A federated graph learning method, system, device, and storage medium

Abstract

The present application relates to the technical field of federated learning, and relates to a federated graph learning method, system, device and storage medium; wherein the federated graph learning method comprises the following steps: a server calculates a weight vector according to the number of global category samples, gives higher weight to a scarce category, and distributes the weight vector to a client; the client establishes a double-encoder architecture, a first encoder aggregates neighbor information based on a graph structure, and a second encoder extracts node features independently of the graph structure; the client maintains a semantic center vector, and updates the semantic center vector stably through a momentum mechanism, and marks reliable nodes based on prediction confidence; the server establishes a cross-domain consensus mapping model, weightedly fuses client parameters, and synthesizes a global semantic center vector; and the client fuses local and global semantic center vectors. The method can fundamentally suppress semantic drift, and improve the classification accuracy of minority class nodes and the robustness of a model.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a federated graph learning method, system, device and storage medium. Background Technology

[0002] Graph Neural Networks (GNNs) exhibit superior performance when processing graph-structured data. However, in practical applications, large amounts of graph data are generated by multiple different data owners, making centralized training impossible due to privacy constraints and commercial competition. Federated learning, as a distributed learning solution, enables multiple data owners to collaboratively train models without sharing private data. However, federated graph learning faces severe data heterogeneity issues, as well as the unique challenge of a small number of nodes exhibiting high heterogeneity.

[0003] Existing federated graph learning techniques primarily employ a paradigm that combines centralized parameter aggregation with local graph neural network training. Typical approaches include: The first type is a federated graph learning framework based on direct parameter aggregation. Each client trains a GNN model locally and only uploads the model parameters to the server, which then performs a federated averaging operation. This approach relies entirely on local graph structures for message passing and is effective for independent, identically distributed data, but it cannot handle the heterogeneity of graph structures across clients.

[0004] The second category: Federated graph learning frameworks based on structure-feature decoupling. To mitigate the impact of structural heterogeneity, some existing technologies employ encoder decoupling strategies, where each client maintains both a structure encoder and a feature encoder, and the server only aggregates the structure encoder parameters.

[0005] The third category: Federated graph learning frameworks based on knowledge distillation and prototype alignment. Some advanced solutions introduce a knowledge distillation mechanism, where each client computes a local category prototype, the server aggregates these prototypes to obtain a global prototype, and the client uses the global prototype as a soft target to guide local training.

[0006] However, existing technologies have the following drawbacks: (1) It cannot effectively alleviate the problem of semantic drift across clients. The distribution of the latent representations learned by nodes of the same semantic category differs significantly on different clients, and the decision boundary of the minority class is easily submerged by the gradient of the majority class.

[0007] (2) Minority nodes suffer from structural assimilation. Minority nodes are aggregated into misleading structural information because their neighbors are mostly heterogeneous, causing node embeddings to be incorrectly pulled toward the majority manifold.

[0008] (3) There is a lack of reliability screening mechanism for cross-client knowledge transfer. Existing methods use undifferentiated knowledge transfer, and the error caused by low confidence prediction will accumulate and be amplified in federated aggregation.

[0009] (4) Lack of a unified global semantic anchoring mechanism. Each client learns incompatible representation spaces due to domain offset, affecting the consistency of the global model. Summary of the Invention

[0010] This application aims to at least address the technical problems existing in the prior art and provide a federated graph learning method, system, device, and storage medium.

[0011] In a first aspect, the present invention provides a federated graph learning method, which is applied to a federated learning system comprising one server and multiple clients, each client holding private graph data, including: The server aggregates the category sample statistics uploaded by each client and calculates the global category distribution based on the category sample statistics uploaded by each client. The server determines the category weight vector based on the global category distribution and broadcasts the category weight vector to each client; Each client establishes a topology context view encoder and an intrinsic semantic view encoder; Each client maintains a semantic center vector for each category and updates the semantic center vector through a momentum mechanism. Each client computes the prediction confidence of the intrinsic semantic view encoder on the validation set, dynamically generates a fidelity mask, and marks reliable nodes; Each client calculates the topology attempt loss and updates the model parameters of the local semantic essence prototype based on the category weight vector, fidelity mask, semantic center vector and node features of reliable nodes, and then uploads the model parameters to the server. The server establishes a cross-domain consensus mapping model, and constrains the semantic center vector of each client to approximate the global semantic manifold based on the model parameters uploaded by each client; The server aggregates the semantic center vectors uploaded by each client, generates a global category semantic center vector, and broadcasts it to each client. Each client integrates the local semantic center vector and the global semantic center vector.

[0012] Optionally, the topology view loss includes mutual consistency alignment loss, classification loss, reverse alignment loss, and regularization loss.

[0013] Optionally, each client trains the topological context view encoder with weighted cross-entropy loss using the category weight vector and calculates the mutual consistency alignment loss under fidelity mask constraints. Each client uses a learnable gating mechanism to fuse semantic center vectors and node features with the internal semantic view encoder, and calculates classification loss, back alignment loss, and regularization loss.

[0014] Optionally, the model parameters uploaded by the client to the server include topology view parameters, intrinsic semantic view parameters, local SEP set, and local label count vector.

[0015] Optionally, the formula for calculating the category weight vector is: ; in, Indicates the first In each round of communication, the server calculates and broadcasts the global calibration adjustment weights corresponding to the category to each client. The weight values; C is the total number of categories; For category The intermediate weights; , For adjustment coefficients, The number of samples in the global category c.

[0016] Optionally, the node representation update formula for the topology context view encoder is: ; in, For point In the The layer's topological context view hides the representation, where AGG is an aggregation function. For activation function, For the first The learnable weight matrix of the layer, Represents a node All neighboring nodes In the The topological context view of a layer hides the collection that it represents; Representing neighboring nodes In the The topology context view of the layer hides the representation. For the first The learnable bias vector of the layer, For the local neighborhood of client k.

[0017] Optionally, the topological context view encoder is a graph neural network encoder, and the intrinsic semantic view encoder is a multilayer perceptron encoder.

[0018] Secondly, this invention provides a federated graph learning system, the system comprising a server and multiple clients. The server includes: The receiving unit is used to receive statistical information on category samples uploaded by each client; The weight generation unit is used to calculate the class weight vector based on the number of global class samples. A broadcasting unit is used to broadcast the category weight vector to each client. The distiller unit is used to establish a cross-domain consensus mapping model; The aggregation unit is used to aggregate the semantic center vectors uploaded by each client; The at least one client includes: A dual encoder unit is used to establish a topological context view encoder and an intrinsic semantic view encoder; The prototype management unit is used to maintain a semantic center vector for each category; An evaluation unit is used to calculate the prediction confidence and generate a fidelity mask. Training unit, used to train the topological context view encoder and the intrinsic semantic view encoder; The fusion unit is used to fuse the local semantic center vector and the global semantic center vector.

[0019] Thirdly, the present invention provides an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the federated graph learning method described above.

[0020] Fourthly, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the federated graph learning method described above.

[0021] In summary, this application includes the following beneficial technical effects: By aggregating global category statistics on the server and calculating category weight vectors based on the number of effective samples, the gradient of scarce categories is proactively amplified from an optimization perspective, fundamentally alleviating cross-client semantic drift and significantly improving the classification accuracy of minority classes. At the same time, each client establishes a dual decoupled architecture of topological context view and internal semantic view in parallel, and maintains a semantic center vector for each category that is smoothly updated by the momentum mechanism, effectively eliminating the structural assimilation problem of minority class nodes and enhancing their immunity to topological noise. The client dynamically evaluates the prediction confidence on the validation set and generates a fidelity mask, allowing only high-confidence reliable nodes to participate in cross-view knowledge distillation, thus blocking the propagation of erroneous knowledge at the source and improving distillation efficiency and training stability. The server establishes a cross-domain consensus mapping model, which uses regularization to constrain the semantic center vectors of each client to approximate the global semantic manifold, compensating for feature offsets and achieving unified cross-client semantic alignment. This method only requires uploading lightweight statistics and prototype vectors, with communication overhead far lower than node-level embedding, and is compatible with secure aggregation and differential privacy protection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the processing steps of a federated graph learning method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the framework of a federated graph learning method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device that implements the federated graph learning method according to an embodiment of the present invention.

[0023] Reference numerals: 10, processor; 11, memory; 12, communication bus; 13, communication interface.

[0024] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0026] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0028] Existing technologies rely solely on parameter aggregation or static prototype alignment, failing to fundamentally address the aforementioned heterogeneity issues arising from coupling. This patent proposes a federated graph learning framework based on dual-view semantic alignment and dynamic fidelity evaluation. Through a collaborative mechanism of global calibration adjustment, dual-view decoupling architecture, dynamic fidelity evaluation, and mutual consistency alignment, it systematically mitigates semantic drift and improves the classification accuracy and robustness of minority class nodes. The core difference lies in the fact that this patent, for the first time, upgrades the server role from a passive parameter aggregator to an active semantic calibrator, upgrades the client from single GNN training to dual-view collaborative distillation, and introduces a fidelity gating mechanism to selectively filter unreliable structural signals, achieving a paradigm shift from parameter alignment to semantic manifold alignment.

[0029] Reference Figure 1 and Figure 2 This invention discloses a federated graph learning method in one embodiment. The method is applied to a federated learning system comprising one server and multiple clients, each client holding private graph data. In this embodiment, the federated graph learning method includes: S1. The server aggregates the statistical information of category samples uploaded by each client and calculates the global category distribution based on the statistical information of category samples uploaded by each client.

[0030] Specifically, category sample statistics refer to the number of samples (nodes) contained in each category within each client's local data. Simply put, each client counts the number of nodes belonging to category 1, category 2, ..., category C in its graph data, forming a count vector of length C. The client uploads this vector to the server, which aggregates it to obtain the global category count, thus revealing the global sample distribution of each category throughout the federated system. This information forms the basis for the server to calculate the Global Calibration Adjustment (GCA) weights, used to identify scarce categories and assign them higher optimization weights to mitigate semantic drift.

[0031] S2. The server determines the category weight vector based on the global category distribution and broadcasts the category weight vector to each client.

[0032] Assuming the total number of categories is C=7, the adjustment coefficient is... .

[0033] The server calculates the category weight vector using the following formula: ; in, Indicates the first In each round of communication, the server calculates and broadcasts the global calibration adjustment weights corresponding to the category to each client. The weight values; C is the total number of categories; For category The intermediate weights; The formula for calculating the effective sample weight is: This formula is based on the effective sample size theory. For adjustment coefficients, The number of samples in global category c; when The smaller the size (the scarcer the category). Approaching 1-β, it is assigned a higher weight; when When the value is large, the weight is saturated to about 1 to avoid overcompensation.

[0034] This weight vector explicitly informs the client of the global scarcity of each class, providing prior protection for minority classes. The server will... Broadcast to all clients.

[0035] S3. Each client establishes a topology context view encoder and an intrinsic semantic view encoder.

[0036] The topological context view encoder is a graph neural network encoder, and the intrinsic semantic view encoder is a multilayer perceptron encoder. S4. Each client maintains a semantic center vector for each category and updates the semantic center vector through a momentum mechanism.

[0037] Each client k initializes two heterogeneous encoders in parallel. In this embodiment, the topological context view encoder employs a 2-layer graph convolutional network (GCN) with hidden dimensions. .node The representation at layer l is updated as follows: The node representation update formula for the topology context view encoder is: ; in, For point In the The layer's topological context view hides the representation, where AGG is an aggregation function. For activation function, For the first The learnable weight matrix of the layer, Represents a node All neighboring nodes In the The topological context view of a layer hides the collection that it represents; Representing neighboring nodes In the The topology context view of the layer hides the representation. For the first The learnable bias vector of the layer, For the local neighborhood of client k.

[0038] S5. Each client calculates the prediction confidence of the intrinsic semantic view encoder on the validation set, dynamically generates a fidelity mask, and marks reliable nodes.

[0039] S6. Each client calculates the topology attempt loss based on the category weight vector, fidelity mask, semantic center vector, and node features of reliable nodes, updates the model parameters of the local semantic essence prototype, and uploads the model parameters to the server.

[0040] The topology view loss includes mutual consistency alignment loss, classification loss, back alignment loss, and regularization loss. Each client trains the topology context view encoder using a weighted cross-entropy loss with a class weight vector and calculates the mutual consistency alignment loss under fidelity mask constraints. For the intrinsic semantic view encoder, each client fuses the semantic center vector and node features through a learnable gating mechanism and calculates the classification loss, back alignment loss, and regularization loss.

[0041] The model parameters uploaded by the client to the server include topology view parameters, intrinsic semantic view parameters, local SEP set, and local label count vector.

[0042] S7. The server establishes a cross-domain consensus mapping model and constrains the semantic center vector of each client to approximate the global semantic manifold based on the model parameters uploaded by each client.

[0043] S8. The server aggregates the semantic center vectors uploaded by each client, generates a global category semantic center vector, and broadcasts it to each client.

[0044] S9. Each client integrates the local semantic center vector and the global semantic center vector.

[0045] To facilitate understanding by those skilled in the art, the actions performed by the client and server are described below. The actions performed by the client are as follows: Before performing the first round of training, the personalized GNN model is initialized. The initialization steps are as follows: Step 101: Initialize the dual-view model Unlike existing technologies that initialize only a single GNN model, this patent initializes two heterogeneous encoders in parallel on each client k, forming a decoupled dual-view architecture: Topological Context View: Initialization The parameter is denoted as This view uses a standard messaging mechanism, and the nodes... In the The formula for updating the hidden representation of a layer is: ; in Let σ represent the local neighborhood of client k, where AGG is the aggregation function (such as MEAN or SUM), and σ is the ReLU activation function. Initial input The final layer output is then processed by Softmax to obtain the predicted distribution: This view is responsible for capturing client-private structural dependencies, but it is susceptible to interference from local topological heterogeneity.

[0046] Intrinsic Semantic View: Initializes the structure-independent MLP encoder The parameter is denoted as Nodes are represented to depend only on their own features, and the update formula is: This view provides a stable semantic baseline by isolating structural noise.

[0047] Existing technologies maintain only a single GNN view, resulting in structural noise and semantic signal coupling contamination. This patent achieves structure-semantic decoupling, enabling each view to perform its own function, thus laying the foundation for subsequent selective knowledge distillation.

[0048] Step 102: Semantic essence prototype initialization Initialize the category-level semantic anchor set Each prototype This represents the semantic essence vector of category c. The initialization strategy is:

[0049] in Let k be the set of labeled nodes of category c in client k. This is the initial output for the intrinsic semantic view. Each prototype is equipped with a momentum update buffer, with momentum coefficients m∈[0,1) initialized to 0.9.

[0050] The SEPs in this patent remain stable through a momentum accumulation mechanism, serving as semantic anchors across rounds and reducing update variance.

[0051] In subsequent training communication rounds, the client operates as follows: Local updates for 201 personalized GNN models 201-a Receive global calibration weights The client receives the Global Calibration Adjustment (GCA) weight vector from the server. This vector is based on global category counts. calculate:

[0052] β∈[0,1) is a hyperparameter (typical value 0.999). This weight explicitly informs the client of the global scarcity of each class, providing prior protection for minority classes.

[0053] 201-b Dynamic Fidelity Evaluation (DFA) Generation Mask The reliability of the prediction results of the intrinsic semantic view is evaluated, and a binary fidelity mask is generated. : 1. In the validation set (Without participating in training) Calculate the confidence score of the semantic view prediction:

[0054] in The category distribution output for the semantic view.

[0055] 2. Dynamically calculate client-specific thresholds:

[0056] For the lowest confidence floor, ∈(0,1] is the scaling factor.

[0057] Generate node-level fidelity mask: in This is an indicator function. It only applies when the node prediction confidence exceeds a threshold. Marked as a reliable knowledge source.

[0058] Existing technologies (such as FedProto) indiscriminately align all nodes, leading to error propagation; this patent achieves selective knowledge distillation through the DFA mechanism, fundamentally blocking noise pollution from low-confidence nodes.

[0059] 201-c Topology View Loss Calculation and Parameter Update Topological view under fidelity mask constraints The training loss consists of three components: 1. Globally calibrated cross-entropy loss: for the set of labeled nodes Adjusting the class gradient using GCA weights:

[0060] This loss applies to global minority classes. Apply stronger optimization signals to prevent the decision boundary from being overwhelmed by the majority class.

[0061] 2. Mutual Consistency Alignment Loss: Distilling knowledge from the intrinsic semantic view to the topological view under fidelity mask filtering:

[0062] Where T is the distillation temperature, and σ is the Softmax function. and These are the logits for the two views, respectively. This loss only applies to the mask. The reliable nodes enable the topological view to absorb stable signals from the semantic view and resist structural noise.

[0063] Overall loss and parameter updates:

[0064] Update topology view parameters using the Adam optimizer:

[0065] Intrinsic semantic view parameters during update Keep frozen.

[0066] Existing technologies (such as FedStar) only decouple the encoder but do not perform collaborative optimization; this patent achieves bidirectional semantic correction through MCA gated distillation and introduces GCA weights to actively correct global prior bias at the loss function level, thereby improving robustness of minority classes.

[0067] Step 202: Local update of the intrinsic semantic view (including SEP fusion and backdistillation) 202-a SEP Enhanced Node Representation Fusion For each node embed its semantics SEPs of the corresponding category Integration through learnable gating mechanisms: 1. Calculate the gating weights: in This represents vector concatenation. and These are trainable parameters.

[0068] 2. Generate fused representation: ⊙ represents element-wise multiplication. This mechanism adaptively balances the individual features of nodes with the semantic prior of categories, strengthens the SEP anchoring effect for minority class nodes, and prevents their embedding drift.

[0069] 3. Generate semantic view logits:

[0070] 202-b Semantic View Loss Calculation and Parameter Update 1. Cross-entropy loss:

[0071] in, 2. Reverse MCA Loss: Distillation from the topological view to the semantic view, allowing the semantic view to absorb structural knowledge.

[0072] This operation is performed under DFA mask constraints to ensure that only high-fidelity structural knowledge is distilled into the semantic view.

[0073] 3. SEP Regularization Loss: Constrains the local SEP to be closer to the global semantic distiller. (See step 302):

[0074] This loss prevents SEP from deviating from the semantic manifold of cross-client consensus.

[0075] 4. Overall Loss and Parameter Update:

[0076] Update the intrinsic semantic view parameters and SEPs:

[0077] Topology view parameters during update Keep frozen.

[0078] 202-c SEP Momentum Update (Optional Stability Enhancement) SEP updates are further smoothed using a momentum buffer:

[0079] in Let m be the set of nodes of category c in the current batch, and m = 0.9 be the momentum coefficient. This mechanism makes the SEP update trajectory smooth and resists local batch noise.

[0080] Difference from existing technologies: Existing technologies (such as FedLit) do not employ category-level dynamic fusion; this patent achieves instance-adaptive semantic enhancement through learnable gated SEP fusion, where minority class nodes automatically obtain higher prototype weights, directly counteracting structural assimilation.

[0081] Step 203: Client upload preparation After each round of local training, client k uploads the following parameters (instead of the complete model or raw data) to the server: topology view parameters. (Participating in federal average), intrinsic semantic view parameters Local SEP collection (Participating in semantic consensus aggregation) and local label counting vector (Used for GCA calculations, protected by secure aggregation); Uploaded data volume is strictly limited to [amount missing]. Far lower than the level of embedding at the transmission node level Expenses.

[0082] The server performed the following actions: Step 301: Global Calibration Adjustment (GCA) Weight Calculation The server receives tag counts from each client. Calculate the global category distribution and generate global calibration adjustment weights: 1. Global category count aggregation (supports secure aggregation protocol):

[0083] 2. Calculation of effective sample weights: This formula is based on the effective sample number theory, when The smaller the size (the scarcer the category). Approaching 1-β, it is assigned a higher weight; when When the value is large, the weight is saturated to about 1 to avoid overcompensation.

[0084] 3. Normalize to generate GCA vectors:

[0085] Normalization ensures that the loss scale is consistent with the standard cross-entropy, which facilitates hyperparameter tuning.

[0086] Broadcast GCA weight: The server will Broadcast to all clients for loss calibration in the next round of local topology view training.

[0087] Existing technologies only reweight locally on the client side, which is limited by high variance in local statistics; the GCA of this patent is calculated on the server side based on global consensus, providing unbiased class scarcity priors and actively protecting minority class gradient flows from the perspective of optimization dynamics.

[0088] Step 302: Global Semantic Distiller Update To unify cross-domain feature-label mapping, the server maintains a lightweight MLP distiller. The parameter is ψ: 1. Parameter aggregation: The weights are weighted by the number and volume of training nodes on the client side, ensuring that clients with larger datasets contribute more.

[0089] Cross-domain semantic manifold construction: SEPs are mapped from local semantic spaces to a globally consistent category probability space, forming a cross-domain consensus anchor. During client-side training, The loss pulls the local SEP toward this global manifold, mitigating the impact. Conditional offset.

[0090] Broadcast Distiller: The server will be updated Broadcast to the client to regularize the next round of SEP updates.

[0091] Existing technologies only aggregate static prototypes. This patent establishes a dynamic cross-domain mapping through a learnable global distiller, achieving an upgrade from "point anchoring" to "manifold constraint", which is more adaptable to feature offset.

[0092] Step 303: Global Aggregation and Synchronization of Semantic Essential Prototypes (SEPs) The server performs consensus aggregation on the SEPs uploaded by the client to generate global semantic anchors: 1. Category-level SEP aggregation:

[0093] The weights are the sample proportions of each client category c, ensuring that the global prototype reflects the overall data distribution.

[0094] Synchronous Update: The server will aggregate the global SEP collection. Broadcast to all clients. Clients can optionally merge global SEPs:

[0095] in ∈[0,1] controls the global synchronization strength (default 0.5), balancing local adaptability and global consistency.

[0096] Optional differential privacy protection: For enhanced privacy, the server can add Gaussian noise to the aggregated SEP, where the noise scale... The privacy budget is calculated using a matrix accounting method.

[0097] This patent employs a soft synchronization mechanism and DP enhancement to protect privacy while preserving client-side personalization space and avoiding excessive global smoothing.

[0098] Through a global calibration adjustment mechanism, the server calculates the effective sample weights based on the global class count and generates a normalized calibration vector. This vector directly affects the cross-entropy loss of the topology view training, amplifying the gradient of the globally scarce class and forming a "repulsive" optimization effect. This forces the topology view to push minority class samples away from the majority class clustering area, thereby protecting the minority class semantic manifold.

[0099] Through a dual-view decoupling architecture of a topological context view and an intrinsic semantic view, the intrinsic semantic view completely isolates structural noise, generating pure feature representations. The semantic essence prototype maintains stable updates through a momentum accumulation mechanism, effectively reducing variance as a cross-round semantic anchor. In the node representation fusion stage, the prototype is adaptively fused with node embeddings through a learnable gating mechanism. When structural noise is strong, the prototype weight is automatically strengthened, providing reliable semantic support for minority class nodes independent of the local topology.

[0100] A dynamic fidelity evaluation mechanism is used to calculate the average confidence of semantic view predictions on the validation set, dynamically generating client-specific thresholds. Only when a node's prediction confidence exceeds the threshold is it marked as a reliable knowledge source by a binary mask. During mutual consistency alignment, the knowledge distillation loss is constrained by the mask, allowing only high-fidelity nodes to participate in cross-view knowledge exchange, thus blocking the propagation of structural noise at its source.

[0101] The server maintains a global semantic distiller, using a lightweight neural network to achieve cross-domain consensus mapping. This distiller performs volume-weighted aggregation on the server side, encapsulating cross-client feature-label association consensus. During client training, the semantic essence prototype regularization loss constrains local prototypes to approach the global semantic manifold, forcing similar prototypes from different clients to converge to the same neighborhood in the latent space.

[0102] Employing a lightweight encoder architecture, the number of parameters is significantly smaller than that of traditional graph neural networks. Each round, the client only uploads category-level semantic essence prototype vectors, reducing communication overhead by two orders of magnitude compared to transmitting node-level embeddings. The global semantic distiller is maintained locally on the server, eliminating the need for client-server transmission. Tag count statistics can be calculated using a standard secure aggregation protocol; the server only receives the aggregation result and cannot access the data distribution of individual clients. Optionally, Gaussian noise can be added to the aggregated prototypes to meet differential privacy requirements.

[0103] Global calibration adjustment provides optimized initial conditions for mutual consistency alignment, improving the efficiency of reliable node screening; semantic essence prototype provides a stable semantic benchmark for dynamic fidelity evaluation, reducing confidence evaluation error; global semantic distiller uses regularization loss to inversely constrain the prototype, slowing down the prototype drift speed and further stabilizing the mutual consistency alignment process.

[0104] Based on the same inventive concept, an embodiment of the present invention provides a federated graph learning system.

[0105] The federated graph learning system described in this invention can be installed in an electronic device. Depending on the functions implemented, the federated graph learning system includes one server and multiple clients. The server includes: The receiving unit is used to receive statistical information on category samples uploaded by each client; The weight generation unit is used to calculate the class weight vector based on the number of global class samples. A broadcasting unit is used to broadcast the category weight vector to each client. The distiller unit is used to establish a cross-domain consensus mapping model; The aggregation unit is used to aggregate the semantic center vectors uploaded by each client; The at least one client includes: A dual encoder unit is used to establish a topological context view encoder and an intrinsic semantic view encoder; The prototype management unit is used to maintain a semantic center vector for each category; An evaluation unit is used to calculate the prediction confidence and generate a fidelity mask. Training unit, used to train the topological context view encoder and the intrinsic semantic view encoder; The fusion unit is used to fuse the local semantic center vector and the global semantic center vector.

[0106] The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0107] The various variations and specific examples of the federated graph learning method provided in the above embodiments are also applicable to the federated graph learning system of this embodiment. Through the foregoing detailed description of the federated graph learning method, those skilled in the art can clearly understand the implementation method of the federated graph learning system in this embodiment. For the sake of brevity, they will not be described in detail here.

[0108] This application also discloses an electronic device, such as Figure 3The diagram shown is a schematic representation of an electronic device for a federated graph learning method according to an embodiment of the present invention. The electronic device may include at least one processor 10, a memory 11 communicatively connected to the at least one processor, a communication bus 12, and a communication interface 13. It may also include a computer program, such as a federated graph learning method program, stored in the memory 11 and executable on the processor 10.

[0109] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., methods for executing federated graph learning) and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0110] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as the code of federated graph learning methods, but also to temporarily store data that has been output or will be output.

[0111] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0112] Communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), and optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0113] Figure 3 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 3 The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0114] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power various components. Preferably, the power supply can be logically connected to at least one processor 10 via a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0115] It should be understood that the embodiments are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0116] Furthermore, if the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile.

[0117] This application provides a computer-readable storage medium, including, for example, any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM). The computer-readable storage medium stores a computer program that can be loaded by a processor and execute the federated graph learning method of the above embodiments.

[0118] In the description of this specification, the references to terms such as "an embodiment," "some embodiments," "example," "specific example," "a implementation," "a preferred implementation," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A federated graph learning method, said method being applied to a federated learning system comprising one server and multiple clients, each client holding private graph data, characterized in that, The method includes: The server aggregates the category sample statistics uploaded by each client and calculates the global category distribution based on the category sample statistics uploaded by each client. The server determines the category weight vector based on the global category distribution and broadcasts the category weight vector to each client; Each client establishes a topology context view encoder and an intrinsic semantic view encoder; Each client maintains a semantic center vector for each category and updates the semantic center vector through a momentum mechanism. Each client computes the prediction confidence of the intrinsic semantic view encoder on the validation set, dynamically generates a fidelity mask, and marks reliable nodes; Each client calculates the topology attempt loss and updates the model parameters of the local semantic essence prototype based on the category weight vector, fidelity mask, semantic center vector and node features of reliable nodes, and then uploads the model parameters to the server. The server establishes a cross-domain consensus mapping model, and constrains the semantic center vector of each client to approximate the global semantic manifold based on the model parameters uploaded by each client; The server aggregates the semantic center vectors uploaded by each client, generates a global category semantic center vector, and broadcasts it to each client. Each client integrates the local semantic center vector and the global semantic center vector.

2. The federated graph learning method as described in claim 1, characterized in that, The topology attempt loss includes mutual consistency alignment loss, classification loss, reverse alignment loss, and regularization loss.

3. The federated graph learning method as described in claim 2, characterized in that, Each client uses the category weight vector to train the topological context view encoder with weighted cross-entropy loss, and calculates the mutual consistency alignment loss under the fidelity mask constraint; Each client uses a learnable gating mechanism to fuse semantic center vectors and node features with the internal semantic view encoder, and calculates classification loss, back alignment loss, and regularization loss.

4. The federated graph learning method as described in claim 1, characterized in that, The model parameters uploaded by the client to the server include topology view parameters, intrinsic semantic view parameters, local SEP set, and local label count vector.

5. The federated graph learning method as described in any one of claims 1 to 4, characterized in that, The formula for calculating the category weight vector is: ； in, Indicates the first In each round of communication, the server calculates and broadcasts the global calibration adjustment weights corresponding to the category to each client. The weight values; C is the total number of categories; For category The intermediate weights; , For adjustment coefficients, The number of samples in the global category c.

6. The federated graph learning method as described in any one of claims 1 to 4, characterized in that, The node representation update formula for the topology context view encoder is: ； in, For point In the The layer's topological context view hides the representation, where AGG is an aggregation function. For activation function, For the first The learnable weight matrix of the layer, Represents a node All neighboring nodes In the The topological context view of a layer hides the collection that it represents; Representing neighboring nodes In the The topology context view of the layer hides the representation. For the first The learnable bias vector of the layer, For the local neighborhood of client k.

7. The federated graph learning method as described in any one of claims 1 to 4, characterized in that, The topological context view encoder is a graph neural network encoder, and the intrinsic semantic view encoder is a multilayer perceptron encoder.

8. A federated graph learning system for implementing the federated graph learning method according to any one of claims 1 to 7, characterized in that, Includes one server and multiple clients. The server includes: The receiving unit is used to receive statistical information on category samples uploaded by each client; The weight generation unit is used to calculate the class weight vector based on the number of global class samples. A broadcasting unit is used to broadcast the category weight vector to each client. The distiller unit is used to establish a cross-domain consensus mapping model; The aggregation unit is used to aggregate the semantic center vectors uploaded by each client; The at least one client includes: A dual encoder unit is used to establish a topological context view encoder and an intrinsic semantic view encoder; The prototype management unit is used to maintain a semantic center vector for each category; An evaluation unit is used to calculate the prediction confidence and generate a fidelity mask. Training unit, used to train the topological context view encoder and the intrinsic semantic view encoder; The fusion unit is used to fuse the local semantic center vector and the global semantic center vector.

9. An electronic device, characterized in that, The electronic device includes: At least one processor (10); and, A memory (11) communicatively connected to the at least one processor (10); The memory (11) stores a computer program that can be executed by the at least one processor (10) to enable the at least one processor (10) to perform the federated graph learning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program; when executed by a processor, the computer program implements the federated graph learning method as described in any one of claims 1 to 7.

Images