A structured graph inference method based on sample-feature cluster integrated graph
By constructing a sample-feature cluster integrated graph and introducing an adaptive correction mechanism for relationship strength, the problem of the difficulty in explicitly expressing cross-feature cluster relationships in existing technologies is solved, and structured graph inference in heterogeneous feature scenarios is realized, improving the interpretability and robustness of graph inference.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-19
AI Technical Summary
Existing graph inference methods lack a unified sample-feature cluster joint modeling structure when dealing with heterogeneous multimodal or multi-source features. They cannot explicitly express cross-feature cluster relationships and lack interpretable sample-level differential contribution interfaces and propagation intensity correction mechanisms, resulting in insufficient graph structure sparsity, interpretability, and task adaptability.
A sample-feature cluster integrated graph is constructed. By expressing the prior relations between feature clusters and the contribution of samples to each feature cluster, a sparse composite relation graph structure is formed. An adaptive correction mechanism for relation strength is introduced to realize structured graph inference.
It enhances the interpretability and stability of the graph inference process, improves the robustness of explicit expression of cross-feature cluster relationships and information propagation, and provides a unified structured graph inference interface suitable for multimodal and multi-source feature scenarios.
Smart Images
Figure CN122242767A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer application technology, specifically relating to a structured graph inference method based on a sample-feature cluster integrated graph. Background Technology
[0002] With the development of artificial intelligence, graph computing, and pattern recognition technologies, an increasing number of inference tasks require the simultaneous processing of heterogeneous feature data from different sources, modalities, or processing procedures. This type of data varies in dimensionality, numerical range, semantic level, statistical distribution, and preprocessing methods. While direct fusion using simple concatenation, unified mapping, or fixed weighting can form a single input representation, the structural relationships between different feature groups are compressed into a unified feature space. This makes it difficult to preserve cross-feature group information pathways and to provide a clear structural interface for subsequent propagation and inference processes.
[0003] To structure and organize the aforementioned heterogeneous features, a set of features sharing similarities in source, semantic relevance, statistical attributes, or processing procedures can be defined as a feature cluster. In multimodal scenarios, one modality can correspond to one or more feature clusters; in multi-source scenarios, related features from different sources can also be further divided into multiple feature clusters. By organizing the original features into feature clusters, dispersed heterogeneous features can be mapped into modelable structural units, thus providing a foundation for modeling relationships between feature clusters, expressing sample-level differences, and graph propagation inference. Existing research focuses on three main aspects regarding modeling needs for this type of problem: explicit structural interfaces, characterizing sample differences, and controlling relationship propagation.
[0004] Existing graph inference methods mainly employ two approaches. The first approach uses samples as nodes, constructs adjacency relationships between samples based on overall representation similarity, and then propagates and infers these relationships on the sample graph. While this approach can describe sample neighborhood relationships, interactions between different modalities or feature groups are collapsed into node attributes, failing to introduce cross-feature cluster dependencies into the graph propagation process as explicit relational structures. The second approach first encodes different modalities or feature groups separately, then fuses them in the representation space before executing downstream tasks. This approach enhances feature representation capabilities, but cross-feature group interactions are mainly achieved through implicit fusion, lacking structured expression and traceable propagation links.
[0005] Furthermore, existing methods have the following shortcomings. First, they lack a unified graph structure for joint modeling of samples and feature clusters, making it difficult to simultaneously express the relationships between "different feature clusters within a sample" and "same or different feature clusters between different samples," thus preventing cross-feature cluster dependencies from entering the propagation and inference process as a relation type. Second, existing fusion methods mostly use globally fixed weights or implicit attention coefficients, lacking sample-level, interpretable weight interfaces that can directly participate in graph construction and convergence, making it difficult to express the differentiated contributions of different samples on different feature clusters. Third, existing adjacency structures are mainly determined by overall similarity rules, and the graph structure only provides neighborhood constraints, without providing a unified definition of "which feature clusters are allowed to propagate, how the propagation direction is defined, and how the propagation intensity is corrected during training," making it difficult for the graph structure to simultaneously meet the requirements of sparsity, interpretability, and task adaptability.
[0006] Building upon this, directly employing a fully connected approach across samples and feature clusters to construct the graph would introduce a large number of redundant connections, increasing propagation noise and computational burden, and weakening the stability of the structural inference process. Conversely, retaining only local similarity adjacency relationships makes it difficult to explicitly express the direction, scope, and type of cross-feature cluster information flow, thus failing to form a unified structured inference link. Therefore, how to maintain the sparsity of the graph structure while introducing prior relationships between feature clusters, the differentiated contribution of samples to feature clusters, and an adaptive correction mechanism for relationship strength has become a key problem that graph inference methods in heterogeneous multimodal scenarios need to solve. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, this invention proposes a structured graph inference method based on a sample-feature cluster integrated graph, addressing the lack of a unified sample-feature cluster joint modeling structure, a lack of sample-level differentiated contribution interfaces that can directly participate in graph construction and aggregation, and a lack of a unified definition and adaptive correction mechanism for cross-feature cluster propagation direction and intensity in heterogeneous multimodal or multi-source feature scenarios. First, the input features are structurally grouped to construct prior relationships between feature clusters and establish the contribution expression of samples to each feature cluster. Second, using sample-feature cluster combinations as integration nodes, a sparse composite relationship graph structure with relationship type distinctions is constructed. Finally, an adaptive correction mechanism for relationship strength is introduced on a fixed sparse relationship skeleton, combining composite relationship graph propagation and sample-level weighted aggregation to achieve structured graph inference output. This method uses feature clusters as structured modeling units and sample-feature cluster combinations as integration nodes. It simultaneously introduces the contribution information of samples to each feature cluster and prior relationships between feature clusters into the integrated graph, and adaptively corrects the propagation intensity on a fixed sparse relationship skeleton, thereby achieving structured relationship modeling, information propagation, and inference output for heterogeneous feature inputs.
[0008] This invention is roughly divided into two parts: (1) Constructing a sample-feature cluster integrated graph. First, the input sample features are structurally grouped to obtain multiple feature clusters; second, prior relationships between feature clusters are constructed based on the feature representations of each feature cluster, and the contribution expression of the sample to each feature cluster is constructed; then, the combination of sample and feature cluster is used as a comprehensive node, and sparse composite relation edges with relation type distinctions are generated according to the prior relationships between feature clusters and the sample contribution, thus forming a sample-feature cluster integrated graph. This part is used to establish a unified graph structure representation for heterogeneous feature inputs.
[0009] (2) Perform structured graph inference based on the comprehensive graph. An adaptive correction mechanism for relation strength is introduced on the fixed sparse relation skeleton of the sample-feature cluster comprehensive graph to adjust the propagation strength of different relation types. Subsequently, composite relation propagation is performed on the comprehensive graph to obtain comprehensive node representations. Finally, sample-level weighted aggregation is performed on multiple comprehensive node representations corresponding to the same sample to obtain sample representations, and the structured graph inference results are output. This part is used to complete information propagation and inference output on the constructed comprehensive graph structure.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A structured graph inference method based on sample-feature cluster integrated graph, assuming a total Each sample is divided into input features. The feature cluster, the first Each sample is represented as ,in Indicates sample In the The input feature vector on the feature cluster, which is derived from the ... It consists of multiple features within a feature cluster. By combining prior relationships between feature clusters, the differentiated contributions of samples to each feature cluster, and a structured propagation mechanism, a mapping function from multi-feature cluster input to target output is learned. ,in, Indicates sample The corresponding target output, the format of which can be determined according to the specific task, can be continuous regression values, discrete category labels, category distributions, risk distributions, rank probability vectors, or other forms of structured prediction results. The specific steps are as follows: Step (1) Construct a sample-feature cluster integrated graph (1.1) Feature Cluster Partitioning. The input sample features are structurally grouped, and features with similar sources, semantic relevance, statistical attributes, or processing procedures are grouped into the same feature cluster, resulting in... The first feature cluster. Let the first feature cluster be... The feature matrix corresponding to each feature cluster is: ,in, Indicates the number of samples. Indicates the first Each feature cluster has a feature dimension. Each feature cluster serves as the basic structural unit for subsequent relation modeling, weight assignment, and graph propagation.
[0011] (1.2) Construction of prior relations between feature clusters. The feature matrices corresponding to each feature cluster are preprocessed to obtain the feature cluster representations used for relation estimation. The calculation formula is: (1) in, This indicates that the feature matrix is standardized column-wise. This indicates that the standardized sample vectors are normalized. Represents a centered matrix. It is the identity matrix. It is a vector consisting entirely of 1s.
[0012] Based on the preprocessed feature cluster representation, calculate any two feature clusters and The strength of prior relationships between feature clusters is used to construct a prior relationship matrix between feature clusters. The calculation formula is: (2) in, Represents the matrix trace operation. Representing feature clusters With feature clusters The strength of the prior relationship between them. The strength of the prior relationship is calculated using the centered linear CKA method. In the formula, the diagonal terms... Indicates the first The intra-cluster relationships between each feature cluster and itself. To highlight the cross-cluster relationships between different feature clusters, the matrix... Suppress the diagonal terms, that is, let .
[0013] (1.3) Modeling the contribution of each sample to each feature cluster. Based on the feature response of each sample to each feature cluster, construct the original contribution intensity of the sample to each feature cluster. The calculation formula is: (3) in, Indicates sample In the In the feature cluster, the th Standardized values of 3D features Indicates sample For feature clusters The intensity of the original contribution.
[0014] Normalization is performed on the original contribution intensities to obtain the contribution matrix of the sample to each feature cluster. The calculation formula is as follows: (4) in, For numerically stable terms, This represents the feature cluster index variable in the summation, used for traversing samples. For all The original contribution strength of each feature cluster, Indicates sample For feature clusters The relative contribution ratio. The contribution matrix As a sample-level differentiated weight interface, it participates in both edge weight calculation and sample-level weighted aggregation in subsequent steps.
[0015] (1.4) Definition of Sample-Feature Cluster Synthesis Node. Each combination of a sample and each feature cluster is treated as a sample-feature cluster synthesis node, constructing a set of synthesis nodes. Specifically, the sample... With feature clusters The combination corresponds to a comprehensive node. The original features, preprocessed features, or combinations thereof corresponding to this combination are used as the input attributes of this synthesis node.
[0016] (1.5) Construction of sparse composite relation edges. Based on the prior relation matrix between the feature clusters... and the contribution matrix Under preset constraints, source feature clusters are selected for each target feature cluster, and candidate sample adjacency relationships are screened within the source feature clusters. Then, sparse composite relationship edges between the comprehensive nodes are generated based on the combination relationships between source samples, source feature clusters, target feature clusters, and target samples. For the source comprehensive nodes... Pointing to the target integrated node The normalized edge weight of the edge is defined as: (5) in, and These represent the source sample index and the target sample index, respectively. and These represent the source feature cluster index and the target feature cluster index, respectively. This represents the source aggregation node index variable in the summation, where For the source sample index variable, This is the source feature cluster index variable, used to traverse the target synthesis nodes. candidate source neighborhood set Each candidate source node in the data; Represents the target integrated node The candidate source neighborhood set, This represents the normalized edge weight corresponding to the sparse composite relation edge. This results in a sample-feature cluster composite graph with relation type distinctions, where the relation type is an ordered combination of the source and target feature clusters. Sure.
[0017] Step (2) Perform structured graph inference based on the comprehensive graph (2.1) Adaptive Correction of Relationship Strength. Without changing the sparse composite relation edge set obtained in step (1), a learnable relation gating matrix between feature clusters is introduced to adaptively correct the propagation strength of different "source feature cluster → target feature cluster" relation types. To enable gating learning to start from prior relations, a trainable parameter matrix is introduced. And based on the prior relation matrix between feature clusters As the initialization baseline, the initialization method is as follows: (6) in, For numerically stable terms, Indicates the source feature cluster Target feature cluster The initial gating parameters. Perform temperature parameterization on the parameter matrix. The column softmax transform is applied, and a self-loop mixing coefficient is introduced. The gating matrix is obtained. The calculation formula is: (7) in, Indicates temperature parameter, Represents the self-loop mixing coefficient. This indicates that a softmax transformation is performed column-wise, which is used to normalize the gating strength of feature clusters from different sources when the target feature cluster is fixed. This indicates that column-wise normalization is performed again to maintain the column sum of the gating weights at 1 after introducing self-loop mixed terms; This represents the identity matrix. Therefore, Indicates updating the target feature cluster Time comes from source feature cluster The proportion of path strength. The gating matrix only adjusts the relative strength of different relationship types, without changing the established sparse structural skeleton.
[0018] (2.2) Composite Relationship Graph Propagation. Given the input features of each sample-feature cluster synthesis node, first apply the input projection to different feature clusters respectively, mapping each synthesis node to a unified hidden space to obtain the initial node representation. Let the source aggregation node be... The target integrated node is ,in, and These represent the source sample index and the target sample index, respectively. and These represent the source feature cluster index and the target feature cluster index, respectively. Therefore, the edge... The corresponding relation type is denoted as In the first In the propagation of the hierarchical relationship graph, for any target synthesis node Incoming messages are aggregated according to relation type. The final propagation coefficient after gating correction is defined as: (8) in, This represents the normalized edge weights obtained in step (1). Indicates "source feature cluster" Target feature clusters "The gating coefficient corresponding to the relation type, Indicates pointing to the target synthesis node The set of incoming edge sources, This represents the source node index variable in the summation, used for traversing the set. Each candidate source node in the process, This represents the final propagation coefficient after gating correction and normalization.
[0019] Based on this, message passing is performed using a parameter matrix determined by the relation type for each relation type, with the target node at the . Layer-based relational aggregation and updates can be represented as: (9) in, Represents the target integrated node In the The hidden representation after the layer update, Represents the target integrated node In the Hidden representation of layers, Indicates the source of the integrated node In the Hidden representation of layers; Indicates the first Layer and relation type The corresponding learnable parameter matrix, Indicates the first Layer self-loop branching transformation matrix; Indicates pointing to the target synthesis node The set of source nodes; Presentation layer normalization operation. After... After propagating the layer relationship graph, the hidden representation of each synthesis node in the last layer is obtained. It serves as the basic representation for subsequent sample-level weighted aggregation and prediction head input.
[0020] (2.3) Sample-level weighted convergence. After obtaining the hidden representation of each sample-feature cluster integrated node, the sample contribution matrix to each feature cluster constructed in step (1) is used as the basis for the convergence. For the same sample The corresponding multiple integrated node representations are weighted and aggregated at the sample level to obtain the sample representation. The calculation formula is: (10) in, Indicates sample In the Each feature cluster corresponds to a comprehensive node after passing through The last hidden representation obtained after propagation of the layer relationship graph Indicates sample The aggregated representation. The sample-level weighted aggregation is represented by the contribution matrix. Differentiated weighting is applied to the comprehensive node representations corresponding to different feature clusters to form a unified representation for sample-level inference.
[0021] (2.4) Output of structured graph inference results. The sample representations obtained in step (2.3) after relational graph propagation learning and sample-level weighted aggregation are... Input the prediction header and output the structured inference results corresponding to the target task.
[0022] The prediction head can be a linear transform layer or a multilayer perceptron, used to represent samples. Mapped to a preset output space. Let the prediction head output be... Then we have: (11) in, Indicates sample The structured graph inference results are supervised by the target output defined in the aforementioned task context and abstract definition. The target output can be a continuous regression value, discrete class label, class distribution, risk distribution, rank probability vector, or other form of structured prediction result. For tasks that require satisfying specific constraints, the prediction head output can be further mapped accordingly.
[0023] During the training phase, the target output is utilized from the samples. Compared with the prediction results A supervised loss matching the task format is constructed, and the input projection parameters, relation gating parameters, relation type parameter matrix, self-loop branch parameters, and prediction head parameters are jointly optimized to achieve end-to-end structured graph inference learning from multi-feature cluster input to target output.
[0024] The beneficial effects of this invention are: This invention constructs a sample-feature cluster integrated graph, organizing heterogeneous multimodal or multi-source features into explicitly modelable graph structure units. This allows cross-feature cluster relationships to enter the propagation and inference process in the form of relationship types and graph paths. Simultaneously, by constructing a contribution matrix of samples to each feature cluster, the relative contributions of different samples to different feature clusters are explicitly quantified and uniformly used for edge weight calculation and sample-level weighted aggregation, thereby enhancing the interpretability and traceability of the structured graph inference process. Furthermore, this invention constructs a sparse composite relationship edge set based on prior relationships between feature clusters and introduces a learnable relationship gating mechanism on a fixed sparse relationship skeleton to adaptively correct the propagation intensity of different relationship types. This reduces redundant connections and propagation noise while improving the stability and task adaptability of the relationship propagation process. Further, this invention adopts a combination of composite relationship graph propagation and sample-level weighted aggregation to complete heterogeneous relationship propagation, sample representation construction, and result output within a unified graph structure. This provides a unified structured graph inference interface for multimodal, multi-source, or heterogeneous feature input scenarios, exhibiting good versatility, scalability, and robustness. Attached Figure Description
[0025] Figure 1 This is an overall framework diagram of the present invention.
[0026] Figure 2 This is a schematic diagram of the sample-feature cluster integrated graph structure of the present invention.
[0027] Figure 3 This is a flowchart of the method of the present invention.
[0028] Figure 4 This is a schematic diagram of the adaptive correction of relation strength and propagation of composite relations according to the present invention. Detailed Implementation
[0029] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.
[0030] This invention can be used for structured graph inference in scenarios with multimodal, multi-source, or heterogeneous feature inputs, and is particularly useful for tasks that predict the risk state distribution of a target at a given time point based on baseline multimodal features. The overall framework is as follows: Figure 1 As shown, the sample-feature cluster integrated graph structure is as follows: Figure 2 As shown, the method flow is as follows: Figure 3 As shown, the adaptive correction of relation strength and the propagation process of composite relations are as follows: Figure 4 As shown below, the specific implementation of this invention is illustrated with an example of risk prediction related to an implantable cardioverter-defibrillator (ICD) intervention.
[0031] This implementation scenario is a task for predicting ICD intervention-related risks based on baseline multimodal cardiovascular-related characteristics. The scenario is built using the UK Biobank database. The input is multimodal cardiovascular-related characteristics collected from the same subject at baseline, and the output is the future risk outcome of ICD intervention-related events occurring at the target follow-up time point for that subject. In this scenario, the future risk outcome is divided into three states: low risk, medium risk, and high risk, and the corresponding three-state distribution is used as the prediction output form. Specifically, the high-risk state corresponds to malignant ventricular arrhythmias, cardiac arrest, or actual ICD implantation events; the medium-risk state corresponds to the baseline structural heart disease state related to ICD primary prevention; and the low-risk state corresponds to the absence of the aforementioned events during the follow-up period. The target label can be constructed using the inpatient diagnosis ICD-10 field and the inpatient procedure OPCS-4 field. The codes related to malignant ventricular arrhythmias may include I47.2, I49.0, I46.1 and I46.9, and the codes related to ICD implantation procedures may include K59.1, K59.2, K59.6 and K72.1.
[0032] In this implementation scenario, the input features are divided into five feature clusters according to modality: demographic features, blood biochemistry features, anthropometric features, past medical history features, and cardiac magnetic resonance-derived phenotypic features. All of these features can be obtained from baseline survey data, biochemical test data, health outcome linked data, and cardiac magnetic resonance-derived phenotypic data from UK Biobank.
[0033] (1) Constructing a sample-feature cluster composite graph. First, the input sample features are structured and grouped to obtain K=5 feature clusters. Let there be N subjects in total, and let the Kth feature cluster be the first feature cluster. The feature matrix corresponding to each feature cluster is: Subsequently, the feature matrices corresponding to each feature cluster are preprocessed, and the feature cluster representation is obtained according to formula (1). The prior relation matrix between feature clusters is constructed according to formula (2). In this implementation scenario, the prior relation matrix is calculated using centered linear CKA, with diagonal terms suppressed, while self-loop pathways are preserved in the subsequent mapping stage. Simultaneously, the contribution matrix of the subjects to each feature cluster is calculated according to formulas (3) and (4). .
[0034] In obtaining the prior relation matrix and contribution matrix Then, the combination of each subject and each feature cluster is treated as a sample-feature cluster synthesis node, with the original features of the corresponding feature cluster as the node input. Subsequently, based on the prior relation matrix between feature clusters... and contribution matrix Construct a sparse composite relation edge set. In this implementation scenario, for each target feature cluster, from the matrix... The two non-originating feature clusters with the highest similarity are selected, and the original feature cluster is also retained. At the same time, within each source feature cluster, 20 nearest neighbor samples are selected for each subject as candidate source nodes. Finally, the normalized edge weights are calculated according to formula (5) to form a sample-feature cluster comprehensive graph with relational type distinction.
[0035] (2) Perform structured graph inference based on the comprehensive graph. After constructing the sample-feature cluster comprehensive graph, a learnable feature cluster relation gating matrix is introduced without changing the sparse composite relation edge set. Specifically, the gating parameters are initialized according to formula (6), and the gating matrix is obtained according to formula (7) to adaptively correct the propagation intensity of different relation types. In this implementation scenario, the gating temperature parameter is gradually reduced from 2.0 to 0.7, and the self-loop mixing coefficient is gradually reduced from 0.50 to 0.20 to balance the stability in the early stage of training and the relation discrimination ability in the later stage of training.
[0036] Subsequently, input projection is applied to the input features of the integrated nodes to map each integrated node to a unified hidden space; then, the final propagation coefficient is calculated according to formula (8), and composite relationship graph propagation is performed according to formula (9) to obtain the hidden representation of each integrated node. Afterwards, the weighted convergence of multiple integrated node representations corresponding to the same subject is performed according to formula (10) to obtain the subject-level representation. Finally, the subject-level representation is input into the prediction head to obtain the unnormalized result of the target output, and then subjected to softmax mapping to obtain the final three-state distribution prediction result of ICD intervention-related risks.
[0037] In this implementation scenario, the maximum number of training epochs is set to 300, and the early stopping patience is set to 20. During training, a priori relation matrix between feature clusters is first constructed from the training set data. and subject-feature cluster contribution matrix The model generates a sparse composite relation edge set. Simultaneously, based on inpatient diagnosis records, inpatient procedure records, and corresponding event date information from the UK Biobank, it determines the risk status label for each subject within the target follow-up window. Subsequently, in each training round, gating correction, relation propagation, subject-level weighted aggregation, and prediction output are performed sequentially, and model parameters are updated via backpropagation. After training, the predicted three-state distribution of ICD intervention-related risks for the target subjects is output during the inference phase.
[0038] Through the aforementioned mapping, propagation, and convergence processes, the model can simultaneously express cross-feature cluster interactions and subject-specific feature cluster importance in ICD intervention-related risk prediction tasks, and output risk distribution prediction results with traceable interpretable interfaces. Experimental results show that the risk prediction results output by this invention have a good correlation with the actual medical outcomes occurring within the target follow-up window. They can reflect whether subjects subsequently experience malignant ventricular arrhythmias, cardiac arrest, or actual ICD implantation events, indicating a pre-existing structural heart disease state related to ICD primary prevention, or whether the aforementioned events did not occur. This provides a data-driven basis for relevant risk stratification, application scenario selection, and subsequent decision support.
Claims
1. A structured graph inference method based on a sample-feature cluster integrated graph, characterized in that, Assume there is a total Each sample is divided into input features. The feature cluster, the first Each sample is represented as ,in Indicates sample In the The input feature vector on the feature cluster, which is derived from the ... It consists of multiple features within a feature cluster; by combining prior relationships between feature clusters, the differential contributions of samples to each feature cluster, and the structured propagation mechanism, a mapping function from multi-feature cluster input to target output is learned. ,in, Indicates sample The corresponding target output is the structured prediction result; the specific steps are as follows: Step (1) Construct a sample-feature cluster integrated graph (1.1) Feature cluster partitioning; (1.2) Construction of prior relations between feature clusters; (1.3) Modeling the contribution of samples to each feature cluster; (1.4) Definition of sample-feature cluster integrated node; (1.5) Construction of sparse composite relation edges; Step (2) Perform structured graph inference based on the comprehensive graph (2.1) Adaptive correction of relationship strength; (2.2) Propagation of composite relationship diagrams; (2.3) Sample-level weighted aggregation; (2.4) Output of structured graph inference results.
2. The structured graph inference method based on sample-feature cluster comprehensive graph according to claim 1, characterized in that, Step (1) is as follows: (1.1) Feature cluster partitioning: The input sample features are structurally grouped, and a group of features with consistent source, semantic relevance, similar statistical attributes, or consistent processing flow are divided into the same feature cluster, resulting in... The feature cluster; let the first feature cluster be denoted as . The feature matrix corresponding to each feature cluster is: ,in, Indicates the number of samples. Indicates the first Each feature cluster has a feature dimension; each feature cluster serves as the basic structural unit for subsequent relation modeling, weight allocation, and graph propagation. (1.2) Construction of prior relations between feature clusters: The feature matrices corresponding to each feature cluster are preprocessed to obtain the feature cluster representations used for relation estimation. ; Based on the preprocessed feature cluster representation, calculate any two feature clusters and The strength of prior relationships between feature clusters is used to construct a prior relationship matrix between feature clusters. ; (1.3) Modeling the contribution of each sample to each feature cluster; constructing the original contribution intensity of each sample to each feature cluster based on the feature response of each sample on each feature cluster. ; Normalization is performed on the original contribution intensities to obtain the contribution matrix of the sample to each feature cluster. ; (1.4) Definition of Sample-Feature Cluster Synthesis Node; Each combination of a sample and each feature cluster is treated as a sample-feature cluster synthesis node, and a set of synthesis nodes is constructed. Specifically, the sample With feature clusters The combination corresponds to a comprehensive node. The original features, preprocessed features, or combined features corresponding to this combination are used as the input attributes of the integrated node. (1.5) Construction of sparse composite relation edges; based on the prior relation matrix between the feature clusters. and the contribution matrix Under preset constraints, source feature clusters are selected for each target feature cluster, and candidate sample adjacency relationships are screened within the source feature clusters. Then, sparse composite relationship edges between comprehensive nodes are generated based on the combination relationship between source samples, source feature clusters, target feature clusters and target samples.
3. The structured graph inference method based on sample-feature cluster comprehensive graph according to claim 1, characterized in that: In step (1.2), the feature cluster representation used for relation estimation is... The calculation formula is: （1） in, This indicates that the feature matrix is standardized column-wise. This indicates that the standardized sample vectors are normalized. Represents a centered matrix. It is the identity matrix. It is a vector consisting entirely of 1s; Prior relation matrix between feature clusters The calculation formula is: （2） in, Represents the matrix trace operation. Representing feature clusters With feature clusters The strength of the prior relation between them; the strength of the prior relation is calculated using the centered linear CKA method; where the diagonal terms Indicates the first The intra-cluster relationships between each feature cluster and itself; to highlight the cross-cluster relationships between different feature clusters, the matrix... Suppress the diagonal terms, that is, let ; In step (1.3), the original contribution intensity of each feature cluster The calculation formula is: （3） in, Indicates sample In the In the feature cluster, the th Standardized values of 3D features Indicates sample For feature clusters The intensity of the original contribution; Contribution matrix of each feature cluster The formula for calculating it is: （4） in, For numerically stable terms, This represents the feature cluster index variable in the summation, used for traversing samples. For all The original contribution strength of each feature cluster, Indicates sample For feature clusters The relative contribution ratio; the contribution matrix As a sample-level differentiated weight interface, it participates in both edge weight calculation and sample-level weighted aggregation in subsequent steps. In step (1.5), for the source synthesis node Pointing to the target integrated node The normalized edge weight of the edge is defined as: （5） in, and These represent the source sample index and the target sample index, respectively. and These represent the source feature cluster index and the target feature cluster index, respectively. This represents the source aggregation node index variable in the summation, where For the source sample index variable, This is the source feature cluster index variable, used to traverse the target synthesis nodes. candidate source neighborhood set Each candidate source node in the data; Represents the target integrated node The candidate source neighborhood set, This represents the normalized edge weight corresponding to the sparse composite relation edge; thus forming a sample-feature cluster composite graph with relation type distinction, where the relation type is an ordered combination of the source feature cluster and the target feature cluster. Sure.
4. The structured graph inference method based on sample-feature cluster comprehensive graph according to claim 1, characterized in that, Step (2) is as follows: (2.1) Adaptive correction of relation strength; without changing the sparse composite relation edge set obtained in step (1), a learnable relation gating matrix between feature clusters is introduced to adaptively correct the propagation strength of different "source feature cluster → target feature cluster" relation types; in order to enable gating learning to start from prior relations, a trainable parameter matrix is introduced. And based on the prior relation matrix between feature clusters As an initialization benchmark; And introduce a self-cyclic mixing coefficient The gating matrix is obtained. ; (2.2) Composite Relationship Graph Propagation: Given the input features of each sample-feature cluster synthesis node, first apply the input projection to different feature clusters respectively, and map each synthesis node to a unified hidden space to obtain the initial node representation. Let the source aggregation node be... The target integrated node is ,in, and These represent the source sample index and the target sample index, respectively. and These represent the source feature cluster index and the target feature cluster index, respectively; edges The corresponding relation type is denoted as ; in the In the propagation of the hierarchical relationship graph, for any target synthesis node Aggregate incoming edge messages according to relation type; the final propagation coefficient after gating correction is defined as... ; For each relation type, message passing is performed using a parameter matrix determined by the relation type; (2.3) Sample-level weighted convergence; After obtaining the hidden representation of each sample-feature cluster integrated node, the contribution matrix of the sample to each feature cluster constructed in step (1) is used to converge the samples. For the same sample The corresponding multiple integrated node representations are weighted and aggregated at the sample level to obtain the sample representation. ; (2.4) Output of structured graph inference results; the sample representation obtained in step (2.3) after relational graph propagation learning and sample-level weighted aggregation. Input the prediction header and output the structured inference results corresponding to the target task.
5. The structured graph inference method based on sample-feature cluster integrated graph according to claim 4, characterized in that: In step (2.1), the initialization method is as follows: （6） in, For numerically stable terms, Indicates the source feature cluster Target feature cluster Initial gating parameters; perform temperature parameterization on the parameter matrix. The column softmax transform is applied, and a self-loop mixing coefficient is introduced. The gating matrix is obtained. The calculation formula is: （7） in, Indicates temperature parameter, Represents the self-loop mixing coefficient. This indicates that a softmax transformation is performed column-wise, which is used to normalize the gating strength of feature clusters from different sources when the target feature cluster is fixed. This indicates that column-wise normalization is performed again to maintain the column sum of the gating weights at 1 after introducing self-loop mixed terms; Represents the identity matrix; Indicates updating the target feature cluster Time comes from source feature cluster The proportion of pathway strength; In step (2.2), the final propagation coefficient after gating correction is defined as: （8） in, This represents the normalized edge weights obtained in step (1). Represents "source feature cluster" Target feature cluster "The gating coefficient corresponding to the relation type, Indicates pointing to the target synthesis node The set of incoming edge sources, This represents the source node index variable in the summation, used for traversing the set. Each candidate source node in the process, This represents the final propagation coefficient after gating correction and re-normalization; The target node is at the Layer-based relational aggregation and updates can be represented as: （9） in, Represents the target integrated node In the The hidden representation after the layer update, Represents the target integrated node In the Hidden representation of layers, Indicates the source of the integrated node In the Hidden representation of layers; Indicates the first Layer and relation type The corresponding learnable parameter matrix, Indicates the first Layer self-loop branch transformation matrix; Indicates pointing to the target synthesis node The set of source nodes; Presentation layer normalization operation; after After propagating the layer relationship graph, the hidden representation of each synthesis node in the last layer is obtained. This serves as the basic representation for subsequent sample-level weighted aggregation and prediction head input; In step (2.3), sample characterization The calculation formula is: （10） in, Indicates sample In the Each feature cluster corresponds to a comprehensive node after passing through The last hidden representation obtained after propagation of the layer relationship graph Indicates sample The aggregated representation; the sample-level weighted convergence is represented by the contribution matrix. Differentiated weighting is applied to the comprehensive node representations corresponding to different feature clusters to form a unified representation for sample-level inference.
6. The structured graph inference method based on sample-feature cluster integrated graph according to claim 4, characterized in that, In step (2.3), the prediction head can be a linear transform layer or a multilayer perceptron, used to represent the samples. Mapped to a preset output space; let the prediction head output be... Then we have: （11） in, Indicates sample The structured graph inference results are supervised by the target output defined in the aforementioned task context and abstract definition. .
7. The structured graph inference method based on sample-feature cluster comprehensive graph according to claim 4, characterized in that, In step (2.3), the target output is a structured prediction result in the form of continuous regression values, discrete category labels, category distribution, risk distribution, and grade probability vectors; for tasks that need to meet specific constraints, the prediction head output can be further mapped accordingly.
8. The structured graph inference method based on sample-feature cluster comprehensive graph according to claim 4, characterized in that, In step (2.3), during the training phase, the target output of the samples is used. Compared with the prediction results A supervised loss matching the task format is constructed, and the input projection parameters, relation gating parameters, relation type parameter matrix, self-loop branch parameters, and prediction head parameters are jointly optimized to achieve end-to-end structured graph inference learning from multi-feature cluster input to target output.