Prediction method combining multi-relation hypergraph modeling and comparative learning

Abstract

The invention is suitable for the technical field of artificial intelligence, and particularly relates to a prediction method combining multi-relation hypergraph modeling and comparative learning, and the method comprises the steps: constructing a hypersubgraph containing an explicit and implicit regulation and control relation through a multi-relation view angle, the construction modes of the hypergraph comprise a data driving mode, a knowledge driving mode and a rule driving mode; performing feature learning on the constructed hyper-subgraphs through a hyper-graph convolution operator, and extracting high-order feature representations of nodes in different hyper-subgraphs; enhancing the consistency and distinguishing ability of node feature representation through comparative learning, and optimizing cross-subgraph feature representation; and fusing node feature representations of different hyper-graphs through an attention mechanism to generate a globally optimized node descriptor. The method provided by the invention not only improves the prediction accuracy, but also provides important theoretical support for mining potential cancer markers.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a prediction method combining multi-relation hypergraph modeling and contrastive learning. Background Art

[0002] At present, the methods for predicting cancer markers mainly include a variety of techniques based on statistical analysis, machine learning and deep learning. Statistical analysis methods usually rely on calculating simple correlations between molecules, which are highly efficient but difficult to capture complex nonlinear relationships; while traditional machine learning methods can effectively process some data through feature engineering, but they are highly dependent on artificial feature selection and are difficult to adapt to the characteristics of high-dimensional, sparse and heterogeneous data. In recent years, graph neural networks have been widely used in the field of bioinformatics. With their advantages in capturing complex relationships and local structural information in molecular networks, they have provided strong support for the identification of cancer markers. However, traditional graph models can often only represent binary relationships and cannot accurately express the regulatory relationships between high-order multi-molecules in ncRNA networks. Summary of the invention

[0003] The purpose of the present invention is to provide a prediction method combining 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 regulatory relationships between high-order multi-molecules in ncRNA networks.

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

[0005] Constructing a super subgraph including explicit and implicit regulatory relations from a multi-relation perspective, wherein the super subgraph is constructed in a data-driven manner, a knowledge-driven manner, and a rule-driven manner;

[0006] The hypergraph convolution operator is used to perform feature learning on the constructed hypergraphs, and high-order feature representations of nodes in different hypergraphs are extracted;

[0007] Enhance the consistency and distinguishability of node feature representation through contrastive learning, and optimize the feature representation across subgraphs;

[0008] The node feature representations of different super-subgraphs are fused through the attention mechanism to generate a globally optimized node descriptor.

[0009] Preferably, when the data-driven approach is adopted, the nearest neighbor hyperedge is generated based on the node embedding method of weighted random walk, and the neighbor sequence of any node is generated by the random walk formula:

[0010]

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

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

[0013]

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

[0015]

[0016] Among them, h is the Haar wavelet basis and r is the residual.

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

[0018] Preferably, the consistency and distinguishing ability of node feature representation are enhanced by contrastive learning. In the step of optimizing the feature representation 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 of different hyper-subgraphs through an attention mechanism, for each node, its attention weight in different subgraphs is calculated respectively, and the attention weight is dynamically allocated according to the feature importance of the node and the correlation strength of the hyper-edge. The feature representation of the node in each subgraph is multiplied by the corresponding attention weight to generate a weighted feature representation, and all weighted features are integrated into a global descriptor through a feature fusion operation.

[0020] The present invention provides a prediction method combining multi-relation hypergraph modeling and contrastive learning, which combines hypergraph modeling, contrastive learning and attention mechanism feature fusion technology to conduct an in-depth analysis of the regulatory mechanism and multi-relation network characteristics of cancer-related ncRNA markers, thereby achieving high-precision prediction of markers. Through the subgraph construction method of multi-relation background, the present invention can comprehensively capture explicit and implicit relationships, reveal high-order dependencies and long-distance regulatory patterns in the ceRNA regulatory network, and significantly improve the modeling ability of complex biological networks and the depth of understanding of key nodes. This refined modeling method based on multiple relationships not only improves the accuracy of prediction, but also provides important theoretical support for mining potential cancer markers. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0022] Figure 2 A flowchart of a prediction method combining multi-relational hypergraph modeling and contrastive learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

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

[0025] Constructing a super subgraph including explicit and implicit regulatory relations from a multi-relation perspective, wherein the super subgraph is constructed in a data-driven manner, a knowledge-driven manner, and a rule-driven manner;

[0026] The hypergraph convolution operator is used to perform feature learning on the constructed hypergraphs, and high-order feature representations of nodes in different hypergraphs are extracted;

[0027] Enhance the consistency and distinguishability of node feature representation through contrastive learning, and optimize the feature representation across subgraphs;

[0028] The node feature representations of different super-subgraphs are fused through the attention mechanism to generate a globally optimized node descriptor.

[0029] Although the existing cancer marker prediction technology has made significant progress, it still has some obvious shortcomings, mainly including the following: the model's high dependence on data, insufficient processing capabilities for multimodal features, difficulty in capturing complex relationships in high-order biological networks, and poor robustness to noisy data.

[0030] Traditional marker prediction methods mainly rely on statistical analysis and machine learning techniques, which usually require manual feature selection to build prediction models. However, manual feature selection may miss potential key features and make it difficult to fully capture the complex regulatory relationships in biological networks. In addition, with the deepening of research on the molecular mechanisms of cancer, marker data have shown high-dimensional and multimodal characteristics. For example, multiple data types such as transcriptomes, epigenetic groups, and proteomes are intertwined, and traditional methods are unable to integrate multimodal features. Although deep learning technology can improve prediction performance by automatically extracting features in some scenarios, it is highly dependent on large-scale high-quality data. In actual biomedical research, high-quality labeled data is often scarce, which greatly limits the performance of deep learning models.

[0031] At the same time, the regulatory network of cancer markers is complex and nonlinear. Existing methods are mostly based on traditional graph models or simple matrix operations, which are difficult to fully express and capture high-order dependencies between biological molecules. Especially in the competitive endogenous RNA regulatory network (ceRNA), the multi-molecule regulatory relationship between circRNA, LncRNA, miRNA and mRNA is complex, and a single binary relationship cannot meet the modeling requirements. In addition, biomedical data is usually accompanied by large noise. Existing technologies perform poorly when faced with noise interference in the data, which can easily lead to deviations in prediction results, thereby reducing the reliability of marker screening.

[0032] In view of the above technical defects, the present invention proposes a multimodal marker prediction method based on hypergraph modeling and contrastive learning to effectively solve the shortcomings of existing methods. This method constructs targeted hyperedges through three structural relationships, and uses a hypergraph model to model high-order regulatory relationships in the ceRNA network, thereby achieving comprehensive and accurate capture of complex relationships between molecules; combined with contrastive learning, it can efficiently process multimodal data, automatically extract key alignment features, and improve the robustness of the model to noisy data through end-to-end training. In addition, the present invention integrates multiple topological structure information through multi-view fusion, effectively alleviating the problem of insufficient representation ability of a single data structure, and improving the comprehensiveness and accuracy of marker prediction. This method provides a new technical framework for cancer marker prediction, while promoting the application of hypergraph modeling and contrastive learning technology in the field of bioinformatics, and has significant innovation, practicality and application prospects.

[0033] This paper proposes an innovative learning method based on multi-relation hypergraph comparison. The system architecture is as follows: Figure 1 As shown in the figure, it is dedicated to the high-precision prediction of cancer-related ncRNA markers. By integrating the multi-relational background of data, knowledge and rules to build subgraphs, and combining the hypergraph Laplace spectrum theory and contrastive learning technology, it realizes the comprehensive modeling of the ceRNA regulatory network, and through the multi-relational enhanced attention mechanism, adaptively optimizes and fuses the spatial features of different subgraphs to generate high-dimensional and interpretable node feature descriptors, which ultimately significantly improves the understanding of complex regulatory networks and the prediction performance of cancer markers.

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

[0035] (1) Subgraph construction module for multi-relational context

[0036] The subgraph construction module of multi-relational context aims to construct a super-subgraph containing explicit and implicit regulatory relationships from a multi-relational perspective, revealing high-order regulatory patterns and potential associations between molecules.

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

[0038] Generate initial relationship matrix and feature matrix through molecular omics data (such as transcriptome data, sequence similarity, functional annotation). Construct direct explicit relationship subgraph based on biological facts. For example, the regulatory relationship between miRNA and mRNA can be directly modeled as explicit edges based on existing experimental data.

[0039] For the association between circRNA and mRNA, the functional similarity or correlation of expression patterns is used to define explicit relationships. These explicit relationships are the basic edges of the super-subgraph, reflecting the direct biological significance between molecules.

[0040] Implicit relationships are an important supplement to subgraphs, which are generated based on statistical inference and rule mining. First, potential molecular grouping relationships are mined from the network through clustering algorithms (such as spectral clustering or density clustering); second, possible indirect regulatory relationships between molecules are extracted in combination with pathway analysis (such as KEGG pathways). 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 relationship. Based on explicit and implicit relationships, hyperedges are used to connect multiple nodes together to form a subgraph. For example, a hyperedge can include multiple circRNA and miRNA nodes, which may act together on a certain mRNA target to form a biologically meaningful regulatory unit. In this way, subgraphs can not only capture one-to-one molecular relationships, but also represent complex associations between multiple molecules.

[0041] There are three ways to build subgraphs: data, knowledge, and rules. The construction principles of each driving method are as follows:

[0042] Data driven:

[0043] The node embedding technology based on weighted random walk generates the nearest neighbor hyperedge. For any node, its neighbor sequence is generated by the following random walk formula:

[0044]

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

[0046] Knowledge Driven:

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

[0048]

[0049] Rule-driven:

[0050] The signal path characteristics are used to construct implicit hyperedges, and multi-scale features are extracted through wavelet transform:

[0051]

[0052] Among them, h is the Haar wavelet basis and r is the residual.

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

[0054] The hypergraph Laplacian spectral representation learning module performs feature learning on the constructed hypersubgraph through the hypergraph convolution operator to extract the high-order feature representation 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 ​​of each molecular node (such as expression value, functional annotation vector, etc.). In addition, the vertex-hyperedge association matrix of the hypergraph is constructed to record the connection relationship between the hyperedge and the node. 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 hyperedge. The purpose of the Laplacian matrix is ​​to retain the topological structure information between nodes during the feature propagation process. Using the hypergraph convolution operator, the node features are propagated from one hyperedge to other connected hyperedges. Through this feature aggregation operation, the node can obtain information from the neighboring nodes and update its own feature representation layer by layer. The updated features not only contain the node's own information, but also integrate the regulatory features of the neighboring nodes, reflecting the global position of the node in the subgraph. Through multi-layer hypergraph convolution operations, high-order features are gradually extracted. The first layer focuses on the local neighborhood information of the node; the second layer begins to capture long-distance dependencies; 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 in the subgraph. The hypergraph convolution operator is defined based on the Laplace spectrum theory, and its propagation mechanism is as follows:

[0055]

[0056] (3) Multi-relational hypergraph contrastive learning

[0057] The goal of contrastive learning in multi-relational hypergraphs is to enhance the consistency and distinguishing ability of node feature representations through contrastive learning, and to optimize the feature representations across subgraphs. For each node, positive and negative samples are generated. Positive samples come from the representations of the same node in different hypergraphs. Since these representations come from the same node, they should have high similarity. Negative samples select feature representations of other nodes to ensure that they are not directly related to the target node. By defining a contrastive loss function, the feature similarity between positive samples is maximized, while the feature similarity between positive and negative samples is minimized. The optimization goal of the loss function is to make the feature representations of the same node in different hypergraphs more consistent, while widening the feature differences between different nodes, thereby improving the distinguishing ability of node features. On the basis of contrastive learning, node features are aligned to make their representations in different relational subgraphs as consistent as possible. Through this alignment operation, nodes can obtain more comprehensive feature information across explicit and implicit relationships. The representation consistency of nodes is optimized through contrastive learning. Its objective function is defined as follows:

[0058]

[0059] Among them, t is the temperature parameter and sim is the click similarity.

[0060] (4) Multi-relation Enhanced Attention Mechanism

[0061] The multi-relation enhanced attention mechanism fuses the node feature representations of different hyper-subgraphs through a dynamic attention mechanism to generate a globally optimized node descriptor. For each node, its attention weight in different subgraphs is calculated separately. The attention weight is dynamically allocated according to the feature importance of the node and the correlation strength of the hyper-edge, with the aim of highlighting the importance of key relationships while suppressing the interference of irrelevant relationships. The feature representation of the node in each subgraph is multiplied by the corresponding attention weight to generate a weighted feature representation. Then, through the feature fusion operation, all weighted features are integrated into a global descriptor. The fused descriptor not only retains the feature differences of each subgraph, but also reflects the global regulatory characteristics of the node. The fused features are further optimized through the multi-head attention mechanism to ensure the accuracy and robustness of the feature fusion. The optimized node descriptor will be used for downstream prediction tasks, such as the prediction and verification of cancer markers. The multi-relation enhanced attention mechanism dynamically calculates the weight distribution of node features. Its core formula is:

[0062]

[0063] Among them, head is the number of attention heads, Q, K, and V are query, key, and value respectively.

[0064] (5) Module Collaboration Work Process

[0065] like Figure 2 As shown in the figure, the above four modules work together to complete the marker prediction task: starting from the subgraph construction module of the multi-relational background, the high-order features of the nodes are extracted through the hypergraph Laplacian spectral representation learning module, and the feature expression of the nodes is optimized by combining the multi-relational hypergraph comparative learning. Finally, the information of all subgraphs is enhanced through the multi-relational attention mechanism to generate accurate and comprehensive node feature representations. These node feature representations are input into the downstream classifier to achieve efficient prediction of cancer-related ncRNA markers.

[0066] Through systematic research and a large number of experimental verifications, the present invention confirms the important role of multi-relational hypergraph contrastive learning in the prediction of cancer-related ncRNA markers. The scheme uses data-driven, knowledge-driven and rule-driven multi-relational modeling methods to construct a multi-channel subgraph 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 the ceRNA network, but also provides solid data support and biological basis for a comprehensive understanding of the diverse regulatory mechanisms between molecules.

[0067] The method based on hypergraph convolution and multi-relation contrastive learning proposed in the present invention demonstrates excellent feature extraction ability and generalization performance when processing multi-relation subgraph features. This method uses the hypergraph convolution operator to strengthen the high-dimensional feature representation of the node, significantly improving the node's ability to capture explicit and implicit relationships. The multi-relation contrastive learning module further optimizes the feature consistency and distinction ability of the nodes, thereby ensuring the robustness and adaptability of the model in different feature spaces. In addition, the multi-head attention mechanism designed in this method dynamically integrates the feature representations of different subgraphs, which can not only accurately capture cross-relation collaborative information, but also improve the model's processing efficiency and prediction ability for high-dimensional complex data.

[0068] The scheme of the present invention combines hypergraph modeling, contrastive learning and attention mechanism feature fusion technology to conduct an in-depth analysis of the regulatory mechanism and multi-relationship network characteristics of cancer-related ncRNA markers, thereby achieving high-precision prediction of markers. Through the subgraph construction method of multi-relationship background, the present invention can comprehensively capture explicit and implicit relationships, reveal high-order dependencies and long-distance regulatory patterns in the ceRNA regulatory network, and significantly improve the modeling ability of complex biological networks and the depth of understanding of key nodes. This refined modeling method based on multiple relationships not only improves the accuracy of prediction, but also provides important theoretical support for mining potential cancer markers.

[0069] The present invention also demonstrates significant advantages in computational efficiency and data processing capabilities. By introducing the hypergraph convolution operator and multi-relational contrastive learning technology, the present invention can efficiently extract high-dimensional representations of nodes in different feature spaces and optimize the robustness and consistency of feature learning. The dynamic feature fusion strategy combined with the multi-head attention mechanism significantly improves the model's processing efficiency for high-dimensional data and sparse data. This fast and efficient feature extraction and fusion capability ensures that the marker prediction task can run stably on large-scale data sets, while being more adaptable to data noise and structural complexity.

[0070] In general, the present invention has demonstrated significant practicality and effectiveness in the field of cancer marker prediction. It not only solves the bottleneck problems of insufficient multi-relationship modeling and insufficient feature fusion in existing methods, but also provides innovative technical support and practical application value for promoting the application of precision medicine in cancer research.

[0071] In the prediction task of cancer-related ncRNA markers, the method proposed in the present invention has been extensively verified in three cancer marker prediction tasks: miRNA, LncRNA and circRNA. The experimental results are shown in Tables 1, 2 and 3. In the experiments, existing deep learning models (such as GCN, GAT, TransformerConv (TC)) were used as comparison benchmarks, and the models were tested and evaluated using real cancer ceRNA network datasets.

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

[0073]

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

[0075]

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

[0077]

[0078] In the prediction task of cancer-related ncRNA markers, the method proposed in the present invention performs well in processing different types of cancer marker prediction tasks, especially in the prediction of miRNA, lncRNA and circRNA markers, which is significantly better than the existing deep learning model. The comparison models used in the experiment include GCN, GAT and TC, which have been widely used in biological network analysis. The experimental results show that the method of the present invention performs well in multiple indicators such as accuracy (ACC), F1 score (F1-score), AUC (area under the receiver operating curve) and AUPR (area under the PR curve), verifying its powerful feature modeling ability and prediction performance. In the miRNA marker prediction task, the F1-score of the method of the present invention is 0.7403, which is significantly higher than GCN (0.7293) and TC (0.6967). In the lncRNA marker prediction task, the AUC of the method of the present invention reaches 0.8516, while GCN is only 0.8214. For the circRNA marker prediction task, the accuracy of the method of the present invention reached 0.7600, which is significantly higher than 0.7215 of GAT. These results show that the present invention has obvious advantages in capturing the complex regulatory relationship between molecules and improving the prediction accuracy, and has good generalization ability and robustness.

[0079] Existing deep learning models have 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 expression of neighborhood information through the attention mechanism, but its performance depends on the quality of the local neighborhood and its ability to capture implicit relationships is weak; although TC introduces the advantages of the Transformer architecture, its performance is limited by the data scale and complex relationship structure when processing long-distance dependencies and multi-relational features in biological networks.

[0080] The method proposed in the present invention combines hypergraph modeling, multi-relationship driven and contrastive learning optimization techniques, and shows excellent performance in processing high-order dependencies and multi-relationship features. Through the multi-relationship driven hypergraph construction method, the present invention can comprehensively capture explicit and implicit relationships and encapsulate high-order regulatory patterns in the ceRNA network in the form of a hypergraph. Subsequently, the hypergraph convolution operator is used to enhance the node feature representation, which significantly improves the modeling ability of long-distance dependencies between molecules. Furthermore, through contrastive learning optimization and multi-head attention mechanism, the present invention can adaptively fuse feature information in different subgraphs to generate globally optimized node feature descriptors. Experimental results show that the method of the present invention has achieved better performance than existing models in three cancer marker prediction tasks.

[0081] In summary, the multi-relationship driven multi-channel hypergraph contrastive learning method proposed in the present invention successfully integrates the multimodal information of biological networks and fully utilizes the technical advantages of hypergraph modeling and deep learning. Compared with existing models, the present invention not only achieves higher performance in multiple evaluation indicators, but also significantly enhances the modeling ability of complex biological networks through dynamic feature fusion and contrastive learning optimization, providing a powerful tool for the prediction and analysis of cancer-related ncRNA markers.

[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 in the protection scope of the present invention.

Claims

1. A prediction method combining multi-relation hypergraph modeling and contrastive learning, the method comprising: Constructing a super subgraph including explicit and implicit regulatory relations from a multi-relation perspective, wherein the super subgraph is constructed in a data-driven manner, a knowledge-driven manner, and a rule-driven manner; The hypergraph convolution operator is used to perform feature learning on the constructed hypergraphs, and high-order feature representations of nodes in different hypergraphs are extracted; Enhance the consistency and distinguishability of node feature representation through contrastive learning, and optimize the feature representation across subgraphs; The node feature representations of different super-subgraphs are fused through the attention mechanism to generate a globally optimized node descriptor.

2. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that: When taking a data-driven approach; The node embedding method based on weighted random walk generates the nearest neighbor hyperedge. For any node, its neighbor sequence is generated by the weighted random walk formula: Among them, A is the adjacency matrix, d(v j ) is the node degree, N(v j ) represents the neighbor set of a node.

3. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that: When using the knowledge-driven approach, hyperedges are generated through the k-nearest neighbor algorithm based on biological facts and experimental data, and their mathematical definition is:

4. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that: When the rule-driven approach is adopted, the signal path characteristics are used to construct implicit hyperedges, and multi-scale features are extracted through wavelet transform: Among them, h is the Haar wavelet basis and r is the residual.

5. The prediction method combining multi-relation hypergraph modeling and contrastive learning according to claim 1, characterized in that: In the hypergraph, the Laplacian matrix is ​​used to model the strength of association between nodes. The Laplacian matrix of the hypergraph is generated by calculating the weight matrix of the hyperedge, the node degree matrix, and the hyperedge degree matrix. The hypergraph convolution operator is used to propagate node features from one hyperedge to other connected hyperedges. Through this feature aggregation operation, high-order features are extracted through multi-layer hypergraph convolution operations.

6. 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 distinguishing ability of node feature representation through contrastive learning and optimizing the feature representation across subgraphs, positive samples and negative samples are generated for each node, the positive samples come from the representations of the same node in different super-subgraphs, and the feature similarity between positive samples is maximized by defining a contrastive loss function.

7. 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 of different hyper-subgraphs through the attention mechanism, for each node, its attention weight in different subgraphs is calculated respectively, and the attention weight is dynamically allocated according to the feature importance of the node and the correlation strength of the hyper-edge. The feature representation of the node in each subgraph is multiplied by the corresponding attention weight to generate a weighted feature representation. Through the feature fusion operation, all weighted features are integrated into a global descriptor.

Images