sparse, and heterogeneous data. In recent years, graph neural networks have been widely used in bioinformatics, providing strong support for cancer biomarker identification due to their advantage in capturing complex relationships and local structural information in molecular networks. However, traditional graph models often only represent binary relationships and cannot accurately express higher-order multi-molecule regulatory relationships in ncRNA networks. Summary of the Invention

[0003] The purpose of this invention is to provide a prediction method that combines multi-relationship hypergraph modeling and contrastive learning, aiming to solve the problem that traditional graph models can often only represent binary relationships and cannot accurately express the high-order regulatory relationships between multiple molecules in ncRNA networks.

[0004] This invention is implemented as follows: a prediction method combining multi-relation hypergraph modeling and contrastive learning, the method comprising:

[0005] A super subgraph containing explicit and implicit regulatory relationships is constructed from a multi-relationship perspective. The construction methods of the super subgraph include data-driven, knowledge-driven, and rule-driven approaches.

[0006] By using hypergraph convolution operators to learn features from the constructed hypergraphs, high-order feature representations of nodes in different hypergraphs are extracted.

[0007] By contrastive learning, the consistency and discriminative power of node feature representations are enhanced, and feature representations across subgraphs are optimized.

[0008] By fusing node feature representations from different super subgraphs through an attention mechanism, a globally optimized node descriptor is generated.

[0009] Preferably, when using a data-driven approach, a node embedding method based on weighted random walks is used to generate nearest neighbor hyperedges. For any node, its neighbor sequence is generated using the random walk formula:

[0010]

[0011] Where A is the adjacency matrix, d(vj ) represents the node degree, N(v) j ) represents the set of neighbors of a node.

[0012] Preferably, when using a knowledge-driven approach, hyperedges are generated based on biological facts and experimental data using the k-nearest neighbor algorithm. Its mathematical definition is:

[0013]

[0014] Preferably, when using a rule-driven approach, implicit hyperedges are constructed using signal path characteristics, and multi-scale features are extracted through wavelet transform:

[0015]

[0016] Where h is the Haar wavelet basis and r is the residual.

[0017] Preferably, in the supergraph, the Laplacian matrix is ​​used to model the association strength between nodes. By calculating the weight matrix, node degree matrix, and hyperedge degree matrix of the hyperedge, the Laplacian matrix of the hypergraph is generated. Using the hypergraph convolution operator, node features are propagated from one hyperedge to other connected hyperedges. Through this feature aggregation operation, and through multi-layer hypergraph convolution operation, high-order features are extracted.

[0018] Preferably, in the step of enhancing the consistency and discriminative ability of node feature representations through contrastive learning and optimizing feature representations across subgraphs, for each node, positive samples and multiple negative samples are generated. The positive samples come from the representations of the same node in different super subgraphs. By defining a contrastive loss function, the feature similarity between positive samples is maximized.

[0019] Preferably, in the step of generating a globally optimized node descriptor by fusing node feature representations from different super subgraphs through an attention mechanism, for each node, its attention weight in different subgraphs is calculated. The attention weight is dynamically allocated based on the feature importance of the node and the correlation strength of the superedge. The feature representation of the node in each subgraph is multiplied by the corresponding attention weight to generate a weighted feature representation. Through feature fusion operation, all weighted features are integrated into a global descriptor.

[0020] This invention provides a prediction method combining multi-relationship hypergraph modeling and contrastive learning. By integrating hypergraph modeling, contrastive learning, and attention mechanism feature fusion techniques, it deeply analyzes the regulatory mechanisms and multi-relationship network characteristics of cancer-related ncRNA biomarkers, thereby achieving high-precision biomarker prediction. Through a subgraph construction method with a multi-relationship background, this invention can comprehensively capture explicit and implicit relationships, revealing high-order dependencies and long-distance regulatory patterns in the ceRNA regulatory network, significantly improving the modeling ability of complex biological networks and the depth of understanding of key nodes. This refined modeling method based on multi-relationships not only improves prediction accuracy but also provides important theoretical support for the discovery of potential cancer biomarkers. Attached Figure Description

[0021] Figure 1 A framework diagram of the prediction method combining multi-relation hypergraph modeling and contrastive learning provided in an embodiment of the present invention;

[0022] Figure 2 The flowchart illustrates a prediction method combining multi-relation hypergraph modeling and contrastive learning, as provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] like Figure 1 and Figure 2 As shown, the method provided in the embodiments of the present invention includes:

[0025] A super subgraph containing explicit and implicit regulatory relationships is constructed from a multi-relationship perspective. The construction methods of the super subgraph include data-driven, knowledge-driven, and rule-driven approaches.

[0026] By using hypergraph convolution operators to learn features from the constructed hypergraphs, high-order feature representations of nodes in different hypergraphs are extracted.

[0027] By contrastive learning, the consistency and discriminative power of node feature representations are enhanced, and feature representations across subgraphs are optimized.

[0028] By fusing node feature representations from different super subgraphs through an attention mechanism, a globally optimized node descriptor is generated.

[0029] While existing cancer biomarker prediction technologies have made significant progress, they still have some obvious shortcomings, mainly including the following: high dependence of models on data, insufficient ability to process multimodal features, difficulty in capturing complex relationships in high-order biological networks, and poor robustness to noisy data.

[0030] Traditional biomarker prediction methods primarily rely on statistical analysis and machine learning techniques, which often require manual feature selection to build predictive models. However, manual feature selection may miss potentially crucial features, making it difficult to comprehensively capture the complex regulatory relationships within biological networks. Furthermore, with the deepening research into cancer molecular mechanisms, biomarker data exhibits high dimensionality and multimodal characteristics, with various data types such as transcriptomics, epigenetics, and proteomics intertwined, proving inadequate for traditional methods in integrating multimodal features. While deep learning techniques can improve predictive performance in certain scenarios by automatically extracting features, their high dependence on large-scale, high-quality data is a significant limitation in actual biomedical research, where high-quality labeled data is often scarce, severely restricting the performance of deep learning models.

[0031] Meanwhile, the regulatory networks of cancer biomarkers are complex and nonlinear. Existing methods, mostly based on traditional graphical models or simple matrix operations, struggle to fully express and capture high-order dependencies between biomolecules. This is particularly true in competitive endogenous RNA regulatory networks (ceRNAs), where the multi-molecular regulatory relationships among circRNAs, lncRNAs, miRNAs, and mRNAs are complex, and simple binary relationships cannot meet modeling requirements. Furthermore, biomedical data is often accompanied by significant noise, and existing techniques perform poorly in handling this noise interference, easily leading to biased predictions and reducing the reliability of biomarker screening.

[0032] To address the aforementioned technical shortcomings, this invention proposes a multimodal biomarker prediction method based on hypergraph modeling and contrastive learning, effectively overcoming the deficiencies of existing methods. This method constructs targeted hyperedges through three structural relationships and employs a hypergraph model to model higher-order regulatory relationships in the ceRNA network, achieving comprehensive and accurate capture of complex intermolecular relationships. Combined with contrastive learning, it can efficiently process multimodal data, automatically extract alignment key features, and improve the model's robustness to noisy data through end-to-end training. Furthermore, this invention integrates multiple topological structure information through multi-view fusion, effectively alleviating the problem of insufficient representation capacity of single data structures and improving the comprehensiveness and accuracy of biomarker prediction. This method provides a novel technical framework for cancer biomarker prediction and promotes the application of hypergraph modeling and contrastive learning techniques in the field of bioinformatics, demonstrating significant innovation, practicality, and application prospects.

[0033] This invention proposes an innovative learning method based on multi-relation hypergraph comparison, with the system architecture as follows: Figure 1 As shown, this method is specifically designed for high-precision prediction of cancer-related ncRNA biomarkers. It constructs subgraphs by integrating multi-relational contexts of data, knowledge, and rules, and combines hypergraph Laplacian spectrum theory with contrastive learning techniques to achieve comprehensive modeling of the ceRNA regulatory network. Through a multi-relational enhanced attention mechanism, it adaptively optimizes and fuses spatial features from different subgraphs to generate high-dimensional, interpretable node feature descriptors, ultimately significantly improving the understanding of complex regulatory networks and the predictive performance of cancer biomarkers.

[0034] The present invention mainly comprises four modules: a subgraph construction module for multi-relational backgrounds, a hypergraph Laplacian spectrum representation learning module, a multi-relational hypergraph contrast learning module, and a multi-relational enhancement module. The detailed technical solution of the present invention is as follows:

[0035] (1) Subgraph construction module with multiple relationship background

[0036] The subgraph construction module with a multi-relationship background aims to construct supergraphs containing explicit and implicit regulatory relationships from a multi-relationship perspective, revealing higher-order regulatory patterns and potential connections between molecules.

[0037] First, molecular nodes and their relationships are extracted from the ceRNA regulatory network. Nodes can be non-coding RNAs (such as miRNAs and circRNAs) or mRNAs, while edges represent the relationships between these molecules (such as co-expression relationships and functional similarities).

[0038] Initial relational and feature matrices are generated using molecular omics data (such as transcriptome data, sequence similarity, and functional annotation). Direct explicit relational subgraphs are then constructed based on biological facts. For example, the regulatory relationship between miRNAs and mRNAs can be directly modeled as explicit edges using existing experimental data.

[0039] For the association between circRNA and mRNA, explicit relationships are defined using functional similarity or correlation of expression patterns. These explicit relationships form the foundational edges of the hypergraph, reflecting the direct biological significance between molecules.

[0040] Implicit relationships are an important complement to subgraphs, generated based on statistical inference and rule mining. First, clustering algorithms (such as spectral clustering or density clustering) are used to mine potential molecular grouping relationships from the network. Second, path analysis (such as KEGG pathway analysis) is combined to extract possible indirect regulatory relationships between molecules. For example, if two molecules share similar functional annotations or participate in the same regulatory pathway, implicit edges can be generated to reflect this potential association. Based on explicit and implicit relationships, hyperedges are used to connect multiple nodes to form a subgraph. For example, a hyperedge can include multiple circRNA and miRNA nodes, which may collectively act on a specific mRNA target, thus forming a biologically significant regulatory unit. In this way, subgraphs can not only capture one-to-one molecular relationships but also represent complex associations between multiple molecules.

[0041] Subgraph construction methods include three approaches: data-driven, knowledge-driven, and rule-driven. The construction principles for each approach are as follows:

[0042] Data-driven:

[0043] A node embedding technique based on weighted random walks is used to generate nearest neighbor hyperedges. For any node, its neighbor sequence is generated using the following random walk formula:

[0044]

[0045] Where A is the adjacency matrix, d(v j ) represents the node degree, N(v) j ) represents the set of neighbors of a node.

[0046] Knowledge-driven:

[0047] Based on biological facts and experimental data, hyperedges are generated using the k-nearest neighbor algorithm. Its mathematical definition is:

[0048]

[0049] Rule-driven:

[0050] Implicit hyperedges are constructed using signal path characteristics, and multi-scale features are extracted using wavelet transform.

[0051]

[0052] Where h is the Haar wavelet basis and r is the residual.

[0053] (2) Hypergraph Laplace Spectral Representation Learning Module

[0054] The hypergraph Laplacian spectral representation learning module performs feature learning on the constructed hypersubgraph using hypergraph convolution operators, extracting high-order feature representations of nodes in different subgraphs. The hypergraph consists of a set of nodes, a set of hyperedges, and a feature matrix. In each subgraph, the node feature matrix contains the initial attribute values ​​(such as expression values, functional annotation vectors, etc.) of each node. Furthermore, a vertex-hyperedge association matrix is ​​constructed to record the connection relationships between hyperedges and nodes. In the hypergraph, the Laplacian matrix is ​​used to model the association strength between nodes. The Laplacian matrix of the hypergraph is generated by calculating the weight matrix, node degree matrix, and hyperedge degree matrix of the hyperedges. The purpose of the Laplacian matrix is ​​to preserve the topological information between nodes during feature propagation. Using hypergraph convolution operators, node features are propagated from one hyperedge to other connected hyperedges. Through this feature aggregation operation, nodes can obtain information from neighboring nodes and update their feature representations layer by layer. The updated features not only contain the node's own information but also incorporate the regulatory features of neighboring nodes, reflecting the node's global position in the subgraph. High-order features are extracted progressively through multi-layer hypergraph convolution operations. The first layer focuses on the local neighborhood information of nodes; the second layer begins to capture long-range dependencies; and the third layer further integrates global topological information. Ultimately, the representation vector of each node contains both its own attributes and its global regulatory features within the subgraph. The hypergraph convolution operator is defined based on Laplace's spectral theory, and its propagation mechanism is as follows:

[0055]

[0056] (3) Multi-relation hypergraph comparison learning

[0057] The goal of contrastive learning in multi-relation hypergraphs is to enhance the consistency and discriminative power of node feature representations and optimize feature representations across subgraphs through contrastive learning. For each node, positive and negative samples are generated. Positive samples come from the representations of the same node in different hypergraphs; these representations should have high similarity since they originate from the same node. Negative samples select feature representations from other nodes, ensuring they are not directly related to the target node. A contrastive loss function is defined to maximize the feature similarity between positive samples while minimizing the feature similarity between positive and negative samples. The optimization objective of the loss function is to make the feature representations of the same node more consistent across different hypergraphs, while widening the feature differences between different nodes, thereby improving the discriminative power of node features. Based on contrastive learning, node features are aligned to make their representations as consistent as possible across different relation subgraphs. Through this alignment operation, nodes can obtain more comprehensive feature information across explicit and implicit relations. Contrastive learning optimizes the consistency of node representations. Its objective function is defined as follows:

[0058]

[0059] Where t is the temperature parameter and sim is the click similarity.

[0060] (4) Multi-relationship enhanced attention mechanism

[0061] The multi-relationship enhanced attention mechanism fuses node feature representations from different super-subgraphs using a dynamic attention mechanism to generate a globally optimized node descriptor. For each node, its attention weights in different subgraphs are calculated separately. These attention weights are dynamically allocated based on the node's feature importance and the correlation strength of its hyperedges, aiming to highlight the importance of key relationships while suppressing interference from irrelevant relationships. The node's feature representation in each subgraph is multiplied by its corresponding attention weight to generate a weighted feature representation. Then, through a feature fusion operation, all weighted features are integrated into a single global descriptor. The fused descriptor retains the feature differences between subgraphs while reflecting the global regulatory features of the nodes. A multi-head attention mechanism further optimizes the fused features, ensuring the accuracy and robustness of the feature fusion. The optimized node descriptor will be used for downstream prediction tasks, such as the prediction and validation of cancer biomarkers. The multi-relationship enhanced attention mechanism dynamically calculates the weight distribution of node features. Its core formula is:

[0062]

[0063] Where head represents the number of attention heads, and Q, K, and V represent the query, key, and value, respectively.

[0064] (5) Module Collaboration Process

[0065] like Figure 2 As shown, the four modules work together to complete the biomarker prediction task: starting with the subgraph construction module with a multi-relational background, the module extracts high-order features of nodes through the hypergraph Laplacian spectrum representation learning module, optimizes the feature representation of nodes by combining multi-relational hypergraph contrastive learning, and finally generates accurate and comprehensive node feature representations by using information from all subgraphs through a multi-relational enhanced attention mechanism. These node feature representations are input into the downstream classifier to achieve efficient prediction of cancer-related ncRNA biomarkers.

[0066] This invention, through systematic research and extensive experimental validation, confirms the crucial role of multi-relationship hypergraph contrastive learning in the prediction of cancer-related ncRNA biomarkers. This approach utilizes data-driven, knowledge-driven, and rule-driven multi-relationship modeling methods to construct multi-channel subgraphs with explicit and implicit relationships. By encapsulating complex regulatory relationships through hyperedges, it not only reveals high-order dependencies and long-distance regulatory patterns in ceRNA networks but also provides solid data support and biological evidence for a comprehensive understanding of diverse intermolecular regulatory mechanisms.

[0067] The proposed method based on hypergraph convolution and multi-relation contrastive learning demonstrates superior feature extraction and generalization performance when processing features of multi-relation subgraphs. This method leverages hypergraph convolution operators to enhance the high-dimensional feature representation of nodes, significantly improving their ability to capture explicit and implicit relationships. The multi-relation contrastive learning module further optimizes the feature consistency and discriminative ability of nodes, ensuring the model's robustness and adaptability across different feature spaces. Furthermore, the multi-head attention mechanism designed in this method dynamically fuses feature representations from different subgraphs, accurately capturing cross-relational collaborative information and improving the model's efficiency and predictive ability for high-dimensional and complex data.

[0068] This invention combines hypergraph modeling, contrastive learning, and attention mechanism feature fusion techniques to deeply analyze the regulatory mechanisms and multi-relationship network characteristics of cancer-related ncRNA biomarkers, thereby achieving high-precision biomarker prediction. Through a subgraph construction method with a multi-relationship background, this invention can comprehensively capture explicit and implicit relationships, revealing high-order dependencies and long-distance regulatory patterns in the ceRNA regulatory network, significantly improving the modeling ability of complex biological networks and the depth of understanding of key nodes. This refined modeling method based on multi-relationships not only improves prediction accuracy but also provides important theoretical support for the discovery of potential cancer biomarkers.

[0069] This invention also demonstrates significant advantages in computational efficiency and data processing capabilities. By introducing hypergraph convolution operators and multi-relational contrastive learning techniques, this invention can efficiently extract high-dimensional representations of nodes in different feature spaces, optimizing the robustness and consistency of feature learning. The dynamic feature fusion strategy combined with multi-head attention significantly improves the model's processing efficiency for high-dimensional and sparse data. This fast and efficient feature extraction and fusion capability ensures that marker prediction tasks can run stably on large-scale datasets, while also exhibiting stronger adaptability to data noise and structural complexity.

[0070] Overall, this invention demonstrates significant practicality and effectiveness in the field of cancer biomarker prediction. It not only addresses the bottlenecks of existing methods, such as insufficient modeling of multiple relationships and inadequate feature fusion, but also provides innovative technical support and practical application value for promoting the application of precision medicine in cancer research.

[0071] In the task of predicting cancer-related ncRNA biomarkers, the method proposed in this invention has been extensively validated experimentally on three cancer biomarker prediction tasks: miRNA, lncRNA, and circRNA. The experimental results are shown in Tables 1, 2, and 3. Existing deep learning models (such as GCN, GAT, and TransformerConv(TC)) were used as benchmarks in the experiments, and the models were tested and evaluated using real cancer ceRNA network datasets.

[0072] Table 1 Comparison of miRNA biomarker prediction results based on deep learning models

[0073]

[0074] Table 2 Comparison of LncRNA biomarker prediction results based on deep learning models

[0075]

[0076] Table 3 Comparison of circRNA biomarker prediction results based on deep learning models

[0077]

[0078] In the prediction task of cancer-related ncRNA biomarkers, the method proposed in this invention performs excellently in handling different types of cancer biomarker prediction tasks, especially in the prediction of miRNA, lncRNA, and circRNA biomarkers, significantly outperforming existing deep learning models. The comparative models used in the experiments include GCN, GAT, and TC, which are widely used in biological network analysis. Experimental results show that the method of this invention outperforms existing models in multiple metrics, including accuracy (ACC), F1 score (F1-score), AUC (area under the receiver operating function), and AUPR (area under the PR curve), validating its powerful feature modeling ability and predictive performance. In the miRNA biomarker prediction task, the F1-score of the method of this invention is 0.7403, significantly higher than GCN (0.7293) and TC (0.6967). In the lncRNA biomarker prediction task, the AUC of the method of this invention reaches 0.8516, while GCN is only 0.8214. For the circRNA biomarker prediction task, the accuracy of the method in this invention reached 0.7600, which is a significant improvement compared to GAT's 0.7215. These results demonstrate that the present invention has significant advantages in capturing complex intermolecular regulatory relationships and improving prediction accuracy, while also exhibiting good generalization ability and robustness.

[0079] Existing deep learning models have demonstrated their own characteristics and limitations in different tasks. The GCN model shows good performance in processing local structural information of biological networks, but its ability to model complex dependencies in multi-relational networks is insufficient; GAT enhances the representation of neighborhood information through attention mechanisms, but its performance depends on the quality of local neighborhoods and has a weak ability to capture implicit relationships; although TC introduces the advantages of the Transformer architecture, its performance in processing long-distance dependencies and multi-relational features in biological networks is limited by the data scale and complex relational structure.

[0080] The proposed method combines hypergraph modeling, multi-relation-driven approaches, and contrastive learning optimization techniques, demonstrating superior performance in handling high-order dependencies and multi-relation features. Through a multi-relation-driven hypergraph construction method, this invention comprehensively captures explicit and implicit relationships, encapsulating high-order regulatory patterns in ceRNA networks in hypergraph form. Subsequently, hypergraph convolution operators are used to enhance node feature representations, significantly improving the modeling ability of long-distance intermolecular dependencies. Furthermore, through contrastive learning optimization and multi-head attention mechanisms, this invention adaptively fuses feature information from different subgraphs to generate globally optimized node feature descriptors. Experimental results show that the proposed method outperforms existing models in three cancer biomarker prediction tasks.

[0081] In summary, the multi-relationship-driven, multi-channel hypergraph contrastive learning method proposed in this invention successfully integrates multimodal information from biological networks, fully leveraging the technical advantages of hypergraph modeling and deep learning. Compared to existing models, this invention not only achieves higher performance on various evaluation metrics but also significantly enhances the modeling capability for complex biological networks through dynamic feature fusion and contrastive learning optimization, providing a powerful tool for the prediction and analysis of cancer-related ncRNA biomarkers.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A prediction method combining multi-relation hypergraph modeling and contrastive learning, the method comprising: Construct a supergraph containing explicit and implicit regulatory relationships from a multi-relationship perspective; Molecular nodes and their relationships were extracted from the ceRNA regulatory network. Nodes were non-coding RNAs or mRNAs, with non-coding RNAs being miRNAs or circRNAs. Edges represented the relationships between these molecules, including co-expression relationships and functional similarities. Initial relational and feature matrices are generated using molecular omics data, including transcriptome data, sequence similarity, and functional annotations. Direct explicit relational subgraphs are constructed based on biological facts. Implicit relationships complement subgraphs, generated based on statistical inference and rule mining. Potential molecular grouping relationships are mined from the network using clustering algorithms, such as spectral clustering or density clustering. Path analysis is combined to extract possible indirect regulatory relationships between molecules, using KEGG pathway analysis. The construction methods of the super subgraph include data-driven, knowledge-driven, and rule-driven approaches; When using a data-driven approach; The node embedding method based on weighted random walk generates nearest neighbor hyperedges. For any node, its neighbor sequence is generated using the weighted random walk formula: Where A is the adjacency matrix, d(v j ) represents the node degree, N(v) j ) represents the set of neighbors of a node; When using a knowledge-driven approach, hyperedges are generated based on biological facts and experimental data using the k-nearest neighbor algorithm, and their mathematical definition is: ； When using a rule-driven approach, implicit hyperedges are constructed using signal path characteristics, and multi-scale features are extracted through wavelet transform. Where h is the Haar wavelet basis and r is the residual; Feature learning is performed on the constructed hypersubgraph using hypergraph convolution operators to extract higher-order feature representations of nodes in different hypersubgraphs. In the hypersubgraph, the Laplacian matrix is ​​used to model the association strength between nodes. By calculating the weight matrix, node degree matrix, and hyperedge degree matrix of the hyperedge, the Laplacian matrix of the hypergraph is generated. Using hypergraph convolution operators, node features are propagated from one hyperedge to other connected hyperedges. Through this feature aggregation operation, higher-order features are extracted through multi-layer hypergraph convolution operations. By contrastive learning, the consistency and discriminative power of node feature representations are enhanced, and feature representations across subgraphs are optimized. By fusing node feature representations from different super subgraphs through an attention mechanism, a globally optimized node descriptor is generated.

2. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that, In the step of enhancing the consistency and discriminative ability of node feature representations through contrastive learning and optimizing feature representations across subgraphs, positive and negative samples are generated for each node. Positive samples come from the representations of the same node in different super subgraphs. By defining a contrastive loss function, the feature similarity between positive samples is maximized.

3. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that, In the step of generating a globally optimized node descriptor by fusing node feature representations from different super subgraphs through an attention mechanism, for each node, its attention weight in different subgraphs is calculated. The attention weight is dynamically allocated based on the node's feature importance and the correlation strength of the hyperedge. The feature representation of the node in each subgraph is multiplied by the corresponding attention weight to generate a weighted feature representation. Through feature fusion, all weighted features are integrated into a global descriptor.

Images