A Graph Convolutional Neural Network Federated Modeling Method

Abstract

The present invention relates to the field of graph neural network technology, specifically a graph convolutional neural network federated modeling method, comprising the following steps: extracting overall graph model structural features; selecting a common node subgraph; training a subgraph model for federated modeling; training an overall graph model; reasoning and prediction tasks; the beneficial effects are as follows: the graph convolutional neural network federated modeling method proposed in the present invention is targeted at real business complex modeling demand scenarios such as node overlap, fully considering the correlation between the graph data owned by the data owners participating in the modeling, and based on the strategy selection, adopting the Privacy Set Intersection (PSI) protocol to determine the common nodes, select the common nodes with high value, and perform secure joint modeling based on the subgraph formed by the selected common nodes, and then integrate them into the overall graph data structure to complete the secure modeling of the overall graph data. Compared with the traditional modeling method, the secure joint modeling method of the common node subgraph graph convolution can better explore the deep connection between institutional data.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and specifically to a graph convolutional neural network federated modeling method. Background Art

[0002] Convolutional Neural Networks (CNNs) can effectively improve the efficiency of reasoning tasks when processing data in Euclidean spaces such as text and images, but their application to non-Euclidean spatial data, such as graph-structured data, is less effective. With the increasing interest in graph-structured data, Graph Neural Networks (GNNs), as a deep learning method for processing graph data, have been widely applied to various neural network models. Among them, Graph Convolutional Networks (GCNs), as a key branch of GNNs, demonstrate their advantages in graph structure processing and have been successfully applied in fields such as recommendation systems, anti-financial fraud, traffic prediction, and genetic testing.

[0003] With the increasing complexity of business scenarios, existing technologies are posing new challenges to business modeling. The limitations of traditional independent modeling approaches are becoming apparent, making them unable to meet the business requirements of complex scenarios. Joint modeling requires integrating cross-institutional, cross-regional, and multi-source data. However, due to the distributed nature of data and the need to protect data privacy, new technical solutions are needed. Therefore, combining secure multi-party computing (MPC) technologies (such as Private Set Intersection (PSI), Oblivious Transfer (OT), and homomorphic encryption) with graph convolutional neural networks has become a research hotspot.

[0004] Currently, federated algorithms such as FedGCN and FedGNN have met some of the needs to a certain extent. However, in real business modeling scenarios, the graph data owned by the data owners involved in the modeling often have various connections, such as node overlap and edge overlap, and each modeling party has its own data labels. In this case, how to effectively integrate privacy computing technology and federated modeling technology to achieve secure and efficient joint modeling of graph convolutional neural networks in complex scenarios with overlapping data while protecting data privacy, and enhance business value has become an urgent problem that needs to be solved. Summary of the Invention

[0005] The purpose of the present invention is to provide a graph convolutional neural network federated modeling method to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a graph convolutional neural network federated modeling method, the method comprising the following steps:

[0007] Extract the structural features of the overall graph model;

[0008] Select the common node subgraph;

[0009] Training a model for federated modeling of subgraph models;

[0010] Train the overall graph model;

[0011] Inference prediction tasks.

[0012] Preferably, in step 101, according to the requirements of the business domain prediction task Pred, data features that affect the prediction task are selected;

[0013] Step 102: Each joint modeling party collects massive data in the business domain of the prediction task Pred, obtains the true label of the prediction task, and forms training data;

[0014] Step 103: Each joint modeling party i processes its own training data through the graph structure feature generator GSGen to generate a graph structure GS-i.

[0015] Preferably, selecting the common node subgraph comprises the following steps:

[0016] Step 201: The PSI module sets the rules for the intersection of private sets, and the public nodes of all participants are used as the result of the private intersection by default.

[0017] Step 202: The node selection module NSC sets a public node selection strategy, which may include parameters such as the number of public nodes, the proportion of public nodes, and priority node selection rules.

[0018] Step 203: Based on the node sets of the graph structure GS-i owned by each joint modeling party i, a secure set intersection of multiple parties is performed through a private set intersection (PSI) module to obtain a public node set.

[0019] Step 204: The node selection module NSC judges the public node set. If it does not meet the requirements of the node selection policy, the private set intersection rule is reset and the private set intersection PSI module reselects.

[0020] Step 205: If the set selection strategy is met, the node selection module NSC selects a public node set SelNodeSet;

[0021] Step 206: Change the public node selection strategy settings to find the optimal node selection strategy for the graph data.

[0022] Preferably, training the subgraph model federation modeling model includes the following steps:

[0023] Step 301: Each joint modeling party i processes its own graph structure GS-i through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i;

[0024] Step 302: The subgraph structure generator SubGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training.

[0025] Step 303: Each joint modeling party i performs a GCN graph convolution operation on the graph structure SelGS-i to form its own node embedding vector and relationship embedding vector;

[0026] Step 304: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method.

[0027] Step 305: input the embedding vector of the security aggregation result into the prediction task Pred output label;

[0028] Step 306: Calculate the prediction result output by the prediction task Pred and the actual label of the prediction task through the loss gradient calculation model update module CL, and back-propagate the error to update the model;

[0029] Step 307: The node vector distance calculation (NVC) module sets a triple loss function to calculate the cosine similarity of the embedding vectors of the common nodes of all parties, so as to minimize the node embedding vector distance between the same nodes and maximize the embedding vector distance between different nodes. The gradient is then back-propagated to update the graph structure generator SubGSGen.

[0030] Step 308: Repeat steps 302 to 307, and each joint modeling party i obtains the final common node subgraph structure SelGS-i, including its node embedding vector and relationship embedding vector.

[0031] Preferably, training the overall graph model comprises the following steps:

[0032] Step 401: Each joint modeling party i processes its own graph structure GS-i and public node subgraph structure SelGS-i through the fusion graph structure generator FusGSGen, and combines it with the selected public node set SelNodeSet to generate a graph structure newGS-i;

[0033] Step 402: The fusion graph structure generator FusGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training.

[0034] Step 403: Each joint modeling party i performs a GCN graph convolution operation on the graph structure newGS-i to form its own node embedding vector and relationship embedding vector;

[0035] Step 404: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method.

[0036] Step 405: Input the embedding vector of the security aggregation result into the prediction task Pred output label;

[0037] Step 406: The loss gradient calculation model update module CL calculates the prediction result output by the prediction task Pred and the actual label of the prediction task, and backpropagates the error to update the model;

[0038] Step 407: The public node vector update module NUPD sets a node embedding vector update gradient threshold and back-propagates the gradient to limit the range of the public node embedding vector of the updated graph structure newGS.

[0039] Step 408: Repeat steps 402 to 407, and each joint modeling party i obtains the final common node subgraph structure newGS-i, including its node embedding vector and relationship embedding vector.

[0040] Preferably,.

[0041] Preferably, the inference prediction task includes the following steps:

[0042] Step 501: Joint modeling party i generates a graph structure GS-i through the graph feature generator GSGen based on real data;

[0043] Step 502: Generate a node set SelNodeSet through the privacy set intersection PSI module and the node selection module;

[0044] Step 503: Joint modeling party i processes the graph structure GS-i owned by each of them through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i;

[0045] Step 504: The joint modeling party i processes the graph structure GS-i and the common node subgraph structure SelGS-i owned by each party through the fusion graph structure generator FusGSGen, and combines the selected common node set SelNodeSet to generate a graph structure newGS-i;

[0046] Step 504: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are subjected to weighted average security aggregation based on the secret sharing method by the security aggregation module SF, and the prediction result is output through the prediction task Pred;

[0047] Step 505: All parties continue to collect data and provide feedback on the correctness of the output results for continuous optimization of the model.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The graph convolutional neural network federated modeling method proposed in this paper targets complex modeling scenarios in real-world businesses, such as node overlap. It fully considers the relevance of the graph data owned by the participating data owners. Based on policy selection, it uses the Privacy Set Intersection (PSI) protocol to determine common nodes, select high-value common nodes, and then securely model the subgraph formed by the selected common nodes. This subgraph is then integrated into the overall graph data structure to complete the secure modeling of the overall graph data. Compared to traditional modeling methods, secure joint modeling using graph convolution of common node subgraphs can better uncover the deep connections between institutional data, fully utilize the data of each participant, and improve the accuracy of model predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a structural diagram of the convolutional neural network federation model of the present invention. DETAILED DESCRIPTION

[0051] In order to clearly and completely describe the objectives and technical solutions of the present invention and make the advantages more clearly understood, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, not all of them, and are only used to explain the embodiments of the present invention, not to limit the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] Example 1

[0053] See also Figure 1 The present invention provides a technical solution: a graph convolutional neural network federated modeling method, the method comprising the following steps:

[0054] Extracting the structural features of the overall graph model; including the following steps:

[0055] Step 101: Select data features that affect the prediction task based on the requirements of the business domain prediction task Pred;

[0056] Step 102: Each joint modeling party collects massive data in the business domain of the prediction task Pred, obtains the true label of the prediction task, and forms training data;

[0057] Step 103: Each joint modeling party i processes its own training data through the graph structure feature generator GSGen to generate a graph structure GS-i;

[0058] Select the common node subgraph; including the following steps:

[0059] Step 201: The PSI module sets the rules for the intersection of private sets, and the public nodes of all participants are used as the result of the private intersection by default.

[0060] Step 202: The node selection module NSC sets a public node selection strategy, which may include parameters such as the number of public nodes, the proportion of public nodes, and priority node selection rules.

[0061] Step 203: Based on the node sets of the graph structure GS-i owned by each joint modeling party i, a secure set intersection of multiple parties is performed through a private set intersection (PSI) module to obtain a public node set.

[0062] Step 204: The node selection module NSC judges the public node set. If it does not meet the requirements of the node selection policy, the private set intersection rule is reset and the private set intersection PSI module reselects.

[0063] Step 205: If the set selection strategy is met, the node selection module NSC selects a public node set SelNodeSet;

[0064] Step 206: Change the public node selection strategy settings to find the optimal node selection strategy for the graph data;

[0065] Training a model for federated subgraph modeling includes the following steps:

[0066] Step 301: Each joint modeling party i processes its own graph structure GS-i through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i;

[0067] Step 302: The subgraph structure generator SubGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training.

[0068] Step 303: Each joint modeling party i performs a GCN graph convolution operation on the graph structure SelGS-i to form its own node embedding vector and relationship embedding vector;

[0069] Step 304: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method.

[0070] Step 305: input the embedding vector of the security aggregation result into the prediction task Pred output label;

[0071] Step 306: Calculate the prediction result output by the prediction task Pred and the actual label of the prediction task through the loss gradient calculation model update module CL, and back-propagate the error to update the model;

[0072] Step 307: The node vector distance calculation (NVC) module sets a triple loss function to calculate the cosine similarity of the embedding vectors of the common nodes of all parties, so as to minimize the node embedding vector distance between the same nodes and maximize the embedding vector distance between different nodes. The gradient is then back-propagated to update the graph structure generator SubGSGen.

[0073] Step 308: Repeat steps 302 to 307, and each joint modeling party i obtains the final common node subgraph structure SelGS-i, including its node embedding vector and relationship embedding vector;

[0074] Training the overall graph model; including the following steps:

[0075] Step 401: Each joint modeling party i processes its own graph structure GS-i and public node subgraph structure SelGS-i through the fusion graph structure generator FusGSGen, and combines it with the selected public node set SelNodeSet to generate a graph structure newGS-i;

[0076] Step 402: The fusion graph structure generator FusGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training.

[0077] Step 403: Each joint modeling party i performs a GCN graph convolution operation on the graph structure newGS-i to form its own node embedding vector and relationship embedding vector;

[0078] Step 404: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method.

[0079] Step 405: Input the embedding vector of the security aggregation result into the prediction task Pred output label;

[0080] Step 406: The loss gradient calculation model update module CL calculates the prediction result output by the prediction task Pred and the actual label of the prediction task, and backpropagates the error to update the model;

[0081] Step 407: The public node vector update module NUPD sets a node embedding vector update gradient threshold and back-propagates the gradient to limit the range of the public node embedding vector of the updated graph structure newGS.

[0082] Step 408: Repeat steps 402 to 407 until each joint modeling party i obtains the final common node subgraph structure newGS-i, including its node embedding vectors and relationship embedding vectors;

[0083] Inference prediction task; includes the following steps:

[0084] Step 501: Joint modeling party i generates a graph structure GS-i through the graph feature generator GSGen based on real data;

[0085] Step 502: Generate a node set SelNodeSet through the privacy set intersection PSI module and the node selection module;

[0086] Step 503: Joint modeling party i processes the graph structure GS-i owned by each of them through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i;

[0087] Step 504: The joint modeling party i processes the graph structure GS-i and the common node subgraph structure SelGS-i owned by each party through the fusion graph structure generator FusGSGen, and combines the selected common node set SelNodeSet to generate a graph structure newGS-i;

[0088] Step 504: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are subjected to weighted average security aggregation based on the secret sharing method by the security aggregation module SF, and the prediction result is output through the prediction task Pred;

[0089] Step 505: All parties continue to collect data and provide feedback on the correctness of the output results for continuous optimization of the model.

[0090] Example 2

[0091] Based on Example 1, a graph convolutional neural network federated modeling method forms a graph structure by a graph structure feature generator based on the correlation of graph data owned by each data owner participating in the modeling, and determines common nodes through the Privacy Set Intersection (PSI) protocol. Common nodes with high value are selected according to the strategy, and then the subgraph formed by the selected common nodes themselves is used for secure joint modeling, and its public graph node embedding vector is calculated to form an initial graph structure representation; the variation range of the public graph embedding vector is set, and the secret sharing technology is combined to realize the global graph convolution operation, and the gradient is calculated and the model parameters are updated through the known node-level task labels and graph structure-level task labels to form a secure joint modeling of a multi-party graph convolutional neural network model; the node selection strategy is adjusted multiple times to form the optimal joint modeling model for application to node-level and graph structure-level reasoning tasks.

[0092] Example 3

[0093] Based on the second embodiment, a graph convolutional neural network federated modeling method is provided, wherein the training data comes from each joint modeling participant and is stored locally on the node of the joint modeling participant; the graph structure GS is a graph relationship structure (V, E, v, e) formed by the training data through a graph structure generator, wherein V represents the nodes of the graph, E represents the relationship between the graph nodes, v is the feature group of the nodes, e is the feature group of the node relationship, and v and e are represented by multi-dimensional feature vectors; the core of the graph structure feature generator GSGen is a neural network, which is responsible for generating the graph structure GS while normalizing the original data to generate feature vectors for the nodes and relationships; the privacy set The intersection PSI module is responsible for calculating and obtaining the public nodes of all participants in the joint modeling; the node selection module NSC is responsible for selecting a public node set SelNodeSet that meets the requirements according to the set strategy. The public node set SelNodeSet is a subset of all public nodes; the subgraph structure generator SubGSGen, whose core is a neural network, is responsible for generating a graph structure SelGS based on the selected public node set SelNodeSet and updating the node features and relationship features of its graph structure; the graph convolution operation GCN adopts the GAT algorithm to perform graph convolution operations on the graph structure data to form node embedding vectors and node relationship embedding vectors; The security aggregation module SF is responsible for sharding the embedding vectors of each participant in the joint modeling through secret sharing, and performing weighted average security aggregation of data at the security node containing the prediction task label; the node vector distance calculation module NVC adopts TripleLoss loss, and calculates the distance between the node embedding vectors of the common nodes of different participants to minimize the node embedding vector distance between the same nodes and maximize the embedding vector distance between different nodes; the prediction task Pred is an analysis and judgment task of node data and graph data, which includes node-level tasks and graph-level tasks, and outputs prediction labels based on node embedding vectors and node relationship embedding vectors; The loss gradient calculation model update module CL is responsible for calculating the loss gradient between the predicted label output by the prediction task Pred and the true label of the prediction task, and updating the neural network model parameters; the fusion graph structure generator FusGSGen, whose core is a neural network, is responsible for generating the graph structure newGS based on the generated graph structure GS and the selected public node set SelNodeSet, and generating the node features and relationship features of the newGS graph structure; the public node vector update module NUPD is responsible for updating the embedding vector of the public node set SelNodeSet to ensure that the node embedding vector update gradient is less than the set threshold.

[0094] For scenarios with complex modeling requirements for real business, such as node overlap, the correlation between the graph data owned by the data owners participating in the modeling is fully considered. According to the strategy selection, the privacy set intersection (PSI) protocol is used to determine the common nodes, select the common nodes with high value, and perform secure joint modeling based on the subgraph formed by the selected common nodes. The subgraph is then integrated into the overall graph data structure to complete the secure modeling of the overall graph data. Compared with the traditional modeling method, the secure joint modeling method of the common node subgraph convolution can better mine the deep connection between the institutional data, and make full use of the data of each participant, thereby improving the accuracy of the model prediction; at the same time, the method of subgraph secure modeling followed by overall joint modeling can reduce the communication volume of the data exchanged by the parties, while accelerating the convergence speed of the joint modeling. In addition, by adjusting the node selection strategy multiple times, the optimal joint modeling model is formed, and feedback is continuously collected to continuously optimize and improve the model accuracy. In practical applications, the method of the present invention has a very broad application prospect, can better cope with scenarios with high business complexity and high privacy requirements, has a wide range of applications, and has very good practical effects.

[0095] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A graph convolutional neural network federated modeling method, characterized by: The method comprises the following steps: Extract the structural features of the overall graph model; Select the common node subgraph; Training a model for federated modeling of subgraph models; Train the overall graph model; Reasoning and prediction tasks; Extracting the structural features of the overall graph model includes the following steps: Step 101: Select data features that affect the prediction task based on the requirements of the business domain prediction task Pred; Step 102: Each joint modeling party collects massive data in the business domain of the prediction task Pred, obtains the true label of the prediction task, and forms training data; Step 103: Each joint modeling party i processes its own training data through the graph structure feature generator GSGen to generate a graph structure GS-i; Selecting a common node subgraph involves the following steps: Step 201: The PSI module sets the rules for the intersection of private sets, and the public nodes of all participants are used as the result of the private intersection by default. Step 202: The node selection module NSC sets a public node selection strategy, including the number of public nodes, the proportion of public nodes, and priority node selection rule parameters. Step 203: Based on the node sets of the graph structure GS-i owned by each joint modeling party i, a secure set intersection of multiple parties is performed through the private set intersection (PSI) module to obtain a public node set. Step 204: The node selection module NSC judges the public node set. If it does not meet the requirements of the node selection policy, the private set intersection rule is reset and the private set intersection PSI module reselects. Step 205: If the set selection strategy is met, the node selection module NSC selects a public node set SelNodeSet; Step 206: Change the public node selection strategy settings to find the optimal node selection strategy for the graph data; Training the subgraph model for federated modeling involves the following steps: Step 301: Each joint modeling party i processes its own graph structure GS-i through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i; Step 302: The subgraph structure generator SubGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training. Step 303: Each joint modeling party i performs a GCN graph convolution operation on the graph structure SelGS-i to form its own node embedding vector and relationship embedding vector; Step 304: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method. Step 305: input the embedding vector of the security aggregation result into the prediction task Pred output label; Step 306: The loss gradient calculation model update module CL calculates the prediction result output by the prediction task Pred and the actual label of the prediction task, and backpropagates the error to update the model; Step 307: The node vector distance calculation NVC module sets the triple loss function to calculate the cosine similarity distance of the embedding vectors of the common nodes of all parties, so as to minimize the node embedding vector distance between the same nodes and maximize the embedding vector distance between different nodes. The gradient is then back-propagated to update the graph structure generator SubGSGen. Step 308: Repeat steps 302 to 307, and each joint modeling party i obtains the final common node subgraph structure SelGS-i, including its node embedding vector and relationship embedding vector; Training the overall graph model involves the following steps: Step 401: Each joint modeling party i processes its own graph structure GS-i and public node subgraph structure SelGS-i through the fusion graph structure generator FusGSGen, and combines it with the selected public node set SelNodeSet to generate a graph structure newGS-i; Step 402: The fusion graph structure generator FusGSGen, the graph convolution operation, the security aggregation module SF, and the prediction task Pred are combined into a model for training. Step 403: Each joint modeling party i performs a GCN graph convolution operation on the graph structure newGS-i to form its own node embedding vector and relationship embedding vector; Step 404: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are aggregated by weighted average security using a secure aggregation module SF based on a secret sharing method. Step 405: Input the embedding vector of the security aggregation result into the prediction task Pred output label; Step 406: The loss gradient calculation model update module CL calculates the prediction result output by the prediction task Pred and the actual label of the prediction task, and backpropagates the error to update the model; Step 407: The public node vector update module NUPD sets a node embedding vector update gradient threshold and back-propagates the gradient to limit the range of the public node embedding vector of the updated graph structure newGS. Step 408: Repeat steps 402 to 407 until each joint modeling party i obtains the final common node subgraph structure newGS-i, including its node embedding vectors and relationship embedding vectors; The inference prediction task includes the following steps: Step 501: Joint modeling party i generates a graph structure GS-i through the graph feature generator GSGen based on real data; Step 502: Generate a node set SelNodeSet through the privacy set intersection PSI module and the node selection module; Step 503: The joint modeling party i processes the graph structure GS-i owned by each party through the subgraph structure feature generator SubGSGen, combines and selects the common node set SelNodeSet, and generates the graph structure SelGS-i; Step 504: The joint modeling party i processes the graph structure GS-i and the common node subgraph structure SelGS-i owned by each party through the fusion graph structure generator FusGSGen, and combines the selected common node set SelNodeSet to generate the graph structure newGS-i; Step 504: The node embedding vectors and relationship embedding vectors formed by each joint modeling party are subjected to weighted average security aggregation based on the secret sharing method by the security aggregation module SF, and the prediction result is output through the prediction task Pred; Step 505: All parties continue to collect data and provide feedback on the correctness of the output results for continuous optimization of the model.

Images