Client subgraph repair method, device and equipment based on beidou timing and medium

Abstract

The application discloses a Beidou timing-based client subgraph repairing method, device, equipment and medium, which comprises the following steps: acquiring a subgraph of a client; using a Beidou timing system to perform time sequence synchronization on local data of the client, so as to ensure that the subgraph and other subgraphs of a global graph are in the same time sequence; deploying a central generator, a discriminator and a local generator; training the local generator, generating first missing node features of the subgraph through the trained local generator; generating second missing node features of the subgraph through a trained central generator; repairing missing neighbor information of the subgraph based on the first missing node features and the second missing node features; training a local node classifier based on the repaired subgraph, generating local parameters, and sending the local parameters to a server, so that the server trains a global node classifier based on the local parameters. The application can repair the client subgraph and improve the performance of the global node classifier.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a client sub-image repair method, apparatus, device and storage medium based on BeiDou time synchronization. Background Technology

[0002] With the rapid development and application of artificial intelligence technology, the scale and complexity of datasets in target systems are constantly increasing. Federated graph learning is a novel paradigm for addressing the privacy protection needs of target systems. Its main function is to solve the data silo problem while utilizing graph neural networks to process complex graph-structured data and better extract features between data points.

[0003] For ease of explanation, we will use a financial system as an example, and refer to [reference needed]. Figure 1 , Figure 1 Figure 1 This is a sample of an existing global graph, which is divided into multiple subgraphs, each corresponding to a bank's dataset. Due to privacy protection requirements, different banks cannot exchange any information, resulting in missing client information in the subgraphs. Therefore, the subgraphs need to be repaired, which requires data from all subgraphs. Thus, how to repair lost information across clients without compromising privacy is a challenging problem.

[0004] Furthermore, how to recover lost information across clients without compromising privacy is also a pressing issue that needs to be addressed, as cross-client information loss in federated learning can have adverse effects on fields such as finance and healthcare, as detailed below:

[0005] Risk assessment and fraud detection: Neighbor relationships in the global graph can provide more comprehensive information, helping banks conduct risk assessment and fraud detection. However, due to privacy protection requirements, these neighbor relationships disappear between subgraphs, resulting in insufficient graph data representation. This may cause banks to miss some important signals and patterns when assessing risk and detecting fraud, thereby reducing their accuracy and effectiveness.

[0006] Cross-bank collaboration and data sharing: Restricted information exchange between banks can pose challenges to cross-bank collaboration and data sharing. In the financial healthcare sector, cross-bank data sharing can help identify potential financial fraud and support precise financial billing management. The lack of clear communication channels between banks, hindering full information sharing, may limit the potential and benefits of cross-bank collaboration.

[0007] Data Analysis and Model Performance: The performance of global and local node classifiers may be affected. Global node classifiers are typically trained on the entire global graph, which contains more comprehensive information. However, due to broken neighbor relationships, global node classifiers may fail to capture cross-bank associations and patterns, thus reducing their predictive and analytical capabilities. Local node classifiers, on the other hand, are trained only on subgraphs, lacking the support of global information, which may lead to performance degradation or increased limitations.

[0008] Challenges in Disease Transmission Surveillance: In the medical field, understanding disease transmission patterns and trends is crucial for effective prevention, control, and treatment. Disconnections in the global graph can lead to inaccurate representation of disease transmission relationships between different institutions within subgraphs, thus impacting epidemic monitoring and control. This can result in underutilization of data in both global and local node classifiers, thereby reducing the ability to understand and respond to disease transmission.

[0009] Lack of Data Mining and Insights: Global graphs offer a broader data perspective, revealing interactions and influences between institutions. However, due to broken links, local node classifiers can only see limited information in subgraphs, leading to a lack of data mining and insights. This can hinder the discovery of new patterns, trends, and correlations, limiting the ability to make accurate predictions and decisions in the financial and healthcare sectors.

[0010] Therefore, how to recover lost information across clients without compromising privacy is a difficult problem and an urgent issue to be addressed. Summary of the Invention

[0011] This application provides a client subgraph repair method, apparatus, device, and storage medium based on BeiDou time synchronization to solve the problems of how to repair lost information across clients without disclosing privacy and how to use the repaired client subgraph to train a global node classifier.

[0012] In a first aspect, embodiments of this application provide a client sub-image repair method based on BeiDou time synchronization, the client sub-image repair method based on BeiDou time synchronization includes:

[0013] Obtain a subgraph of the client, wherein the subgraph is formed by partitioning the global graph of the target system, and the target system is a federated learning-based system;

[0014] The BeiDou timing system is used to synchronize the local data of this client in time series, ensuring that the subgraph is in the same time series as other subgraphs of the global graph;

[0015] A central generator of a generative adversarial neural network is deployed on the server side participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client side participating in the federated learning, and a local generator of a graph neural network is deployed on the client side.

[0016] The local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator.

[0017] The central generator and the discriminator are trained together, and the trained central generator generates the second missing node feature of the subgraph.

[0018] Based on the first missing node features and the second missing node features of the subgraph, the missing neighbor information of the subgraph is repaired;

[0019] The local node classifier of the federated learning is trained on the subgraph after the neighbor information is repaired, and local parameters of the local node classifier are generated. The local parameters are sent to the server so that the server can train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier on the node classification task.

[0020] As an optional implementation, training the local generator by occluding the subgraph and restoring the occluded subgraph, and generating the first missing node feature of the subgraph using the trained local generator, includes:

[0021] A occluded subgraph is input into the untrained local generator, and a graph representation is obtained through the graph neural network. The graph representation is then sent to a multilayer perceptron, which performs classification training on the number of occluded nodes, the number of predicted nodes, the occluded node features, and the predicted node features of the occluded subgraph.

[0022] After training is complete, the unmasked subgraph is input into the trained local generator, and the trained local generator generates the first missing node feature of the subgraph.

[0023] As an optional implementation, the step of co-training the central generator and the discriminator to generate the second missing node feature of the subgraph using the trained central generator includes:

[0024] Gaussian white noise is input into the central generator, which generates the initial missing node features of the subgraph. The initial missing node features are then input into the discriminator.

[0025] The evaluation result is obtained through the discriminator, and the discriminator is updated at the same time. The evaluation result includes the probability of identifying the initial missing node feature as the real node feature.

[0026] The central generator is updated based on the evaluation results. The loss function between the updated discriminator and the updated central generator is obtained. Multiple loss functions are aggregated. The central generator is iteratively trained using the aggregated loss function until a preset condition is met, at which point the training of the central generator is stopped.

[0027] The second missing node feature of the subgraph is generated by the trained central generator.

[0028] As an optional implementation, the preset conditions include one or a combination of the following: the number of training iterations has reached a preset number, the training time has reached a preset time, and the loss function has been minimized to make the adversarial neural network reach a convergence state.

[0029] As an optional implementation, the global node classifier is a node classifier used to predict user behavior and user preferences in the financial system.

[0030] As an optional implementation, the global node classifier is a node classifier used to predict user behavior and user preferences in the medical system.

[0031] As an alternative implementation, the target system includes one or a combination of financial systems and medical systems.

[0032] Secondly, this application also provides a client sub-image repair device based on BeiDou time synchronization. The client sub-image repair device based on BeiDou time synchronization includes a processor and a memory. The memory stores a computer program. When the processor calls the computer program in the memory, it executes the above-mentioned client sub-image repair method based on BeiDou time synchronization.

[0033] Thirdly, embodiments of this application also provide a device, which includes the client sub-image repair device based on BeiDou time synchronization as described above.

[0034] Fourthly, embodiments of this application also provide a storage medium for storing a computer program, which, when executed by a processor, enables the processor to implement the aforementioned client sub-graph repair method based on BeiDou time synchronization.

[0035] This application provides a client sub-image repair method, apparatus, device, and storage medium based on BeiDou time synchronization. The method includes:

[0036] Obtain a subgraph of the client, wherein the subgraph is formed by partitioning the global graph of the target system, and the target system is a federated learning-based system;

[0037] The BeiDou timing system is used to synchronize the local data of this client in time series, ensuring that the subgraph is in the same time series as other subgraphs of the global graph;

[0038] A central generator of a generative adversarial neural network is deployed on the server side participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client side participating in the federated learning, and a local generator of a graph neural network is deployed on the client side.

[0039] The local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator.

[0040] The central generator and the discriminator are trained together, and the trained central generator generates the second missing node feature of the subgraph.

[0041] Based on the first missing node features and the second missing node features of the subgraph, the missing neighbor information of the subgraph is repaired;

[0042] The local node classifier of the federated learning is trained on the subgraph after the neighbor information is repaired, and local parameters of the local node classifier are generated. The local parameters are sent to the server so that the server can train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier on the node classification task.

[0043] The beneficial effects of this application's embodiments are twofold. Firstly, by using the BeiDou time synchronization system, it ensures that data from different clients are in the same time series, thereby more accurately predicting user behavior and preferences, which is beneficial for improving the accuracy of client subgraph repair based on BeiDou time synchronization. Secondly, the local node classifier of the federated learning is trained based on the subgraph after repairing the neighbor information, generating local parameters of the local node classifier. These local parameters are then sent to the server, enabling the server to train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier in node classification tasks. Since only local parameters are uploaded and not local client data, cross-client lost information repair can be performed without disclosing privacy, solving the problem of how to perform cross-client lost information repair without disclosing privacy, and also solving the problem of how to train a global node classifier using a subgraph after repairing client lost information. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is a sample diagram of an existing global graph;

[0046] Figure 2 This is a flowchart illustrating the client-side sub-graph repair method based on BeiDou time synchronization provided in this application embodiment;

[0047] Figure 3 This is a flowchart illustrating the training of a local generator provided in an embodiment of this application;

[0048] Figure 4 This is a flowchart of the training central generator provided in an embodiment of this application;

[0049] Figure 5 This is a schematic block diagram of a client-side sub-image repair device based on BeiDou time synchronization provided in the embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0051] It should be noted that the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0052] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," "in neighbor information embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0053] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0054] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or neighbor information merged, so the actual execution order may change depending on the actual situation.

[0055] Please see Figure 2 , Figure 2 This is a flowchart illustrating the client subgraph repair method based on BeiDou time synchronization provided in this application embodiment. This method can be applied to devices participating in federated learning. The device can be any of the following: server, mobile phone, camera, tablet computer, wearable device, vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, personal computer (PC), netbook, personal digital assistant (PDA). No restrictions are imposed in this application embodiment.

[0056] like Figure 2 As shown in the embodiments of this application, the client sub-image repair method based on BeiDou time synchronization includes the following steps, detailed below:

[0057] S201, Obtain the client's subgraph, which is formed by partitioning the global graph of the target system, wherein the target system is a federated learning-based system;

[0058] The target system is a federated learning-based system, comprising a server and multiple clients. The server, also known as the central server, maintains a global node classification model but does not store data.

[0059] The target system includes either a federated learning-based financial system or a federated learning-based healthcare system, or a combination thereof.

[0060] S202, use the BeiDou timing system to synchronize the local data of this client in time sequence to ensure that the subgraph is in the same time series as other subgraphs of the global graph;

[0061] The process involves the client sending its local data timing sequence to the BeiDou timing system. Once the BeiDou timing system confirms that the client's timing sequence is consistent, it sends a response signal to the client.

[0062] S203, Deploy a central generator of the generative adversarial neural network on the server side participating in the federated learning, deploy a discriminator of the generative adversarial neural network on the client side participating in the federated learning, and deploy a local generator of the graph neural network on the client side.

[0063] Specifically, a central generator of a generative adversarial neural network is deployed on the central server participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client participating in the federated learning, and a local generator of a graph neural network is deployed on the client.

[0064] The relationship between the central generator and the discriminator is one-to-many, meaning that one central generator corresponds to multiple discriminators.

[0065] S204, the local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator;

[0066] S205, the central generator and the discriminator are trained together, and the second missing node feature of the subgraph is generated by the trained central generator;

[0067] S205 includes:

[0068] Obtain the node features of the subgraph, train the generative adversarial neural network constructed by the central generator and the discriminator using the node features of the subgraph, and generate the first missing node features of the subgraph using the central generator of the trained generative adversarial neural network.

[0069] The process of generating the first missing node feature of the subgraph through the central generator of the trained adversarial neural network includes:

[0070] Gaussian white noise is input into the central generator, which generates the initial missing node features of the subgraph. The initial missing node features are then input into the discriminator.

[0071] The evaluation result is obtained through the discriminator, and the discriminator is updated at the same time. The evaluation result includes the probability of identifying the initial missing node feature as the real node feature.

[0072] The central generator is updated based on the evaluation results. A loss function between the updated discriminator and the updated central generator is obtained. Multiple loss functions are aggregated. The central generator is iteratively trained using the aggregated loss function until a preset condition is met, at which point the training of the central generator stops. The second missing node feature of the subgraph is generated using the trained central generator.

[0073] The preset conditions include one or a combination of the following: the number of training iterations has reached a preset number, the training time has reached a preset time, and the loss function has been minimized to make the adversarial neural network reach a convergence state.

[0074] The server uses a broadcast method to send the first missing node feature of the subgraph to the client.

[0075] Among them, the first missing node feature is the potential first node feature.

[0076] Among them, the second missing node feature is the potential second node feature.

[0077] S206, Based on the first missing node features and the second missing node features of the subgraph, repair the missing neighbor information of the subgraph;

[0078] It should be noted that S206 includes:

[0079] The client generates missing node features of the subgraph based on the missing node feature generation model, the first missing node feature of the subgraph, and the second missing node feature of the subgraph, and repairs the missing neighbor information of the subgraph based on the missing node features of the subgraph.

[0080] The missing node feature generation model includes:

[0081]

[0082] Among them, Fnew jβ represents the missing node feature of the j-th missing node in the client. j D(β) represents the first missing node feature of the j-th missing node of the client generated by the central generator. j ) represents the β j The probability of identifying a feature as a real node; α j D(α) represents the second missing node feature of the j-th missing node generated by the local generator for the client. j ) represents the α j The probability of identifying a feature as a real node, wherein the local generator is a generator of a graph neural network deployed on the client.

[0083] S207, the local node classifier of the federated learning is trained based on the subgraph after the neighbor information is repaired, local parameters of the local node classifier are generated, and the local parameters are sent to the server so that the server trains the global node classifier of the federated learning based on the local parameters to improve the performance of the global node classifier on the node classification task.

[0084] The global node classifier is a node classifier used to predict user behavior and user preferences in the financial system.

[0085] The global node classifier is a node classifier used to predict user behavior and user preferences in the medical system.

[0086] Compared with existing solutions, the embodiments of this application have the following advantages:

[0087] 1. Improve the accuracy of client sub-graph repair based on BeiDou time synchronization: By using the BeiDou time synchronization system, it is ensured that data from different clients are in the same time series, thereby more accurately predicting user behavior and preferences.

[0088] 2. Effectively solves the problem of cross-client information loss: Collaborative training generates missing node features of the global graph, and enhances the data of local subgraphs through missing node features, thereby improving the performance of the global node classifier.

[0089] 3. Reduce the risk of privacy leaks: No local data transfer is required, which reduces the possibility of data leaks and improves data privacy performance.

[0090] Please see Figure 3 , Figure 3 The flowchart for training a local generator provided in this application embodiment is described in detail below:

[0091] S301, a occluded subgraph is input to the untrained local generator, a graph representation is obtained through the graph neural network, the graph representation is sent to a multilayer perceptron, and the multilayer perceptron is used to classify and train the number of occluded nodes, the number of predicted nodes, the occluded node features, and the predicted node features of the occluded subgraph.

[0092] S302, after training is completed, the unmasked sub-image is input into the trained local generator, and the first missing node feature of the sub-image is generated by the trained local generator.

[0093] In this embodiment, based on the first missing node feature and the second missing node feature of the subgraph, the missing neighbor information of the subgraph is repaired. The subgraph after repairing the neighbor information is more complete and comprehensive than the subgraph before repair. Therefore, the local parameters trained based on the subgraph after repairing the neighbor information are also more complete and comprehensive. Thus, using the local parameters trained based on the subgraph after repairing the neighbor information in the training of the global node classifier is beneficial to improving the performance of the global node classifier in the node classification task.

[0094] Please see Figure 4 , Figure 4 The flowchart of the training central generator provided in this application embodiment is described in detail below:

[0095] S401, Input Gaussian white noise into the central generator, generate the initial missing node features of the subgraph through the central generator, and input the initial missing node features into the discriminator;

[0096] S402, obtain the evaluation result through the discriminator, and update the discriminator at the same time. The evaluation result includes the probability of identifying the initial missing node feature as the real node feature.

[0097] S403, update the central generator based on the evaluation result, obtain the loss function between the updated discriminator and the updated central generator, aggregate multiple loss functions, and use the aggregated loss function to iteratively train the central generator until a preset condition is met, then stop training the central generator.

[0098] S404, the second missing node feature of the subgraph is generated by the trained central generator.

[0099] In this embodiment, the second missing node feature of the subgraph is generated by the trained central generator, which is beneficial for subsequently repairing the missing neighbor information of the subgraph based on the missing node features of the global graph. Then, the generated missing neighbor information is fused with the local subgraph data to obtain a more complete and comprehensive node feature representation.

[0100] Please see Figure 5 , Figure 5 This is a schematic block diagram of a client-side sub-image repair device based on BeiDou time synchronization provided in the embodiments of this application.

[0101] like Figure 5 As shown, the client sub-map repair device 200 based on BeiDou time synchronization may include a processor 211 and a memory 212. The processor 211 and the memory 212 are connected by a bus, such as an I2C (Inter-integrated Circuit) bus.

[0102] Specifically, the processor 211 can be a microcontroller unit (MCU), a central processing unit (CPU), or a digital signal processor (DSP), etc.

[0103] Specifically, the memory 212 can be a Flash chip, a read-only memory (ROM) disk, an optical disk, a USB flash drive, or a portable hard drive, etc. The memory 212 stores various computer programs for the processor 211 to execute.

[0104] The processor 211 is used to run a computer program stored in the memory processor 211, and performs the following steps when executing the computer program:

[0105] Obtain a subgraph of the client, wherein the subgraph is formed by partitioning the global graph of the target system, and the target system is a federated learning-based system;

[0106] The BeiDou timing system is used to synchronize the local data of this client in time series, ensuring that the subgraph is in the same time series as other subgraphs of the global graph;

[0107] A central generator of a generative adversarial neural network is deployed on the server side participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client side participating in the federated learning, and a local generator of a graph neural network is deployed on the client side.

[0108] The local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator.

[0109] The central generator and the discriminator are trained together, and the trained central generator generates the second missing node feature of the subgraph.

[0110] Based on the first missing node features and the second missing node features of the subgraph, the missing neighbor information of the subgraph is repaired;

[0111] The local node classifier of the federated learning is trained on the subgraph after the neighbor information is repaired, and local parameters of the local node classifier are generated. The local parameters are sent to the server so that the server can train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier on the node classification task.

[0112] In some embodiments, processor 211 is configured to implement:

[0113] Gaussian white noise is input into the central generator, which generates the initial missing node features of the subgraph. The initial missing node features are then input into the discriminator.

[0114] The evaluation result is obtained through the discriminator, and the discriminator is updated at the same time. The evaluation result includes the probability of identifying the initial missing node feature as the real node feature.

[0115] The central generator is updated based on the evaluation results. The loss function between the updated discriminator and the updated central generator is obtained. Multiple loss functions are aggregated. The central generator is iteratively trained using the aggregated loss function until a preset condition is met, at which point the training of the central generator is stopped.

[0116] The second missing node feature of the subgraph is generated by the trained central generator.

[0117] In some embodiments, processor 211 is configured to implement:

[0118] The preset conditions include one or a combination of the following: the number of training iterations reaches a preset number, the training time reaches a preset time, and the loss function is minimized to make the adversarial neural network reach a convergence state.

[0119] In some embodiments, processor 211 is configured to implement:

[0120] A occluded subgraph is input into the untrained local generator, and a graph representation is obtained through the graph neural network. The graph representation is then sent to a multilayer perceptron, which performs classification training on the number of occluded nodes, the number of predicted nodes, the occluded node features, and the predicted node features of the occluded subgraph.

[0121] After training is complete, the unmasked subgraph is input into the trained local generator, and the trained local generator generates the first missing node feature of the subgraph.

[0122] In some embodiments, processor 211 is configured to implement:

[0123] The global node classifier is a node classifier used to predict user behavior and user preferences in the financial system.

[0124] In some embodiments, processor 211 is configured to implement:

[0125] The global node classifier is a node classifier used to predict user behavior and user preferences in the medical system.

[0126] In some embodiments, processor 211 is configured to implement:

[0127] The target system includes one or a combination of financial systems and medical systems.

[0128] The embodiments of this application also provide a device, the type of which includes, but is not limited to, mobile phones, cameras, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, personal computers (PCs), netbooks, personal digital assistants (PDAs), etc., and no limitations are made in the embodiments of this application.

[0129] The device includes a client sub-image repair apparatus based on BeiDou time synchronization. For example, this client sub-image repair apparatus based on BeiDou time synchronization can be the client sub-image repair apparatus 200 described in the above embodiments. The device can execute any of the client sub-image repair methods based on BeiDou time synchronization provided in the embodiments of this application. Therefore, it can achieve the beneficial effects that any of the client sub-image repair methods based on BeiDou time synchronization provided in the embodiments of this application can achieve, as detailed in the preceding embodiments, and will not be repeated here.

[0130] This application also provides a storage medium storing a computer program, which includes program instructions. The processor executes these program instructions to implement the steps of the client sub-image repair method based on BeiDou time synchronization provided in the above embodiments. For example, when the computer program is loaded by the processor, it can perform the following steps:

[0131] Obtain a subgraph of the client, wherein the subgraph is formed by partitioning the global graph of the target system, and the target system is a federated learning-based system;

[0132] The BeiDou timing system is used to synchronize the local data of this client in time series, ensuring that the subgraph is in the same time series as other subgraphs of the global graph;

[0133] A central generator of a generative adversarial neural network is deployed on the server side participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client side participating in the federated learning, and a local generator of a graph neural network is deployed on the client side.

[0134] The local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator.

[0135] The central generator and the discriminator are trained together, and the trained central generator generates the second missing node feature of the subgraph.

[0136] Based on the first missing node features and the second missing node features of the subgraph, the missing neighbor information of the subgraph is repaired;

[0137] The local node classifier of the federated learning is trained on the subgraph after the neighbor information is repaired, and local parameters of the local node classifier are generated. The local parameters are sent to the server so that the server can train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier on the node classification task.

[0138] The beneficial effects of this application's embodiments are twofold. Firstly, by using the BeiDou time synchronization system, it ensures that data from different clients are in the same time series, thereby more accurately predicting user behavior and preferences, which is beneficial for improving the accuracy of client subgraph repair based on BeiDou time synchronization. Secondly, the local node classifier of the federated learning is trained based on the subgraph after repairing the neighbor information, generating local parameters of the local node classifier. These local parameters are then sent to the server, enabling the server to train the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier in node classification tasks. Since only local parameters are uploaded and not local client data, cross-client lost information repair can be performed without disclosing privacy, solving the problem of how to perform cross-client lost information repair without disclosing privacy, and also solving the problem of how to train a global node classifier using a subgraph after repairing client lost information.

[0139] The storage medium can be the internal storage unit of the client sub-map repair device or equipment based on BeiDou time synchronization in the aforementioned embodiments, such as the hard drive or memory of the client sub-map repair device or equipment based on BeiDou time synchronization. The storage medium can also be an external storage device of the client sub-map repair device or equipment based on BeiDou time synchronization, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the client sub-map repair device or equipment based on BeiDou time synchronization.

[0140] Since the computer program stored in the storage medium can execute any of the client sub-image repair methods based on BeiDou time synchronization provided in the embodiments of this application, it can achieve the beneficial effects that any of the client sub-image repair methods based on BeiDou time synchronization provided in the embodiments of this application can achieve. For details, please refer to the previous embodiments, which will not be repeated here.

[0141] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A client-side sub-image repair method based on BeiDou time synchronization, characterized in that, The client-side sub-image repair method based on BeiDou timing includes: Obtain the client's subgraph, which is formed by partitioning the global graph of the target system, where the target system is a financial system or a medical system based on federated learning; The BeiDou time synchronization system is used to synchronize the local data of this client to ensure that the subgraph is in the same time series as other subgraphs of the global graph. A central generator of a generative adversarial neural network is deployed on the server side participating in the federated learning, a discriminator of the generative adversarial neural network is deployed on the client side participating in the federated learning, and a local generator of a graph neural network is deployed on the client side. The local generator is trained by occluding the subgraph and restoring the occluded subgraph, and the first missing node feature of the subgraph is generated by the trained local generator. The central generator and the discriminator are trained together, and the trained central generator generates the second missing node feature of the subgraph. Based on the first missing node features and the second missing node features of the subgraph, the missing neighbor information of the subgraph is repaired; The local node classifier of the federated learning is trained on the subgraph after the neighbor information is repaired, local parameters of the local node classifier are generated, and the local parameters are sent to the server so that the server trains the global node classifier of the federated learning based on the local parameters, thereby improving the performance of the global node classifier on the node classification task. The global node classifier is either a node classifier used to predict user behavior and user preferences in the financial system, or a node classifier used to predict user behavior and user preferences in the medical system.

2. The client sub-image repair method based on BeiDou time synchronization according to claim 1, characterized in that, The process of training the local generator by occluding and restoring the occluded subgraph, and generating the first missing node feature of the subgraph using the trained local generator, includes: A occluded subgraph is input into the untrained local generator, and a graph representation is obtained through the graph neural network. The graph representation is then sent to a multilayer perceptron, which performs classification training on the number of occluded nodes, the number of predicted nodes, the occluded node features, and the predicted node features of the occluded subgraph. After training is complete, the unmasked subgraph is input into the trained local generator, and the trained local generator generates the first missing node feature of the subgraph.

3. The client sub-image repair method based on BeiDou time synchronization according to claim 1, characterized in that, The step of co-training the central generator and the discriminator, and generating the second missing node feature of the subgraph using the trained central generator, includes: Gaussian white noise is input into the central generator, which generates the initial missing node features of the subgraph. The initial missing node features are then input into the discriminator. The evaluation result is obtained through the discriminator, and the discriminator is updated at the same time. The evaluation result includes the probability of identifying the initial missing node feature as a real node feature. The central generator is updated based on the evaluation results. The loss function between the updated discriminator and the updated central generator is obtained. Multiple loss functions are aggregated. The central generator is iteratively trained using the aggregated loss function until a preset condition is met, at which point the training of the central generator is stopped. The second missing node feature of the subgraph is generated by the trained central generator.

4. The client sub-image repair method based on BeiDou time synchronization according to claim 3, characterized in that, The preset conditions include one or a combination of the following: the number of training iterations reaches a preset number, the training time reaches a preset time, and the loss function is minimized to make the adversarial neural network reach a convergence state.

5. A client-side sub-image repair device based on BeiDou time synchronization, characterized in that, The client sub-image repair device based on BeiDou time synchronization includes a processor and a memory. The memory stores a computer program that can be executed by the processor. When the computer program is executed by the processor, it implements the client sub-image repair method based on BeiDou time synchronization as described in any one of claims 1 to 4.

6. A computer device, characterized in that, The device includes the client sub-map repair device based on BeiDou time synchronization as described in claim 5.

7. A storage medium for computer-readable storage, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the steps of the client sub-graph repair method based on BeiDou time synchronization as described in any one of claims 1 to 4.

Images