A cross-institutional collaboration method and system based on graph federated learning

Abstract

The application discloses a cross-institution cooperation method and system based on graph federated learning, and the method comprises the following steps: repairing the loss of graph signals caused by cross-institution cutting edges through space-spectrum dual-domain compensation; eliminating multi-source heterogeneity through semantic-structure dual-axis alignment; improving the cross-institution generalization ability of the federated model through sharpness perception optimization; and realizing joint modeling under the premise that the original data does not leave the local by means of three-level privacy protection of differential privacy, homomorphic encryption and secure aggregation; and finally generating a cross-institution scientific research graph data set with larger scale and higher confidence by using global model reasoning and multi-source confidence weighted fusion. The application can effectively alleviate the precision decline problem of sub-graph federated learning under the privacy compliance constraint, and improve the convergence speed and generalization ability between heterogeneous participants.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of federated learning and graph neural networks, specifically to a cross-institutional collaboration method and system based on graph federated learning. Background Technology

[0002] Graph Neural Networks (GNNs) have demonstrated significant advantages in AI for Science (AI4S) scenarios such as drug target prediction, protein-protein interaction analysis, and disease association discovery. Their core lies in using graph structures to model complex relationships between entities. However, high-value labeled scientific graph data is scattered across multiple hospitals, research institutes, and pharmaceutical companies. The data standards, ontology definitions, and node attribute dimensions of these institutions are inconsistent, creating data silos that are difficult to share.

[0003] Current typical solutions for cross-institutional graph data collaboration fall into two categories: centralized data aggregation and anonymized sharing. Centralized data aggregation copies the raw data from all parties to a unified platform for retraining the GNN. However, medical and research data often involve personal privacy and sensitive information, and are subject to increasingly stringent data privacy protection and compliance requirements. Raw data is usually not allowed to leave the local institution, making centralized aggregation solutions difficult to implement. Anonymized sharing reduces privacy risks by anonymizing or generalizing node attributes, but edge relationships and higher-order topological patterns in the graph structure can still be used for reverse inference, leading to privacy re-identification and limited protection effectiveness.

[0004] The most relevant prior art to this invention is the federated graph learning method within the Federated Graph Neural Network (FedGraphNN) framework. This method allows each participant to retain the original graph data locally, interacting only with model gradients or node embeddings, while a central server aggregates and updates the global GNN parameters. However, this framework has two key drawbacks: First, when graph data is partitioned by institution, cross-client edges are cut off, causing spatial domain message passing to fail and graph signal propagation to be hindered, resulting in a significant decrease in the model accuracy of subgraph-level federated learning. Second, the feature dimensions and topological structures of graph data from different institutions vary greatly (heterogeneity), making it difficult for a simple weighted aggregation global model to accommodate different data distributions, resulting in slow federated convergence and poor generalization. Therefore, how to simultaneously solve the two core problems of graph signal loss caused by cross-client edge cutting and alignment of multi-source heterogeneous data distributions, and generate a large-scale, high-confidence cross-institutional research graph dataset under privacy constraints, has become a technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-institutional collaboration method and system based on graph federated learning. It aims to solve the problems of existing federated graph learning methods in cross-institutional collaboration, such as the failure of spatial domain message passing and the loss of graph signal due to the severing of cross-institutional edges caused by the division of institutional boundaries, and the difficulty of federated aggregation, slow convergence, and poor generalization due to the highly heterogeneous feature dimensions and topological structure of multi-source graph data, which makes it impossible to generate large-scale, high-confidence cross-institutional research graph datasets without leaving the original data domain.

[0006] To address the aforementioned technical problems, this invention provides a cross-institutional collaboration method based on graph federation learning, comprising:

[0007] Local raw scientific research data is mapped to node attributes and edge relationships to construct a standardized local subgraph. A shared ontology tree is used to perform graph structure verification on the standardized local subgraph and extract the local subgraph summary to obtain the verified local subgraph and the local subgraph summary.

[0008] To address the graph signal loss caused by the mechanism boundary segmentation in the local subgraph after verification, compensation operations in the spatial domain and the spectral domain are performed respectively. The results of the compensation operations in the spatial domain and the spectral domain are then fused to form a dual-domain compensated graph substructure feature.

[0009] The dual-domain compensated graph substructure features are input into the local graph neural network model for sharpness-aware local training, and the enhanced local gradient is output. The enhanced local gradient is then encrypted for privacy protection and encapsulated into a privacy-protected encrypted gradient package.

[0010] The privacy-protected encrypted gradient packages uploaded by each participating institution are aggregated and decrypted using a secure aggregation protocol to obtain aggregated gradient plaintext. The aggregated gradient plaintext is then combined with the local subgraph digest to perform dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes to generate global graph neural network model parameters.

[0011] Based on the parameters of the global graph neural network model, inference and enhanced prediction are performed on the verified local subgraph, outputting multiple enhanced prediction values. The multiple enhanced prediction values ​​are then summarized, multi-source confidence weighted fusion is performed, and redundancy is removed to construct a cross-institutional research graph dataset.

[0012] Preferably, the step of mapping node attributes and edge relationships to local original scientific research data to construct a standardized local subgraph, performing graph structure verification on the standardized local subgraph using a shared ontology tree, and extracting a local subgraph summary to obtain a verified local subgraph and a local subgraph summary includes: mapping local heterogeneous node attributes to a unified semantic space to obtain standardized node attributes through a shared ontology tree pre-negotiated across institutions, mapping edge relationships to a standardized set of relationship types to obtain standardized edge relationships, constructing the standardized local subgraph based on the standardized node attributes and the standardized edge relationships; performing graph structure verification on the standardized local subgraph based on the shared ontology tree, identifying and correcting edges that violate ontology constraints, and obtaining the verified local subgraph; calculating subgraph topological statistics and feature distribution summaries for the verified local subgraph, performing differential privacy perturbation on the feature distribution summaries to obtain perturbed feature distribution summaries, and constructing the local subgraph summary from the subgraph topological statistics and the perturbed feature distribution summaries.

[0013] Preferably, the step of performing spatial and spectral domain compensation operations respectively on the graph signal loss caused by the mechanism boundary segmentation in the verified local subgraph, and fusing the results of the spatial and spectral domain compensation operations to form a dual-domain compensated graph substructure feature, includes: identifying boundary nodes in the verified local subgraph, wherein the boundary nodes are nodes connected to cross-mechanism edges; establishing virtual anchor nodes for each boundary node, obtaining the embedding representation of the corresponding position of the boundary node using the global graph neural network model parameters generated in the previous iteration, and using the embedding representation as the initial feature of the virtual anchor node, performing message passing aggregation based on the initial feature to obtain the spatial domain compensation result; calculating the adjacency matrix and degree matrix based on the verified local subgraph, and based on the adjacency matrix... A normalized Laplacian matrix is ​​constructed using the degree matrix and the given matrix. A graph Fourier transform is then performed on the spatial domain compensation result based on the normalized Laplacian matrix to obtain the frequency domain components. High-frequency components with corresponding eigenvalues ​​greater than a preset frequency threshold are smoothed, while low-frequency components with corresponding eigenvalues ​​less than or equal to the preset frequency threshold are gained to obtain the compensated frequency domain components. An inverse graph Fourier transform is then performed on the compensated frequency domain components to reconstruct node features, resulting in the spectral domain compensation result. The topology density is calculated for the verified local subgraph, and adaptive fusion weights are determined based on the topology density. The spatial domain compensation result and the spectral domain compensation result are then weighted and fused to form the dual-domain compensation graph substructure feature.

[0014] Preferably, the step of inputting the dual-domain compensated graph substructure features into a local graph neural network model for sharpness-aware local training, outputting enhanced local gradients, and performing privacy-preserving encryption processing on the enhanced local gradients to encapsulate them into a privacy-preserving encrypted gradient package includes: using a sharpness-aware minimization method, training the local graph neural network model with the dual-domain compensated graph substructure features as input, performing forward propagation to calculate the prediction loss, and performing backpropagation calculation based on the prediction loss to obtain an initial gradient; determining the perturbation direction of the parameters of the local graph neural network model based on the initial gradient, and adjusting the parameters of the local graph neural network model in the perturbation direction. After applying a perturbation of a preset step size, forward and backward propagation are performed again to calculate the perturbation gradient. The initial gradient and the perturbation gradient are weighted and fused to generate the enhanced local gradient. Gradient pruning is performed on the enhanced local gradient to bind sensitivity. Differential privacy noise variance is calculated based on a preset privacy budget and the sensitivity. Random noise that satisfies the differential privacy noise variance is generated and added to the pruned gradient to obtain the privacy-enhanced gradient. The privacy-enhanced gradient is encrypted using a preset homomorphic encryption public key to obtain gradient ciphertext. The gradient ciphertext and the identifier of the current iteration round are encapsulated into the privacy-protected encrypted gradient package.

[0015] Preferably, the step of aggregating and decrypting the privacy-preserving encrypted gradient packets uploaded by each participating institution using a secure aggregation protocol to obtain aggregated gradient plaintext; and performing dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes on the aggregated gradient plaintext in conjunction with the local subgraph digest to generate global graph neural network model parameters, includes: summing the encrypted gradient packets uploaded by each participating institution using a secure aggregation protocol to obtain aggregated ciphertext; decrypting the aggregated ciphertext to obtain the aggregated gradient plaintext, and obtaining the local subgraph digest uploaded by each participating institution, wherein the aggregated gradient plaintext includes aggregated gradient of graph neural network feature mapping layer parameters and aggregated gradient of graph neural network message passing layer parameters; the local subgraph digest includes perturbed feature distribution digest and subgraph topological statistics; and calculating the perturbed feature distribution digest... The semantic offsets of each participating institution corresponding to the feature distribution summary are used to compensate and correct the aggregated gradient of the graph neural network feature mapping layer parameters, resulting in the semantically aligned feature mapping layer parameter gradient. The structural topology differences of each participating institution are calculated based on the subgraph topology statistics, and the aggregated gradient of the graph neural network message passing layer parameters is scaled and aligned based on these structural topology differences, resulting in the structurally aligned message passing layer parameter gradient. The semantically aligned feature mapping layer parameter gradient and the structurally aligned message passing layer parameter gradient are fused to obtain the aligned global gradient. The adaptive aggregation weights of each participating institution are determined by combining the semantic offsets and the structural topology differences, and the aligned global gradient is weighted and aggregated based on these adaptive aggregation weights to generate the global graph neural network model parameters.

[0016] Preferably, the step of calculating the semantic offset of each participating institution corresponding to the perturbed feature distribution summary, and using the semantic offset to compensate and correct the aggregated gradient of the graph neural network feature mapping layer parameters to obtain the semantically aligned feature mapping layer parameter gradient includes: calculating the distribution distance between the perturbed feature distribution summary of each participating institution and the global feature distribution reference summary, and determining the distribution distance as the semantic offset, wherein the global feature distribution reference summary is generated by the coordination server by aggregating the perturbed feature distribution summaries uploaded by each participating institution; constructing a semantic compensation coefficient based on the semantic offset, and using the semantic compensation coefficient to scale and compensate the aggregated gradient of the graph neural network feature mapping layer parameters to obtain the semantically aligned feature mapping layer parameter gradient.

[0017] Preferably, the step of calculating the structural topology difference of each participating institution based on the subgraph topology statistics, and scaling and aligning the aggregated gradient of the graph neural network message passing layer parameters based on the structural topology difference to obtain the structure-aligned message passing layer parameter gradient includes: calculating the topology divergence between the subgraph topology statistics of each participating institution and the global topology statistical reference, and determining the topology divergence as the structural topology difference, wherein the global topology statistical reference is generated by the coordination server aggregating the subgraph topology statistics uploaded by each participating institution; constructing a structural scaling factor based on the structural topology difference, and using the structural scaling factor to scale and align the aggregated gradient of the graph neural network message passing layer parameters to obtain the structure-aligned message passing layer parameter gradient.

[0018] Preferably, the step of performing inference and augmentation prediction on the verified local subgraph based on the global graph neural network model parameters, outputting multiple augmentation prediction values, and summarizing, multi-source confidence weighted fusion, and redundancy removal of the multiple augmentation prediction values ​​to construct a cross-institutional research graph dataset includes: updating the local graph neural network model using the global graph neural network model parameters; performing different graph data augmentation operations on the verified local subgraph to generate an augmentation variant set; inputting each augmentation variant in the augmentation variant set into the updated local graph neural network model for inference, outputting the prediction probability distribution corresponding to each augmentation variant, and generating multiple augmentation prediction values ​​based on the prediction probability distribution; and generating multiple augmentation prediction values ​​based on the prediction probability distribution corresponding to each augmentation variant. The predicted probability distribution is calculated, the information entropy of each enhanced graph variant is calculated, and the multi-source confidence score of the enhanced prediction value corresponding to each enhanced graph variant is determined based on the information entropy. The multi-source confidence score is used as the fusion weight to perform weighted fusion of the multiple enhanced prediction values ​​to obtain an initial fusion prediction result. The feature similarity and structural overlap of each predicted entity node and relation edge in the initial fusion prediction result are calculated. Based on the feature similarity and structural overlap, redundant and conflicting prediction terms are identified and eliminated to obtain the target prediction result. Based on the node attribute prediction value and edge relation prediction value output in the target prediction result, the node attribute completion and edge relation expansion are performed on the verified local subgraph to construct the cross-institutional research graph dataset.

[0019] Preferably, the step of performing different graph data augmentation operations on the verified local subgraph to generate an augmentation variant set includes: performing at least two of the following operations on the verified local subgraph: node discarding, edge perturbation, feature masking, and subgraph sampling, to generate corresponding augmented graph variants to constitute the augmentation variant set; the node discarding operation is used to randomly remove some nodes and edges associated with some nodes in the verified local subgraph based on a preset discarding probability to generate a node discarding augmented graph variant; the edge perturbation operation includes an edge deletion sub-operation and an edge addition sub-operation, randomly deleting existing edges in the verified local subgraph based on the edge deletion sub-operation, and randomly adding new edges that do not exist in the verified local subgraph based on the edge addition sub-operation to generate an edge perturbation augmented graph variant; the feature masking operation is used to set the feature vectors of some nodes in the verified local subgraph to zero vectors based on a preset masking probability to generate a feature mask augmented graph variant; the subgraph sampling operation is used to extract locally connected subgraphs in the verified local subgraph based on a random walk algorithm to generate a subgraph sampling augmented graph variant.

[0020] Preferably, determining the multi-source confidence score of the enhanced prediction value corresponding to each enhanced graph variant based on the information entropy includes: transforming the information entropy of each enhanced graph variant using a negative correlation mapping function to obtain the initial confidence score of the enhanced prediction value corresponding to each enhanced graph variant, wherein the value of the information entropy is negatively correlated with the value of the initial confidence score; performing cross-variant normalization processing on the initial confidence score of the enhanced prediction value corresponding to each enhanced graph variant, and determining the result of the cross-variant normalization processing as the multi-source confidence score of the enhanced prediction value corresponding to each enhanced graph variant.

[0021] This invention also provides a cross-institutional collaboration system based on graph federation learning, comprising:

[0022] The data preprocessing module is used to map node attributes and edge relationships to local raw scientific research data to construct a standardized local subgraph. The shared ontology tree is used to perform graph structure verification on the standardized local subgraph and extract the local subgraph summary to obtain the verified local subgraph and the local subgraph summary.

[0023] The dual-domain compensation module is used to perform spatial domain and spectral domain compensation operations respectively for the graph signal loss caused by the mechanism boundary segmentation in the verified local subgraph, and to fuse the results of the spatial domain and spectral domain compensation operations to form a dual-domain compensated graph substructure feature.

[0024] The training and encryption module is used to input the dual-domain compensated graph substructure features into the local graph neural network model for sharpness-aware local training, output the enhanced local gradient, and perform privacy-preserving encryption processing on the enhanced local gradient, encapsulating it into a privacy-preserving encrypted gradient package.

[0025] The global aggregation module is used to aggregate and decrypt the privacy-protected encrypted gradient packages uploaded by each participating institution through a secure aggregation protocol to obtain aggregated gradient plaintext; and to perform dual-axis alignment and adaptive weighted aggregation on the aggregated gradient plaintext along the semantic axis and structural axis in combination with the local subgraph digest to generate global graph neural network model parameters.

[0026] The enhanced inference and graph construction module is used to perform inference and enhanced prediction on the verified local subgraph based on the parameters of the global graph neural network model, output multiple enhanced prediction values, and summarize the multiple enhanced prediction values, perform multi-source confidence weighted fusion and redundancy removal processing to construct a cross-institutional scientific research graph dataset.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. The present invention effectively repairs graph signal loss caused by cross-mechanism truncation edges through spatial-spectral dual-domain compensation: By compensating for cross-mechanism neighbor message passing of boundary nodes through virtual anchor nodes (spatial domain) and recovering the truncated low-frequency global topology signal (spectral domain) through graph Laplace spectral decomposition, and adaptively fusing with topology density, the cross-mechanism neighbor information of boundary nodes can be recovered. The message passing path coverage of subgraph federated learning is improved compared with the FedGraphNN baseline that only performs local message passing. Under the exemplary configuration, the low-frequency signal reconstruction error can be reduced to about one-third of that in the uncompensated case, thereby alleviating the problem of accuracy degradation in subgraph federated learning.

[0029] 2. This invention improves the quality and convergence speed of federated aggregation by heterogeneous perceptual semantic-structural dual-axis alignment: Since the Wasserstein distance of each institution's feature distribution is calculated based on differential privacy feature quantile summaries along the semantic axis and the gradient of the feature mapping layer is compensated, and the topological difference is calculated based on graph kernel similarity along the structural axis and the gradient of the message passing layer is scaled, the bias of the global model to the heterogeneous data distribution is corrected. Under the exemplary configuration, the semantic offset of each institution's feature distribution can be reduced by about 40% to 60% after compensation. The number of communication rounds required for federated convergence is reduced compared to simple weighted aggregation, and the generalization ability of the global model among heterogeneous participants is enhanced.

[0030] 3. The present invention suppresses local overfitting and enhances cross-mechanism generalization by sharpness-aware optimization: By introducing sharpness-aware minimization (SAM) into local training, each client is forced to update parameters along the flat direction of the loss surface. The local model no longer overfits its own special data distribution, the flatness of the loss surface of the global model on the cross-mechanism validation set is improved, and the cross-mechanism generalization error is reduced.

[0031] 4. The three-level privacy protection mechanism of this invention enables joint modeling without leaving the domain of the original data: By using differential privacy gradient perturbation to provide statistical privacy guarantee, homomorphic encryption to ensure computational privacy of transmission and aggregation, and secure aggregation to prevent the server from snooping on the unilateral gradient, the three complement each other to protect different attack surfaces. The original scientific research nodes and edge data are always kept locally, which meets the relevant requirements of data privacy protection and compliance supervision.

[0032] 5. This invention generates a high-quality dataset based on multi-source confidence weighted fusion of federated inference: By using the converged global model to generate enhanced predictions in local inference at each institution, and by using prediction confidence and topological consistency for quality assessment and multi-source weighted fusion, and by using a graph isomorphic network to remove redundancy, the final output cross-institutional scientific research graph dataset is much larger than that of a single institution, and provides higher prediction confidence through cross-institutional generalization ability and sharpness perception optimization. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is an overall flowchart of a cross-institutional collaboration method based on graph federation learning according to the present invention.

[0035] Figure 2 This is a flowchart of the graph substructure feature extraction method with spatial-spectral dual-domain compensation according to the present invention;

[0036] Figure 3 This is a flowchart of the heterogeneous perceptual semantic-structural dual-axis aligned federated aggregation method of the present invention;

[0037] Figure 4 This is a schematic diagram of the structure of a cross-institutional collaboration system based on graph federation learning according to the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0039] Example 1

[0040] like Figure 1 As shown, this invention provides a cross-institutional collaboration method based on graph federation learning, comprising:

[0041] Step S1: Map node attributes and edge relationships to the local original scientific research data to construct a standardized local subgraph. Use the shared ontology tree to perform graph structure verification on the standardized local subgraph and extract the local subgraph summary to obtain the verified local subgraph and the local subgraph summary.

[0042] To construct a standardized local subgraph for federated training, participating institutions (such as XX Hospital, XX Research Institute, and XX Pharmaceutical Company) first perform a unified mapping of node attributes and edge relationships on their locally held raw research data. Each institution's data includes a node attribute table (e.g., gene expression profiles, clinical indicators), an edge relationship table (e.g., protein-protein interactions, disease-target associations), and metadata recording data sources and acquisition protocols. Since the naming conventions, ontology definitions, and attribute dimensions used by different institutions are not uniform, direct concatenation would lead to semantic misalignment. Therefore, multiple institutions pre-negotiate a shared ontology tree to map heterogeneous local node attributes to a unified semantic space and edge relationships to a standardized set of relation types, thus constructing a standardized local subgraph. After standardization, the shared ontology tree is used to perform consistency checks on the standardized local subgraph, identifying and correcting edges that violate ontology constraints (e.g., illegal cross-type connections that are not allowed to be directly connected), resulting in a validated local subgraph. Finally, the topological features (node ​​degree distribution histogram, edge type distribution) of the verified local subgraph are statistically analyzed, and quantile sketches of each dimension attribute are extracted as feature distribution summaries. After differential privacy perturbation, the feature distribution summaries form a privacy-protected summary that can be publicly disclosed, which together with the un-de-sensitized local subgraph primitive information constitutes the local subgraph summary.

[0043] Step S2: For the graph signal loss caused by the mechanism boundary segmentation in the local subgraph after verification, perform compensation operations in the spatial domain and the spectral domain respectively, and fuse the results of the compensation operations in the spatial domain and the spectral domain to form a dual-domain compensated graph substructure feature.

[0044] When the global graph is segmented according to organizational boundaries, the edges connecting different organizations are truncated, causing boundary nodes to lose cross-organization neighbor information. The spectral structure of the subgraph also deviates from the complete global graph. These two types of losses fall into the spatial and spectral domains, respectively. To repair the spatial domain loss, virtual anchor nodes are established for boundary nodes connected to cross-organization edges. The corresponding embeddings from the previous round of the global model are used as the initial features of the anchor nodes to perform message passing, recovering the missing cross-organization neighbor messages. To repair the spectral domain loss, the normalized Laplacian matrix of the local subgraph is eigenvalued to obtain the graph Fourier transform basis. After projecting the node embeddings into the spectral domain, the truncated low-frequency global signal is amplified, and high-frequency noise is smoothed and suppressed. Then, an inverse transform is performed back to the node domain. Finally, the compensation results from the two domains are adaptively fused based on the topological density of the local subgraph: topologically sparse subgraphs rely more on spectral compensation, while topologically dense subgraphs rely more on spatial compensation. The spatial domain compensation, spectral domain compensation, and adaptive fusion are linked to form a dual-domain compensated graph substructure feature, which serves as the input for training the local model.

[0045] Step S3: Input the dual-domain compensated graph substructure features into the local graph neural network model for sharpness-aware local training, output the enhanced local gradient, and perform privacy-preserving encryption processing on the enhanced local gradient, encapsulating it into a privacy-preserving encrypted gradient package.

[0046] To complete model training locally and protect gradient privacy, step S3 feeds the dual-domain compensated graph substructure features into the local graph neural network model for training and encryption. Unlike ordinary training, sharpness-aware minimization is used here: first, the perturbation direction that makes the loss rise steepest is found in the neighborhood of the loss surface; the parameters are perturbed along this perturbation direction, and then the gradient is calculated and backpropagated, forcing the model to converge to a flat region, thereby suppressing overfitting to the local special data distribution and training to obtain enhanced local gradients. The enhanced local gradients then undergo three steps of processing in sequence: gradient clipping, differential privacy noise perturbation, and homomorphic encryption. Gradient clipping is bound to an upper sensitivity bound, differential privacy noise provides statistical privacy guarantees, and homomorphic encryption encapsulates the gradients into ciphertext. The three together ensure that the original data does not leave the local environment. After sharpness-aware training and three levels of privacy processing, the enhanced local gradients are encapsulated into a privacy-preserving encrypted gradient package, which can be uploaded to the federated coordination server.

[0047] Step S4: The privacy-protected encrypted gradient packets uploaded by each participating institution are aggregated and decrypted using a secure aggregation protocol to obtain aggregated gradient plaintext; combined with the local subgraph digest, the aggregated gradient plaintext is subjected to dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes to generate global graph neural network model parameters.

[0048] Each organization uploads its privacy-preserving encrypted gradient packets to the federal coordination server. Because the gradient packets are encapsulated using homomorphic encryption, the server can directly sum all the ciphertexts without decrypting the individual ciphertexts to obtain the aggregated ciphertext, and then decrypt the aggregated ciphertext to obtain the aggregated gradient plaintext. The existence of the secure aggregation protocol ensures that the server can only obtain the sum of the ciphertexts of all gradients, and cannot reconstruct the gradient plaintext of any single organization. The aggregated gradient plaintext can then be used for subsequent dual-axis alignment processing.

[0049] After obtaining the aggregated gradient plaintext and the local subgraph summaries of each institution, they are aligned along two axes. On the semantic axis, the offset between feature distributions is measured based on the perturbed feature distribution summaries of each institution, and the gradient of the feature mapping layer is compensated to make the features of each party converge towards a unified semantic space. On the structural axis, the differences in graph structure are measured based on the topological statistics of each institution, and the gradient of the message passing layer is scaled to make groups of institutions with similar topologies more coordinated in the aggregation. The gradients after the two axes are aligned are then weighted and aggregated according to the data size of each institution, the local loss, and the dynamically determined weights of the alignment deviation. The aggregation result is the parameter of the global graph neural network model, which is sent to each participant for the next round of local training.

[0050] Step S5: Based on the parameters of the global graph neural network model, perform inference and enhanced prediction on the verified local subgraph, output multiple enhanced prediction values, and summarize, multi-source confidence weighted fusion and redundancy removal of the multiple enhanced prediction values ​​to construct a cross-institutional scientific research graph dataset.

[0051] Each institution updates its local graph neural network model using the distributed global model parameters, performs inference on the validated local subgraph, and obtains enhanced predictions (e.g., potential disease-target associations). Since the confidence level of a single local inference is limited, this invention further applies multiple graph data augmentations to the local subgraph to obtain several augmentation variants. Inference is performed separately on each variant, and the confidence score of each prediction is determined based on the uncertainty (information entropy) of the inference result. Multi-source predictions are then weighted and fused according to their confidence levels, and redundant or conflicting predictions across institutions are removed. Finally, a larger and more reliable cross-institutional research graph dataset is compiled than that of any single institution.

[0052] Example 2

[0053] Based on Example 1, the process of mapping node attributes and edge relationships to local original scientific research data to construct a standardized local subgraph, performing graph structure verification on the standardized local subgraph using a shared ontology tree, and extracting a local subgraph summary to obtain the verified local subgraph and its summary includes:

[0054] Step S1.1: Through a shared ontology tree pre-negotiated across institutions, local heterogeneous node attributes are mapped to a unified semantic space to obtain standardized node attributes, and edge relationships are mapped to a standardized set of relationship types to obtain standardized edge relationships. Based on the standardized node attributes and the standardized edge relationships, the standardized local subgraph is constructed.

[0055] The local data preprocessing module receives local raw research data as input, which includes a node attribute table, an edge relation table, and metadata. Due to the distinct naming conventions, ontology definitions, and attribute dimensions developed by different institutions through long-term independent development, the same type of biological entity may use different local codes, different units of measurement, and even different attribute fields in different institutions. Directly concatenating these without unified processing will lead to semantic misalignment during cross-institutional collaboration. To eliminate semantic misalignment, the local data preprocessing module performs two types of mapping based on a pre-negotiated shared ontology tree across institutions: one is node attribute mapping, which aligns local heterogeneous node attributes to the unified semantic identifier registered in the shared ontology tree and normalizes the values ​​of each dimension to a unified unit of measurement according to semantic space conventions, resulting in standardized node attributes; the other is edge relation mapping, which aligns local edge relation names to a standardized set of relation types, resulting in standardized edge relations. For example, three institutions represent the same type of biological entity using local codes "gene_A", "protein_001", and "target_X" respectively. These are mapped to a unified semantic identifier "ENT:0001" via a shared ontology tree. The dimensions of each institution's expression are unified into standardized values ​​within the interval [0,1] through semantic space mapping. Edge relations "interacts_with" and "binds_to" are uniformly mapped to the standardized relation type "REL:PPI". A standardized local subgraph with consistent semantic scope is constructed using standardized node attributes as the node set and standardized edge relations as the edge set.

[0056] Step S1.2: Perform graph structure verification on the standardized local subgraph based on the shared ontology tree, identify and correct edges that violate ontology constraints, and obtain the verified local subgraph.

[0057] Graph structure verification is used to remove any illegal connections that may remain after ontology mapping, ensuring that the subgraph structure truly conforms to domain knowledge. Verification is based on two types of constraints registered in the shared ontology tree: type connectivity constraints, which specify which entity types are allowed to have edges; and relation domain constraints, which specify the legal first and last entity types for each relation type. The local data preprocessing module checks each edge in the standardized local subgraph. Edges whose first and last entity types conflict with ontology constraints are identified as violating ontology constraints and corrected according to the nature of the conflict: obviously illegal connections are directly deleted, and edges mistakenly connected due to encoding ambiguity are redirected to the correct entity. For example, a direct connection between a "disease" type node and a "gene" type node is not allowed; such cross-type connections in the standardized local subgraph will be identified and deleted. Similarly, if an edge mistakenly connects "drug" to "clinical indicator" due to misaligned source data fields, it will be redirected to the target entity allowed by the ontology constraints. After verification and correction, the standardized local subgraph no longer contains edges violating ontology constraints, becoming a structurally consistent verified local subgraph.

[0058] Step S1.3: Calculate the subgraph topology statistics and feature distribution summary for the verified local subgraph, perform differential privacy perturbation on the feature distribution summary to obtain the perturbed feature distribution summary, and construct the local subgraph summary from the subgraph topology statistics and the perturbed feature distribution summary.

[0059] The local subgraph summary provides the necessary distribution and structural information for subsequent alignment to the coordination server without requiring the original data to leave the local machine. Therefore, the summary must be both statistically representative and privacy-preserving. The summary consists of two parts: first, subgraph topological statistics, including a histogram of node degree distribution and edge type distribution, used to characterize the density and relationships of the subgraph's connections; second, a feature distribution summary, i.e., a quantile sketch extracted from each dimension's attributes, used to characterize the overall distribution of feature values. In this embodiment, the quantile sketch uses a nine-quartile approximation, taking the minimum value, the 10th to 90th percentiles (step size 10), and the maximum value—a total of nine statistics—to summarize the distribution with minimal communication overhead. To prevent the quantile sketch from being used to deduce the original feature values, a differential privacy perturbation is applied to the feature distribution summary: a Laplace mechanism is used, injecting random noise following a Laplace distribution into each statistic. The perturbation scale increases with the summary sensitivity and decreases with the privacy budget. The perturbation-processed feature distribution summary no longer leaks information about individual records, and together with the subgraph topology statistics, it constitutes a local subgraph summary that can be securely uploaded.

[0060] Example 3

[0061] like Figure 2As shown, based on Example 1, for the graph signal loss caused by the mechanism boundary segmentation in the verified local subgraph, spatial domain and spectral domain compensation operations are performed respectively. The results of the spatial domain and spectral domain compensation operations are fused to form a dual-domain compensated graph substructure feature, including:

[0062] Step S2.1: Identify the boundary nodes in the verified local subgraph, where the boundary nodes are nodes connected to cross-mechanism edges.

[0063] The dual-domain compensation module receives the verified local subgraph and its summary as input. When the complete global graph is divided into local subgraphs according to organizational boundaries, edges connecting nodes of different organizations are cut. Nodes at the cut-off points lose their cross-organization neighbors that should have participated in message passing; these nodes are called boundary nodes. Identifying boundary nodes is a prerequisite for dual-domain compensation: the local subgraph metadata records cut-edge detection information, indicating which nodes in the local subgraph have edges pointing to other organizational subgraphs in the global graph. The dual-domain compensation module compares each node according to the cut-edge detection information. Nodes with cross-organization neighbors in the global graph but whose corresponding edges have been cut in the local subgraph are marked as boundary nodes. It also records the positions of the cross-organization neighbors that each boundary node should have connected to in the global graph, for reconstructing the message source during spatial domain compensation. For example, if an organization's local subgraph has 1000 nodes, comparison reveals that approximately 120 nodes have cut cross-organization edges in the global graph; these 120 nodes are identified as boundary nodes.

[0064] Step S2.2: Establish a virtual anchor node for each of the boundary nodes, obtain the embedding representation of the corresponding position of the boundary node using the global graph neural network model parameters generated in the previous iteration, and use the embedding representation as the initial feature of the virtual anchor node. Based on the initial feature, perform message passing aggregation to obtain the spatial domain compensation result.

[0065] Each participating institution is distinguished by a number k, where k ranges from 1 to N, and N is the total number of participating institutions. The core of spatial domain compensation lies in reintroducing the severed cross-institutional neighbor information into the local subgraph in a virtual form, thereby repairing the missing message passing at boundary nodes. A virtual anchor node is established for each boundary node, acting as a local substitute for the severed cross-institutional neighbor. The initial features of the virtual anchor node come from the parameters of the global graph neural network model generated in the previous iteration: the global model from the previous iteration is used to infer the corresponding position of the boundary node in the global graph, obtaining the embedding representation of the corresponding position of the boundary node, and assigning the embedding representation to the virtual anchor node as its initial feature, enabling the virtual anchor node to carry neighbor information from a global perspective. After the virtual anchor node is reconnected to the local subgraph according to the original connection relationship of the severed edges, message passing aggregation is performed within the local subgraph: when aggregating neighborhood information, the boundary node receives messages from both its local real neighbors and cross-institutional neighbor messages from the virtual anchor node, thus compensating for the missing cross-institutional neighbor information. The node embedding obtained from message passing aggregation is the spatial compensation node embedding, denoted as the embedding of the kth institution. , as input for spectral domain compensation.

[0066] Step S2.3: Calculate the adjacency matrix and degree matrix based on the verified local subgraph, construct a normalized Laplace matrix based on the adjacency matrix and the degree matrix, and perform a graph Fourier transform on the spatial domain compensation result based on the normalized Laplace matrix to obtain the frequency domain components.

[0067] Spatial domain compensation alone is insufficient to recover the damage to the global topological pattern caused by subgraph partitioning. Spectral domain compensation further repairs the loss from the perspective of graph signal frequency. Therefore, it is necessary to first establish a graph Fourier analysis foundation for the local subgraph. Based on the verified local subgraph, the adjacency matrix A and the degree matrix D are calculated (the diagonal elements of the degree matrix D are equal to the sum of the corresponding row elements of the adjacency matrix A), and a symmetric normalized Laplace matrix is ​​constructed accordingly (the product of the inverse square root of the degree matrix and the adjacency matrix minus the identity matrix). The symmetric normalized Laplace matrix is ​​a symmetric positive semi-definite matrix, which can be decomposed into an orthogonal matrix composed of eigenvectors and a diagonal matrix of eigenvalues. The magnitude of the eigenvalues ​​corresponds to the frequency of the graph signal: the smaller the eigenvalue, the lower the frequency, reflecting a global, smooth topological pattern; the larger the eigenvalue, the higher the frequency, reflecting local, abrupt structural details. The column vectors of the resulting orthogonal matrix are the graph Fourier transform basis. Because they are orthogonal, their transpose is equal to the inverse matrix, thus also serving as the inverse transform basis. The spatial compensation nodes are embedded and projected onto the spectral basis using the graph Fourier transform basis. The graph Fourier transform is then performed to obtain the frequency domain components, which prepares for subsequent compensation by frequency band.

[0068] Step S2.4: Perform a smoothing operation on the high-frequency components whose corresponding feature values ​​are greater than the preset frequency threshold in the frequency domain components, and perform a gain operation on the low-frequency components whose corresponding feature values ​​are less than or equal to the preset frequency threshold in the frequency domain components to obtain the compensated frequency domain components; perform a graph inverse Fourier transform on the compensated frequency domain components to reconstruct the node features to obtain the spectrum domain compensation result.

[0069] Subgraph segmentation primarily truncates low-frequency signals reflecting the global topology pattern. Simultaneously, the segmentation boundaries introduce high-frequency noise that is not part of the actual structure. Therefore, frequency domain components need to be processed differently according to frequency segments. Frequency bands are divided using a preset frequency threshold (0.5 in this embodiment): For low-frequency components with eigenvalues ​​less than or equal to the preset threshold, a gain coefficient greater than 1 is applied to enhance the truncated global low-frequency signal; for high-frequency components with eigenvalues ​​greater than the preset threshold, Tikhonov regularization smoothing is applied, i.e., multiplied by an attenuation coefficient between 0 and 1, to suppress the high-frequency noise introduced by segmentation, resulting in compensated frequency domain components that balance global signal enhancement and noise suppression. The compensated frequency domain components are still expressed on the spectral basis and need to be returned to the node domain via inverse graph Fourier transform: the compensated frequency domain components are left-multiplied by the graph Fourier transform basis to reconstruct node features, forming a spectral compensation node embedding. The spectral compensation node embedding restores global topology information in the low-frequency band and filters out segmentation noise in the high-frequency band, yielding the spectral domain compensation result.

[0070] Step S2.5: Calculate the topological density of the verified local subgraph, determine the adaptive fusion weight based on the topological density, and perform weighted fusion of the spatial domain compensation result and the spectral domain compensation result to form the dual-domain compensation graph substructure feature.

[0071] Spatial compensation node embedding focuses on recovering local cross-institutional neighbor messages, while spectral compensation node embedding focuses on recovering global low-frequency topological signals. The relative importance of the two types of compensation varies with the density of the subgraph itself; therefore, the fusion weights are adaptively determined by the topological density. The topological density is calculated for the verified local subgraph, defined as the number of edges divided by the number of nodes, denoted as... The spatial domain fusion weights are obtained by monotonically mapping the difference between the topological density and a preset reference density threshold using the Sigmoid function. The spectral domain fusion weights are the complements of the spatial domain fusion weights.

[0072]

[0073] in, For spatial domain fusion weights, For spectral domain fusion weights, For the Sigmoid function, This is the fusion sensitivity coefficient (1.0 in the example). The topological density (number of edges divided by number of nodes) of the local subgraph of the k-th institution. This is a preset reference density threshold (2.0 in the example).

[0074] Sparse subgraphs have few local neighbors and insufficient local messages, so fusion is biased towards spectral compensation to utilize global low-frequency signals; dense subgraphs have sufficient local neighbors, so fusion is biased towards spatial compensation to preserve local message passing. The substructure features of a dual-domain compensated graph are fused using the following formula:

[0075]

[0076] in, The substructure features of the two-domain compensation graph for the k-th mechanism are as follows: Embed the spatial compensation node for the k-th institution. Embed the spectrum compensation node for the k-th institution. For spatial domain fusion weights, For spectral domain fusion weights.

[0077] Assuming a subgraph of an organization has 1000 nodes and 3000 edges, its topology density is 3.0, the reference density threshold is 2.0, and the fusion sensitivity coefficient is 1.0. The calculated spatial domain fusion weight is approximately 0.731, indicating a fusion bias towards spatial compensation. For another organization with a topology density of 1.0, the spatial domain fusion weight is approximately 0.269, indicating a fusion bias towards spectral compensation. The spatial compensation, spectral compensation, and adaptive fusion are interconnected, resulting in a dual-domain compensated graph substructure feature that simultaneously includes cross-organization neighbor information and global low-frequency signals, forming the aforementioned dual-domain compensated graph substructure feature.

[0078] Example 4

[0079] Based on Example 1, the step of inputting the dual-domain compensated graph substructure features into a local graph neural network model for sharpness-aware local training, outputting enhanced local gradients, and performing privacy-preserving encryption processing on the enhanced local gradients, encapsulating them into a privacy-preserving encrypted gradient package, includes:

[0080] Step S3.1: Using the sharpness-aware minimization method, the local graph neural network model is trained with the dual-domain compensated graph substructure features as input, forward propagation is performed to calculate the prediction loss, and backpropagation is performed based on the prediction loss to obtain the initial gradient.

[0081] This map neural network model employs a Graph Convolutional Network (GCN). Each layer of the GCN performs symmetric normalization aggregation on the node embeddings of the previous layer using an adjacency matrix with added self-loops and the corresponding degree matrix. This is followed by linear transformation of the learnable weight matrix and non-linear activation to obtain the node embeddings for the current layer (the initial layer node embeddings are derived from dual-domain compensated graph substructure features). The GCN model contains two layers with a hidden dimension of 128, and the output dimension equals the number of task categories (disease-target association prediction uses two categories). In this embodiment, the training samples are the graph substructure features and corresponding annotations obtained from the local subgraphs of each institution through dual-domain compensation. In an exemplary configuration, a local subgraph of an institution contains approximately 12,000 nodes and 45,000 edges, divided into training, validation, and test sets in an 8:1:1 ratio. The training process uses the Adam optimizer with the following hyperparameters: an initial learning rate of 1×10⁻⁶. -3 The weight decay coefficient is 5×10 -4 The batch size was 256 (using node-based mini-batch training), the training epochs were 100, the random dropout rate was 0.5, and the learning rate was scheduled using cosine annealing. The loss function used was cross-entropy loss, which measures the difference between the model's predicted probability distribution and the true label distribution. After calculating the prediction loss through forward propagation, the initial gradient was calculated through backpropagation based on the prediction loss. .

[0082] Step S3.2: Determine the perturbation direction of the parameters of the local graph neural network model based on the initial gradient. After applying a perturbation of a preset step size to the parameters of the local graph neural network model in the perturbation direction, perform forward propagation and backward propagation again to calculate the perturbation gradient.

[0083] Employing a sharpness-aware minimization method, for the current local model parameters, we first search within the neighborhood for the perturbation direction that causes the loss to increase most steeply: based on the initial gradient. Determine the perturbation direction, and apply a preset step size in the perturbation direction to obtain the perturbation vector:

[0084]

[0085] in, Let be the perturbation vector. The preset step size (perturbation radius, 0.05 in the example) is used. For the initial gradient, Let L2 be the initial gradient norm.

[0086] The perturbation vector is applied to the model parameters to obtain the perturbed parameters. Then, forward and backward propagation are performed again to obtain the perturbed gradient. .

[0087] Step S3.3: Weighted fusion of the initial gradient and the perturbation gradient to generate the enhanced local gradient.

[0088] The initial gradient and the perturbation gradient are weighted and fused to generate a sharpness-aware enhanced local gradient. The fusion uses a convex combination method to merge the two gradients element-wise, allowing the enhanced local gradient to retain the standard descent direction while absorbing information about flat regions carried by the perturbation direction.

[0089]

[0090] in, For the enhanced local gradient generated by weighted fusion of the k-th institution, For the initial gradient, α is the perturbation gradient, and α is the fusion weight (ranging from 0 to 1; the standard sharpness perception is minimized to 0, meaning only the perturbation gradient is used; in this embodiment, it is 0.5).

[0091] In this embodiment, the fusion weight α is set to 0.5, which means that the initial gradient and the perturbation gradient participate with equal weight. The two gradients are in the same parameter space and can be merged element by element to generate the enhanced local gradient.

[0092] Step S3.4: Perform gradient pruning on the enhanced local gradient to bind sensitivity, calculate differential privacy noise variance based on the preset privacy budget and the sensitivity, generate random noise that satisfies the differential privacy noise variance and add it to the pruned gradient to obtain the privacy-enhanced gradient.

[0093] Gradient pruning is performed on the enhanced local gradient to bind sensitivity: pruning is performed according to a gradient norm pruning threshold C (the upper bound of sensitivity, set to 1.0 in this embodiment). If the L2 norm of the enhanced local gradient exceeds the threshold C, it is scaled down to an L2 norm equal to C, resulting in the pruned gradient. Differential privacy noise employs a Gaussian mechanism. The noise standard deviation is jointly calibrated by the pruning threshold C, the privacy budget, and the privacy default probability. In this embodiment, C is set to 1.0, the privacy budget to 2.0, and the privacy default probability to 1 × 10⁻⁶. -5 The noise standard deviation, as determined by this standard deviation, is approximately 2.42. Gaussian random noise with a mean of 0 and a covariance determined by this standard deviation is generated and added to the clipped gradient to obtain the privacy-enhancing gradient. Differential privacy noise provides statistical privacy guarantees for the gradient itself.

[0094] Step S3.5: Encrypt the privacy-enhancing gradient using a preset homomorphic encryption public key to obtain gradient ciphertext, and encapsulate the gradient ciphertext and the identifier of the current iteration round into the privacy-protected encrypted gradient package.

[0095] The privacy-enhancing gradient is encrypted using a pre-defined homomorphic encryption public key pk to obtain gradient ciphertext. In this embodiment, the homomorphic encryption uses the Paillier additive homomorphic encryption scheme, and the process is as follows: Before training begins, a public-private key pair is generated, and the public key pk is distributed to each participating institution. The private key sk paired with the public key pk is stored by the federated coordination server. Each institution quantizes each component of the privacy-enhancing gradient into an integer and then encrypts it component by component using the public key pk to obtain the gradient ciphertext. The Paillier ciphertext satisfies the additive homomorphic property, meaning that multiplication in the ciphertext field is equivalent to addition in the corresponding plaintext field. This property allows the gradient ciphertexts of multiple institutions to be directly summed without decryption, which is the basis for ciphertext summation of the gradient ciphertexts of each institution in subsequent secure aggregation. The gradient ciphertext is then encapsulated together with the identifier of the current iteration round to form a privacy-preserving encrypted gradient package.

[0096] Example 5

[0097] like Figure 3 As shown, based on Example 1, the privacy-preserving encrypted gradient packets uploaded by each participating institution are aggregated and decrypted using a secure aggregation protocol to obtain aggregated gradient plaintext; combined with the local subgraph digest, the aggregated gradient plaintext is subjected to dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes to generate global graph neural network model parameters, including:

[0098] Step S4.1: The privacy-protected encrypted gradient packets uploaded by each participating institution are summed using a secure aggregation protocol to obtain aggregated ciphertext.

[0099] Each participating organization uploads its privacy-preserving encrypted gradient packets to the federal coordination server. Leveraging the additive homomorphic property of homomorphic encryption, the server sums the ciphertexts of the gradients uploaded by each party to obtain the aggregated ciphertext. The secure aggregation protocol ensures that the server can only obtain the ciphertext sum of all gradients and cannot reconstruct the plaintext gradient of any single organization.

[0100] Step S4.2: Decrypt the aggregated ciphertext to obtain the aggregated gradient plaintext, and obtain the local subgraph digest uploaded by each participating institution. The aggregated gradient plaintext includes the aggregated gradient of the graph neural network feature mapping layer parameters and the aggregated gradient of the graph neural network message passing layer parameters. The local subgraph digest includes the perturbed feature distribution digest and the subgraph topology statistics.

[0101] The aggregated ciphertext is decrypted to obtain the aggregated gradient plaintext (equal to the sum of the privacy-enhancing gradients of each party). The coordination server uses a private key paired with the homomorphic encryption public key to complete the decryption. The decryption result only contains the aggregated values ​​of the parameters of each layer, without exposing the original gradients of any single institution. Simultaneously, the local subgraph digests uploaded by each participating institution are obtained. The aggregated gradient plaintext includes the aggregated gradients of the graph neural network feature mapping layer parameters and the graph neural network message passing layer parameters; the local subgraph digests include the perturbed feature distribution digest and the subgraph topological statistics.

[0102] Step S4.3: Calculate the semantic offset of each participating mechanism corresponding to the perturbed feature distribution summary, and use the semantic offset to compensate and correct the aggregate gradient of the graph neural network feature mapping layer parameters to obtain the semantically aligned feature mapping layer parameter gradient.

[0103] The coordination server receives perturbed feature distribution summaries uploaded by participating institutions. It first constructs a global feature distribution reference summary by taking a weighted average of the corresponding quantiles of each institution's summary. Then, it compares the difference between each institution's summary and the global reference summary at each quantile to quantify the semantic offset of the corresponding institution. A larger semantic offset indicates that the feature distribution of the corresponding institution deviates further from the global semantic space. Subsequently, based on the semantic compensation coefficient, the aggregate gradient of the feature mapping layer parameters in the aggregate gradient plaintext of the corresponding institution is scaled and compensated: institutions with large offsets have smaller compensation coefficients, and their feature mapping layer gradients are correspondingly enhanced to align with the global semantic space; institutions with small offsets retain their original gradients. After compensation, the deviations of each institution at the feature semantic level are pre-smoothed before aggregation, and the feature mapping layer gradients are aligned with the global semantic space, resulting in semantically aligned feature mapping layer parameter gradients.

[0104] Step S4.4: Calculate the structural topology differences of each participating mechanism based on the subgraph topology statistics, and scale and align the aggregate gradient of the message passing layer parameters of the graph neural network according to the structural topology differences to obtain the structurally aligned message passing layer parameter gradients.

[0105] The coordination server receives subgraph topology statistics (node ​​degree distribution histogram and edge type distribution) uploaded by each participating institution. First, it aggregates the topology statistics of all institutions to construct a global topology statistical reference. Then, it measures the difference between each institution's subgraph and the global reference structure based on graph kernel similarity, quantifying the structural topology differences of the corresponding institutions. The greater the structural topology difference, the more unique the graph topology of the corresponding institution and the less similar it is to other institutions. Subsequently, based on the structure scaling factor, it scales and aligns the message-passing layer parameter aggregation gradient in the plaintext of the aggregation gradient of the corresponding institution: topologically similar institutions receive more coordinated weights in the aggregation, while topologically unique institutions are correspondingly suppressed. The message-passing layer gradient is thus aligned with the global topology structure, and heterogeneous structural deviations are absorbed before aggregation, resulting in the structure-aligned message-passing layer parameter gradient.

[0106] Step S4.5: Fuse the gradients of the semantically aligned feature mapping layer parameters with the gradients of the structurally aligned message passing layer parameters to obtain the aligned global gradient.

[0107] The feature mapping layer and message passing layer of the graph convolutional network have two different sets of parameters, corresponding to node feature transformation and neighborhood aggregation, respectively. The gradients of the semantically aligned feature mapping layer parameters and the gradients of the structurally aligned message passing layer parameters are concatenated and merged according to their corresponding parameter layers to obtain the aligned global gradient of the k-th structure. This aligned global gradient corrects the heterogeneity deviation in both feature semantics and topology, thus obtaining the aligned global gradient.

[0108] Step S4.6: Combine the semantic offset and the structural topology difference to determine the adaptive aggregation weight of each participating mechanism, and perform weighted aggregation of the aligned global gradient based on the adaptive aggregation weight to generate the global graph neural network model parameters.

[0109] By combining semantic offsets, structural topology differences, and the data scale and local model training loss of each institution, adaptive aggregation weights for each participating institution are determined (each weight is non-negative and their sum is 1):

[0110]

[0111] in, For the adaptive aggregation weight of the k-th institution, Let be the data size of the k-th institution (in terms of the number of nodes), N be the total number of participating institutions, γ be the alignment deviation sensitivity coefficient (0.5 in the example), and λ be the local loss sensitivity coefficient (0.1 in the example). The training loss for the local model of the k-th institution is... Let be the semantic offset of the k-th institution. Let represent the structural topological difference of the k-th institution, and j be the participating institution number, j=1,2,…,N.

[0112] Organizations with large data scales and small distribution offsets receive higher weights. Based on adaptive aggregation weights, the global gradient is aligned and weighted, and the global model parameters from the previous round are updated according to the aggregation learning rate (0.1 in the example). This generates the global model parameters for the current round and distributes them to each participant for local training in the next round, thus generating the global graph neural network model parameters.

[0113] Example 6

[0114] Based on Example 5, the step of calculating the semantic offset of each participating mechanism corresponding to the perturbed feature distribution summary, and using the semantic offset to compensate and correct the aggregate gradient of the graph neural network feature mapping layer parameters to obtain the semantically aligned feature mapping layer parameter gradient, includes:

[0115] Step S4.3.1: Calculate the distribution distance between the perturbed feature distribution summary of each participating institution and the global feature distribution reference summary, and determine the distribution distance as the semantic offset. The global feature distribution reference summary is generated by the coordination server by aggregating the perturbed feature distribution summaries uploaded by each participating institution.

[0116] The global aggregation module calculates the semantic offset based on the perturbed feature distribution summaries (quantile sketches) uploaded by each institution. The global feature distribution reference summary is generated by the coordination server by taking the weighted average of the quantiles of the perturbed feature distribution summaries uploaded by each institution. The semantic offset of the k-th institution from the global reference in the feature distribution is approximated by the mean absolute deviation of the quantile sketches to approximate the first-order Wasserstein distance (Wasserstein distance, also known as bulldozer distance or Earth-Mover distance, measures the transmission cost between two distributions):

[0117]

[0118] in, Let Q be the semantic offset of the k-th institution, and let Q be the number of statistics in the quantile sketch (9 in the example). The q-th quantile of the summary of the distribution of the characteristics of the k-th institution. Let q be the q-th quantile of the global feature distribution reference summary, where q is the quantile index, and q = 1, 2, ..., Q.

[0119] This process relies solely on quantile summaries protected by differential privacy, without revealing the original feature values. Numerical example: The differences between an institution and the global reference at the nine quantiles are 0.05, 0.08, 0.02, 0.10, 0.03, 0.07, 0.04, 0.06, and 0.05, respectively. Under uniform weighting, the semantic offset is approximately 0.0556.

[0120] Step S4.3.2: Construct semantic compensation coefficients based on the semantic offset, and use the semantic compensation coefficients to scale and compensate the gradient of the aggregated gradient of the feature mapping layer parameters of the graph neural network to obtain the semantically aligned gradient of the feature mapping layer parameters.

[0121] A semantic compensation coefficient is constructed based on the semantic offset. The larger the semantic offset and the smaller the compensation coefficient, the stronger the scaling compensation for the gradient of the feature mapping layer.

[0122]

[0123] in, Let η be the semantic compensation coefficient for the k-th institution, and η be the compensation strength coefficient (10.0 in the example). This is the semantic offset of the k-th institution.

[0124] In one implementation, the aggregate gradient of the graph neural network feature mapping layer parameters of the k-th mechanism is scaled and compensated using a compensation coefficient:

[0125]

[0126] in, The gradient of the feature map layer parameters after semantic alignment of the k-th institution. Aggregate gradients for the parameters of the k-th institutional feature mapping layer. is the semantic compensation coefficient for the k-th institution.

[0127] In the example above, the semantic offset is approximately 0.0556 and the compensation intensity coefficient is 10.0, so the semantic compensation coefficient is approximately 0.573. The feature mapping layer gradient is scaled and compensated according to this coefficient and then moved closer to the global semantic space to obtain the semantically aligned feature mapping layer parameter gradient.

[0128] In a preferred implementation, a distribution alignment transformation matrix can also be constructed along the semantic axis. The gradient of the feature mapping layer of the k-th institution is projected onto a unified semantic space and then a semantic compensation coefficient is applied. The transformation matrix is ​​obtained by performing principal component alignment (PCA-whitening) on ​​the feature distribution summaries of each institution, so that the semantic offset of different institutions is minimized after projection.

[0129] Example 7

[0130] Based on Example 5, the step of calculating the structural topology differences of each participating mechanism according to the subgraph topology statistics, and scaling and aligning the aggregated gradient of the graph neural network message passing layer parameters according to the structural topology differences to obtain the structurally aligned message passing layer parameter gradients includes:

[0131] Step S4.4.1: Calculate the topological divergence between the subgraph topological statistics of each participating institution and the global topological statistical reference, and determine the topological divergence as the structural topological difference. The global topological statistical reference is generated by the coordination server by aggregating the subgraph topological statistics uploaded by each participating institution.

[0132] The global aggregation module calculates the structural topology differences based on the subgraph topology statistics (including node degree distribution histograms and edge type distributions) uploaded by each institution. The global topology statistical reference is generated by the coordination server aggregating the subgraph topology statistics uploaded by each institution. The graph kernel similarity function (in this embodiment, the Weisfeiler-Lehman kernel, with values ​​between 0 and 1) is used to measure the kernel similarity between the k-th institution and all other institutions. The structural topology difference of the k-th institution is calculated using the following formula:

[0133]

[0134] in, For the structural topological differences of the k-th mechanism, Let N be the kernel similarity between the k-th subgraph and the j-th subgraph (using the Weisfeiler-Lehman kernel, with a value between 0 and 1), where N is the total number of participating institutions, and N-1 is the number of participating institutions other than the k-th institution.

[0135] Numerical example: The kernel similarity between the k-th mechanism and the other 4 mechanisms (N=5 in total) are 0.82, 0.65, 0.90 and 0.71 respectively. The average similarity is 0.77 and the structural topological difference is 1−0.77=0.23.

[0136] Step S4.4.2: Construct a structural scaling factor based on the structural topology differences, and use the structural scaling factor to scale and align the aggregate gradient of the message passing layer parameters of the graph neural network to obtain the structurally aligned gradient of the message passing layer parameters.

[0137]

[0138] in, Let θ be the structural scaling factor for the k-th mechanism, and θ be the structural alignment strength coefficient (2.0 in the example). The structural topology difference of the k-th mechanism.

[0139] Furthermore, the aggregate gradients of the graph neural network message-passing layer parameters of the k-th mechanism are scaled and aligned using a scaling factor:

[0140]

[0141] in, The gradient of the message passing layer parameters after alignment of the k-th structure. Aggregate gradients for the parameters of the k-th mechanism message passing layer. is the structural scaling factor for the k-th mechanism.

[0142] Example 8

[0143] Based on Example 1, the inference and augmentation prediction based on the global graph neural network model parameters are performed on the verified local subgraph, outputting multiple augmentation prediction values. These multiple augmentation prediction values ​​are then aggregated, multi-source confidence weighted fusion is performed, and redundancy is removed to construct a cross-institutional research graph dataset, including:

[0144] Step S5.1: Update the local graph neural network model using the parameters of the global graph neural network model, and perform different graph data augmentation operations on the verified local subgraph to generate an augmentation variant set.

[0145] After updating their local graph neural network models with global model parameters, each institution applies various graph data augmentation operations to the validated local subgraphs, generating several augmentation variants. Single inference is sensitive to minor perturbations in the subgraph structure, and predictions from a single view are subject to chance. By constructing augmentation variants from multiple perspectives through at least two methods, such as node dropping, edge perturbation, feature masking, and subgraph sampling, the association patterns of the same scientific object can be characterized from different angles, reducing the random error of single-view predictions.

[0146] Step S5.2: Input each augmented graph variant in the augmented variant set into the updated local graph neural network model for inference, output the prediction probability distribution corresponding to each augmented graph variant, and generate multiple augmented prediction values ​​based on the prediction probability distribution.

[0147] Each augmented graph variant in the augmented variant set is input into the updated local graph neural network model for inference, outputting the predicted probability distribution corresponding to each augmented graph variant, and generating multiple augmented prediction values ​​(such as potential disease-target association and association category probability) based on the predicted probability distribution. Inference example: Taking a candidate disease-target association as an example, the corresponding augmented graph variant is input into the updated GCN model. After two layers of message passing, the output layer obtains the probability of each category through the normalized exponential function (Softmax). Let the output logit be z1=2.1 and z2=-0.3, then the positive class probability is about 0.91 and the negative class probability is about 0.09. The association is determined to be positive (association exists), with a confidence level of 0.91, which is used as an augmented prediction value. Inference is performed on each augmented graph variant one by one, producing the corresponding predicted probability distribution and multiple augmented prediction values ​​respectively.

[0148] Step S5.3: Based on the predicted probability distribution corresponding to each augmented graph variant, calculate the information entropy of each augmented graph variant, and determine the multi-source confidence score of the augmented prediction value corresponding to each augmented graph variant according to the information entropy.

[0149] Predictions from a single augmentation variant may be distorted due to local structural perturbations, necessitating a measurement of the reliability of each prediction. Based on the prediction probability distribution corresponding to each augmentation variant, the information entropy of the corresponding augmentation variant is calculated to measure the uncertainty of the prediction (the specific calculation of information entropy and the determination of multi-source confidence scores are described in Example 11): the larger the information entropy, the more dispersed and uncertain the predictions are across categories; the smaller the information entropy, the more concentrated and reliable the predictions are in a particular category. For example, one augmentation variant outputs a positive class probability of 0.91 and a negative class probability of 0.09, with an information entropy of approximately 0.39, corresponding to a higher confidence level; another variant outputs a positive class probability of 0.55 and a negative class probability of 0.45, with an information entropy of approximately 0.69, corresponding to a lower confidence level. Based on the negative correlation between information entropy and confidence, the information entropy of each variant is mapped to its respective confidence weight to determine the multi-source confidence score of the augmentation prediction value for each augmentation variant.

[0150] Step S5.4: Use the multi-source confidence score as the fusion weight to perform weighted fusion of the multiple enhanced prediction values ​​to obtain the initial fusion prediction result.

[0151] The multi-source confidence scores are used as fusion weights to perform weighted fusion of the multiple enhanced prediction values, resulting in an initial fusion prediction result. This ensures that the predictions of high-confidence variants have a greater weight in the final fusion. Taking a candidate disease-target association as an example: the node dropout variant gives a positive probability of 0.88 and a confidence score of 0.55, while the edge perturbation variant gives a positive probability of 0.94 and a confidence score of 0.45. The weighted average of these two variants yields a fusion positive probability of 0.88 × 0.55 + 0.94 × 0.45 ≈ 0.907, which is the initial fusion prediction value for this candidate association. The prediction values ​​of the same candidate entity or relationship under different enhanced variants are all weighted averaged according to their corresponding confidence scores to obtain the initial fusion prediction result.

[0152] Step S5.5: Calculate the feature similarity and structural overlap between each predicted entity node and relation edge in the initial fusion prediction result. Based on the feature similarity and structural overlap, identify and remove redundant and conflicting prediction terms to obtain the target prediction result.

[0153] The initial fusion prediction results often contain duplicate or contradictory predictions across institutions. The feature similarity and structural overlap of each predicted entity node and relation edge in the initial fusion prediction results are calculated: for multiple predictions pointing to the same entity or the same association, the one with the highest confidence is retained, and other duplicates are removed; prediction edges that are determined to be contradictory based on known topological patterns (e.g., conflicting with known edges with high confidence) are removed to obtain the target prediction result. For example, if both Institution A and Institution B predict the association of "disease D—target T" with confidence levels of 0.91 and 0.87 respectively, then Institution A's prediction is retained, and the two are merged into a multi-source cross-validation.

[0154] Step S5.6: Based on the node attribute prediction values ​​and edge relationship prediction values ​​output in the target prediction results, perform node attribute completion and edge relationship expansion on the verified local subgraph to construct the cross-institutional research graph dataset.

[0155] Based on the node attribute predictions and edge relationship predictions output by the target prediction results, the local subgraph after verification is augmented with node attributes and edge relationships. Specifically, newly added node attributes with high confidence from the target prediction results are backfilled to the corresponding nodes, and newly added edge relationships with high confidence (such as potential disease-target associations, unobserved protein interactions) are added as new edges to the subgraph. This supplements the local subgraph with scientific associations inferred by the model in addition to the original measured data, forming enhanced subgraph datasets for each institution. The coordination server further aggregates the enhanced subgraph datasets and quality score annotations from each institution, performs cross-institutional multi-source confidence weighted fusion and redundancy removal processing based on graph isomorphic networks, merges the same associations inferred by different institutions, and removes contradictory terms. As a result, the size and confidence of the cross-institutional scientific graph dataset exceed that of any single institution.

[0156] Example 9

[0157] Based on Example 8, the step of performing different graph data augmentation operations on the verified local subgraph to generate an augmentation variant set includes:

[0158] Step S5.1.1: Perform at least two of the following operations on the verified local subgraph: node discarding operation, edge perturbation operation, feature masking operation, and subgraph sampling operation, respectively, to generate corresponding enhanced graph variants to form the enhanced variant set.

[0159] The enhanced inference and graph construction module receives the verified local subgraph as input and performs at least two graph data augmentation operations on the verified local subgraph to generate an enhanced variant set. Each selected operation is applied independently to the same original subgraph without interference, each generating a corresponding enhanced graph variant. At least two operations are required because a single augmentation perspective has limited dimensions for perturbing the subgraph. Multiple operations can characterize the same subgraph from different dimensions such as nodes, edges, features, and substructures, enabling subsequent multi-source fusion to have richer complementary perspectives and improving the robustness of inference to local perturbations. The enhanced graph variants generated by at least two augmentation operations are collected and summarized to form the enhanced variant set, which is then used for subsequent inference.

[0160] Step S5.1.2: The node discarding operation is used to randomly remove some nodes and some edges associated with the nodes in the verified local subgraph based on a preset discarding probability, and generate a node discarding enhanced graph variant.

[0161] The node discarding operation is used to randomly remove a portion of nodes and their associated edges from the validated local subgraph based on a preset discarding probability. By breaking the original connection structure of the subgraph, it forces the model to avoid over-reliance on the adjacency relationships of specific nodes. For subgraphs containing virtual anchor nodes, which represent cross-agency connections, they are generally retained during node discarding to maintain the cross-agency information obtained from the dual-domain compensation. In this embodiment, the preset discarding probability is set to 0.1, and approximately 10% of the nodes and their associated edges are randomly deleted to generate a node discarding enhanced graph variant.

[0162] Step S5.1.3: The edge perturbation operation includes an edge deletion sub-operation and an edge addition sub-operation. Based on the edge deletion sub-operation, existing edges in the verified local subgraph are randomly deleted, and based on the edge addition sub-operation, new edges that do not exist in the verified local subgraph are randomly added to generate an edge perturbation enhanced graph variant.

[0163] The edge perturbation operation includes edge deletion and edge addition sub-operations: the edge deletion sub-operation randomly deletes existing edges in the verified local subgraph, and the edge addition sub-operation randomly adds new edges that do not exist in the verified local subgraph. Both perturb the topological connectivity without changing the node set—edge deletion simulates missing connections in a real-world scenario, while edge addition simulates potentially existing implicit associations. Deletion and addition are sampled independently on the same edge set to avoid the same edge being both deleted and added, thus canceling each other out. In this embodiment, the edge deletion probability is set to 0.1, and the edge addition probability is set to 0.1, thereby generating an edge perturbation-enhanced graph variant.

[0164] Step S5.1.4: The feature masking operation is used to set the feature vectors of some nodes in the verified local subgraph to zero vectors based on a preset masking probability, thereby generating a feature masking enhanced graph variant.

[0165] The feature masking operation is used to set the feature vectors of some nodes in the verified local subgraph to zero based on a preset masking probability. This forces the model to learn an expression that does not rely on a single feature dimension, improving robustness to missing features or noise. The feature vectors of nodes that are not masked remain unchanged, and the subgraph structure and adjacency relationships are unaffected. In this embodiment, the preset masking probability is 0.2, which sets the feature vectors of approximately 20% of the nodes to zero, generating a feature-masked enhanced graph variant.

[0166] Step S5.1.5: The subgraph sampling operation is used to extract locally connected subgraphs from the local subgraph after verification based on the random walk algorithm, and generate a subgraph sampling enhanced graph variant.

[0167] Subgraph sampling is used to extract locally connected subgraphs from a validated local subgraph based on a random walk algorithm. Starting from a random seed node, it continuously jumps to fixed steps according to adjacency relationships, collecting the nodes passed through and the edges connected to these nodes to form a locally connected subgraph, providing a local perspective at a different scale from the full graph. Multiple sampled subgraphs are sampled with different seed nodes, covering different regions of the subgraph. In this embodiment, the random walk length is 20 and the number of sampled subgraphs is 5, thereby extracting 5 locally connected subgraphs and generating a subgraph sampling enhancement graph variant.

[0168] Example 10

[0169] Based on Example 8, the step of determining the multi-source confidence score of the augmentation prediction value corresponding to each augmentation map variant according to the information entropy includes:

[0170] Step S5.3.1: The information entropy of each augmentation graph variant is transformed using a negative correlation mapping function to obtain the initial confidence level of the augmentation prediction value corresponding to each augmentation graph variant, wherein the value of the information entropy is negatively correlated with the value of the initial confidence level.

[0171] The augmented inference and graphing module receives the predicted probability distribution of each augmented graph variant as input. Information entropy is used to measure the uncertainty of the augmented graph variant predictions: when the predicted probabilities are concentrated in a single class, the distribution is sharp and the entropy is low, indicating a more certain prediction by the model for the augmented graph variant; when the predicted probabilities tend to be uniform across classes, the distribution is flat and the entropy is high, indicating greater uncertainty in the model. The higher the information entropy (the more uncertain the prediction), the lower the initial confidence, and the two are negatively correlated. In this embodiment, a negative exponential form of negative correlation mapping is used, so that the higher the information entropy, the lower the resulting initial confidence, with the highest initial confidence achieved when the information entropy is 0.

[0172] Step S5.3.2: Perform cross-variant normalization on the initial confidence scores of the augmented prediction values ​​corresponding to each augmented graph variant, and determine the result of the cross-variant normalization as the multi-source confidence score of the augmented prediction values ​​corresponding to each augmented graph variant.

[0173] The initial confidence scores of the augmentation prediction values ​​corresponding to each augmentation map variant are normalized across variants. Since the initial confidence scores of different augmentation map variants are unnormalized quantities with inconsistent dimensions and scales, making direct comparison impossible, softmax normalization is used to map the initial confidence scores of all variants to the same scale, ensuring that the scores of each variant are non-negative and their sum is 1, thus becoming weights that can be directly used for weighted fusion. The multi-source confidence score of the v-th augmentation map variant is determined by the following formula:

[0174]

[0175] in, For the v-th augmented graph variant, the multi-source confidence score is... Let be the information entropy of the v-th augmented graph variant, V be the total number of augmented graph variants in the augmented variant set, and u be the index of the augmented graph variant, u=1,2,…,V.

[0176] Example 11

[0177] like Figure 4 As shown, this invention provides a cross-institutional collaboration system based on graph federation learning, comprising:

[0178] The data preprocessing module 401 is deployed in the local environment of each participating institution. It receives local raw scientific research data and performs two types of operations: First, based on the shared ontology tree negotiated across institutions, it maps the heterogeneous node attributes of the local data to a unified semantic space and maps the edge relationships to a standardized set of relation types, thus constructing a standardized local subgraph with consistent semantics. Second, it performs graph structure verification on the standardized local subgraph based on the shared ontology tree, identifies and corrects edges that violate ontology constraints, obtains the verified local subgraph, and extracts the subgraph topological statistics and feature distribution summaries after differential privacy perturbation to form a local subgraph summary.

[0179] The dual-domain compensation module 402, deployed locally, receives the verified local subgraph and its summary. It repairs the graph signal loss caused by the boundary segmentation of the mechanism in two domains: In the spatial domain, it identifies boundary nodes connected to cross-mechanism edges and establishes a virtual anchor node for each boundary node, using the previous round of embedding in the global model as its initial feature. It compensates for missing cross-mechanism neighbor information through local message passing. In the spectral domain, it performs eigenvalue decomposition on the normalized Laplace matrix of the local subgraph to obtain the graph Fourier transform basis. It then applies gain to the truncated low-frequency global signal, smooths and suppresses the high-frequency noise introduced by the segmentation, and performs an inverse transform back to the node domain. The dual-domain compensation module adaptively fuses the compensation results from both domains according to the topological density of the local subgraph, forming a dual-domain compensated graph substructure feature.

[0180] The training and encryption module 403 is deployed in a local environment. It receives the dual-domain compensated graph substructure features and uses a graph convolutional network as the local model for sharpness-aware training. It finds the perturbation direction that makes the loss rise steepest in the neighborhood of the loss surface and backpropagates along the parameters after perturbation, forcing the model to converge to a flat region to suppress overfitting and obtain the enhanced local gradient. Then, it sequentially performs gradient clipping, differential privacy noise perturbation and homomorphic encryption encapsulation on the enhanced local gradient to obtain the privacy-preserving encrypted gradient package and upload it.

[0181] The global aggregation module 404, deployed on the federated coordination server, receives privacy-preserving encrypted gradient packets and local subgraph summaries uploaded by various institutions. First, it sums the ciphertext using the additive homomorphic property of homomorphic encryption and decrypts it via a secure aggregation protocol to obtain the aggregated gradient plaintext. Then, it measures the feature distribution offset of each institution along the semantic axis based on the perturbed feature distribution summary and compensates for the gradient of the feature mapping layer. Along the structural axis, it measures the graph structure differences based on topological statistics and scales the gradient of the message passing layer, correcting heterogeneous biases before aggregation. Finally, it weights and aggregates the gradients after dual-axis alignment according to adaptive weights determined by the data size of each institution, local loss, and alignment bias, generating global graph neural network model parameters and distributing them.

[0182] The enhanced inference and graph construction module 405 is deployed in the local environment. It receives the global graph neural network model parameters and the verified local subgraph. First, it updates the local graph neural network model with the global model parameters. Then, it applies at least two types of graph data enhancement, such as node dropping, edge perturbation, feature masking, and subgraph sampling, to the verified local subgraph to generate multiple enhancement variants. Each enhancement variant is input into the model for inference to obtain the predicted probability distribution. The confidence score of each prediction is obtained based on the information entropy mapping. The multi-source predictions are weighted and fused according to the confidence scores, and cross-institutional duplicate and contradictory predictions are removed. Finally, the local subgraph is augmented with node attributes and edge relationships. The coordination server merges and removes redundancy across institutions to construct a cross-institutional research graph dataset.

[0183] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A cross-institutional collaboration method based on graph federation learning, characterized in that, include: Local raw scientific research data is mapped to node attributes and edge relationships to construct a standardized local subgraph. A shared ontology tree is used to perform graph structure verification on the standardized local subgraph and extract the local subgraph summary to obtain the verified local subgraph and the local subgraph summary. To address the graph signal loss caused by the mechanism boundary segmentation in the local subgraph after verification, compensation operations in the spatial domain and the spectral domain are performed respectively. The results of the compensation operations in the spatial domain and the spectral domain are then fused to form a dual-domain compensated graph substructure feature. The dual-domain compensated graph substructure features are input into the local graph neural network model for sharpness-aware local training, and the enhanced local gradient is output. The enhanced local gradient is then encrypted for privacy protection and encapsulated into a privacy-protected encrypted gradient package. The privacy-protected encrypted gradient packages uploaded by each participating institution are aggregated and decrypted using a secure aggregation protocol to obtain aggregated gradient plaintext. The aggregated gradient plaintext is then combined with the local subgraph digest to perform dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes to generate global graph neural network model parameters. Based on the parameters of the global graph neural network model, inference and enhanced prediction are performed on the verified local subgraph, outputting multiple enhanced prediction values. The multiple enhanced prediction values ​​are then summarized, multi-source confidence weighted fusion is performed, and redundancy is removed to construct a cross-institutional research graph dataset.

2. The method according to claim 1, characterized in that, The process involves mapping node attributes and edge relationships to local raw scientific research data to construct a standardized local subgraph. A shared ontology tree is then used to perform graph structure verification on the standardized local subgraph and extract a local subgraph summary, resulting in a verified local subgraph and its summary. By pre-negotiating a shared ontology tree across institutions, local heterogeneous node attributes are mapped to a unified semantic space to obtain standardized node attributes, and edge relationships are mapped to a standardized set of relationship types to obtain standardized edge relationships. The standardized local subgraph is then constructed based on the standardized node attributes and the standardized edge relationships. Based on the shared ontology tree, the standardized local subgraph is subjected to graph structure verification, edges that violate ontology constraints are identified and corrected, and the verified local subgraph is obtained. The topological statistics and feature distribution summary of the verified local subgraph are calculated. The feature distribution summary is subjected to differential privacy perturbation to obtain the perturbed feature distribution summary. The local subgraph summary is composed of the subgraph topological statistics and the perturbed feature distribution summary.

3. The method according to claim 1, characterized in that, The method addresses the graph signal loss caused by the mechanism boundary segmentation in the verified local subgraph by performing spatial and spectral domain compensation operations respectively, and then fusing the results of the spatial and spectral domain compensation operations to form a dual-domain compensated graph substructure feature, including: Identify the boundary nodes in the verified local subgraph, where the boundary nodes are nodes connected to cross-mechanism edges; For each of the boundary nodes, a virtual anchor node is established. The embedding representation of the corresponding position of the boundary node is obtained using the parameters of the global graph neural network model generated in the previous iteration. The embedding representation is used as the initial feature of the virtual anchor node. Message passing aggregation is performed based on the initial feature to obtain the spatial domain compensation result. The adjacency matrix and degree matrix are calculated based on the verified local subgraph, and a normalized Laplace matrix is ​​constructed based on the adjacency matrix and the degree matrix. The spatial domain compensation result is then subjected to graph Fourier transform based on the normalized Laplace matrix to obtain the frequency domain components. A smoothing operation is performed on the high-frequency components whose corresponding feature values ​​are greater than a preset frequency threshold, and a gain operation is performed on the low-frequency components whose corresponding feature values ​​are less than or equal to the preset frequency threshold to obtain the compensated frequency components; a graph inverse Fourier transform is performed on the compensated frequency components to reconstruct the node features to obtain the spectrum domain compensation result. The topological density is calculated for the verified local subgraph, and the adaptive fusion weight is determined based on the topological density. The spatial domain compensation result and the spectral domain compensation result are weighted and fused to form the dual-domain compensation graph substructure feature.

4. The method according to claim 1, characterized in that, The step involves inputting the dual-domain compensated graph substructure features into a local graph neural network model for sharpness-aware local training, outputting enhanced local gradients, and performing privacy-preserving encryption processing on the enhanced local gradients, encapsulating them into a privacy-preserving encrypted gradient package, including: The sharpness-aware minimization method is used to train the local graph neural network model with the substructure features of the dual-domain compensated graph as input. Forward propagation is performed to calculate the prediction loss, and backpropagation is performed based on the prediction loss to obtain the initial gradient. Based on the initial gradient, the perturbation direction of the parameters of the local graph neural network model is determined. After applying a perturbation of a preset step size to the parameters of the local graph neural network model in the perturbation direction, forward propagation and backward propagation are performed again to calculate the perturbation gradient. The initial gradient and the perturbation gradient are weighted and fused to generate the enhanced local gradient; Gradient clipping is performed on the enhanced local gradient to bind sensitivity. Differential privacy noise variance is calculated based on a preset privacy budget and the sensitivity. Random noise that satisfies the differential privacy noise variance is generated and added to the clipped gradient to obtain the privacy-enhanced gradient. The privacy-enhancing gradient is encrypted using a preset homomorphic encryption public key to obtain gradient ciphertext, and the gradient ciphertext and the identifier of the current iteration round are encapsulated into the privacy-protected encrypted gradient package.

5. The method according to claim 1, characterized in that, The privacy-preserving encrypted gradient packets uploaded by each participating institution are aggregated and decrypted using a secure aggregation protocol to obtain aggregated gradient plaintext. The aggregated gradient plaintext is then combined with the local subgraph digest to perform dual-axis alignment and adaptive weighted aggregation along the semantic and structural axes, generating global graph neural network model parameters, including: The privacy-protected encrypted gradient packets uploaded by each participating institution are summed using a secure aggregation protocol to obtain aggregated ciphertext. The aggregated ciphertext is decrypted to obtain the aggregated gradient plaintext, and the local subgraph digests uploaded by each participating institution are obtained. The aggregated gradient plaintext includes the aggregated gradient of the graph neural network feature mapping layer parameters and the aggregated gradient of the graph neural network message passing layer parameters. The local subgraph digest includes the perturbed feature distribution digest and the subgraph topological statistics. Calculate the semantic offset of each participating institution corresponding to the perturbed feature distribution summary, and use the semantic offset to compensate and correct the aggregate gradient of the graph neural network feature mapping layer parameters to obtain the semantically aligned feature mapping layer parameter gradient. The structural topology differences of each participating institution are calculated based on the subgraph topology statistics, and the aggregate gradient of the message passing layer parameters of the graph neural network is scaled and aligned based on the structural topology differences to obtain the structure-aligned message passing layer parameter gradient. By fusing the gradients of the feature mapping layer parameters after semantic alignment with the gradients of the message passing layer parameters after structural alignment, the alignment global gradient is obtained. The adaptive aggregation weights of each participating mechanism are determined by combining the semantic offset and the structural topology difference. The aligned global gradient is then weighted and aggregated based on the adaptive aggregation weights to generate the parameters of the global graph neural network model.

6. The method according to claim 5, characterized in that, The calculation of the semantic offset of each participating mechanism corresponding to the perturbed feature distribution summary, and the compensation and correction of the aggregate gradient of the graph neural network feature mapping layer parameters using the semantic offset to obtain the semantically aligned feature mapping layer parameter gradient, includes: Calculate the distribution distance between the perturbed feature distribution summary of each participating institution and the global feature distribution reference summary, and determine the distribution distance as the semantic offset. The global feature distribution reference summary is generated by the coordination server by aggregating the perturbed feature distribution summaries uploaded by each participating institution. Based on the semantic offset, a semantic compensation coefficient is constructed. The semantic compensation coefficient is then used to scale and compensate the gradient of the aggregated gradient of the feature mapping layer parameters of the graph neural network, thereby obtaining the semantically aligned gradient of the feature mapping layer parameters.

7. The method according to claim 5, characterized in that, The step of calculating the structural topology differences of each participating institution based on the subgraph topology statistics, and scaling and aligning the aggregated gradients of the graph neural network message passing layer parameters based on the structural topology differences to obtain the structurally aligned message passing layer parameter gradients, includes: Calculate the topological divergence between the subgraph topological statistics of each participating institution and the global topological statistical reference, and determine the topological divergence as the structural topological difference. The global topological statistical reference is generated by the coordination server by aggregating the subgraph topological statistics uploaded by each participating institution. Based on the structural topology differences, a structural scaling factor is constructed. The structural scaling factor is then used to scale and align the aggregate gradients of the message passing layer parameters of the graph neural network to obtain the structurally aligned gradients of the message passing layer parameters.

8. The method according to claim 1, characterized in that, The inference and augmentation prediction based on the global graph neural network model parameters are performed on the verified local subgraph, outputting multiple augmentation prediction values. These multiple augmentation prediction values ​​are then aggregated, weighted fused using multi-source confidence, and redundancy removed to construct a cross-institutional research graph dataset, including: The local graph neural network model is updated using the parameters of the global graph neural network model, and different graph data augmentation operations are performed on the verified local subgraph to generate an augmentation variant set. Each augmented graph variant in the augmented variant set is input into the updated local graph neural network model for inference, and the predicted probability distribution corresponding to each augmented graph variant is output. Multiple augmented prediction values ​​are generated based on the predicted probability distribution. Based on the predicted probability distribution corresponding to each augmented graph variant, the information entropy of each augmented graph variant is calculated, and the multi-source confidence score of the augmented predicted value corresponding to each augmented graph variant is determined according to the information entropy. The multi-source confidence scores are used as fusion weights to perform weighted fusion of the multiple enhanced prediction values ​​to obtain the initial fusion prediction result. Calculate the feature similarity and structural overlap between each predicted entity node and relation edge in the initial fusion prediction result. Based on the feature similarity and structural overlap, identify and remove redundant and conflicting prediction terms to obtain the target prediction result. Based on the node attribute prediction values ​​and edge relationship prediction values ​​output in the target prediction results, the verified local subgraph is augmented with node attributes and edge relationships to construct the cross-institutional research graph dataset.

9. The method according to claim 8, characterized in that, The step of performing different graph data augmentation operations on the verified local subgraph to generate an augmentation variant set includes: Perform at least two of the following operations on the verified local subgraph: node discarding operation, edge perturbation operation, feature masking operation, and subgraph sampling operation, respectively, to generate corresponding enhanced graph variants to form the enhanced variant set; The node discarding operation is used to randomly remove some nodes and some edges associated with nodes in the verified local subgraph based on a preset discarding probability, generating a node discarding enhanced graph variant. The edge perturbation operation includes an edge deletion sub-operation and an edge addition sub-operation. The edge deletion sub-operation randomly deletes existing edges in the verified local subgraph, and the edge addition sub-operation randomly adds new edges that do not exist in the verified local subgraph to generate an edge perturbation enhanced graph variant. The feature masking operation is used to set the feature vectors of some nodes in the verified local subgraph to zero vectors based on a preset masking probability, thereby generating a feature masked enhanced graph variant. The subgraph sampling operation is used to extract locally connected subgraphs from the local subgraph after verification based on the random walk algorithm, and generate a subgraph sampling enhanced graph variant.

10. The method according to claim 8, characterized in that, The step of determining the multi-source confidence score of the augmented prediction value corresponding to each augmented map variant based on the information entropy includes: The information entropy of each augmentation graph variant is transformed using a negative correlation mapping function to obtain the initial confidence level of the augmentation prediction value corresponding to each augmentation graph variant, wherein the value of the information entropy is negatively correlated with the value of the initial confidence level; The initial confidence scores of the augmented prediction values ​​corresponding to each augmented graph variant are normalized across variants, and the results of the cross-variant normalization are determined as the multi-source confidence scores of the augmented prediction values ​​corresponding to each augmented graph variant.

11. A cross-institutional collaborative system based on graph federated learning, characterized in that, include: The data preprocessing module is used to map node attributes and edge relationships to local raw scientific research data to construct a standardized local subgraph. The shared ontology tree is used to perform graph structure verification on the standardized local subgraph and extract the local subgraph summary to obtain the verified local subgraph and the local subgraph summary. The dual-domain compensation module is used to perform spatial domain and spectral domain compensation operations respectively for the graph signal loss caused by the mechanism boundary segmentation in the verified local subgraph, and to fuse the results of the spatial domain and spectral domain compensation operations to form a dual-domain compensated graph substructure feature. The training and encryption module is used to input the dual-domain compensated graph substructure features into the local graph neural network model for sharpness-aware local training, output the enhanced local gradient, and perform privacy-preserving encryption processing on the enhanced local gradient, encapsulating it into a privacy-preserving encrypted gradient package. The global aggregation module is used to aggregate and decrypt the privacy-protected encrypted gradient packages uploaded by each participating institution through a secure aggregation protocol to obtain aggregated gradient plaintext; and to perform dual-axis alignment and adaptive weighted aggregation on the aggregated gradient plaintext along the semantic axis and structural axis in combination with the local subgraph digest to generate global graph neural network model parameters. The enhanced inference and graph construction module is used to perform inference and enhanced prediction on the verified local subgraph based on the parameters of the global graph neural network model, output multiple enhanced prediction values, and summarize the multiple enhanced prediction values, perform multi-source confidence weighted fusion and redundancy removal processing to construct a cross-institutional scientific research graph dataset.

Images