Method, apparatus, node device and computer readable storage medium for missing hole filling

Abstract

This application discloses a missing data completion method, apparatus, node device, and computer-readable storage medium. The method includes: performing longitudinal federated learning based on sample time-series data locally on the node device and a federated specified model to obtain a federated target model; performing federated imputation based on the federated target model of the node device in the federated learning system to obtain an initial generator for the node device to perform missing data completion on the time-series data; performing generative adversarial training on the missing data completion task based on the initial generator and sample time-series data to obtain a target discriminator and a target generator for the node device; performing missing data completion on the test time-series data based on the target generator to obtain test completed data, and determining whether to use the target generator as the missing data completion model of the node device based on the first discrimination result of the target discriminator on the test completed data. This method can improve the accuracy of missing data completion.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a missing completion method, apparatus, node device, and computer-readable storage medium. Background Technology

[0002] With the successive promulgation of the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law, my country has established a relatively complete legal framework for data security and personal information protection. Increasingly, due to privacy concerns, data can no longer leave the local area, making data collection increasingly difficult and resulting in a large number of missing values. Summary of the Invention

[0003] The main technical problem addressed by this application is to provide a missing completion method, apparatus, node device, and computer-readable storage medium that can improve the accuracy of missing completion.

[0004] To address the aforementioned technical problems, the first aspect of this application provides a missing data completion method, which includes: performing longitudinal federated learning based on sample time-series data locally on a node device and a federated specified model to obtain a federated target model; wherein, the sample time-series data is marked with whether or not a masking value is displayed at the sample time position; performing federated imputation based on the federated target model of the node device in the federated learning system to obtain an initial generator for the node device to complete missing data in the time-series data; performing generative adversarial training on the missing data completion task based on the initial generator and the sample time-series data to obtain a target discriminator and a target generator for the node device; performing missing data completion on the test time-series data based on the target generator to obtain test completed data, and determining whether to use the target generator as the missing data completion model of the node device based on the first discrimination result of the target discriminator on the test completed data.

[0005] To address the aforementioned technical problems, a second aspect of this application provides a missing completion device, comprising a federated learning module, a federated filling module, a generative adversarial module, and a completion testing module. The federated learning module performs longitudinal federated learning based on local sample time-series data of a node device and a federated specified model to obtain a federated target model. The sample time-series data contains a label indicating whether a value is masked at the sample time position. The federated filling module performs federated filling based on the federated target model of the node device in the federated learning system to obtain an initial generator for the node device to perform missing completion on the time-series data. The generative adversarial module performs generative adversarial training on the missing completion task based on the initial generator and the sample time-series data to obtain a target discriminator and a target generator for the node device. The completion testing module performs missing completion on the test time-series data based on the target generator to obtain test completed data, and determines whether to use the target generator as the missing completion model of the node device based on the first discrimination result of the target discriminator on the test completed data.

[0006] To address the aforementioned technical problems, a third aspect of this application provides a node device, which includes a communication circuit, a memory, and a processor. The memory stores program instructions, and the processor executes the program instructions to implement the aforementioned missing completion method.

[0007] To address the aforementioned technical problems, a fourth aspect of this application provides a computer-readable storage medium storing program instructions executable by a processor, the program instructions being used to implement the aforementioned missing completion method.

[0008] The above technical solution utilizes an initial generator built upon a federated, multi-party, multi-track federated training model. Specifically, it uses local sample time-series data to generate an adversarial initial generator. Therefore, the initial generator can complete missing data for multiple different time-track data points, enabling the subsequent missing data completion model derived from the initial generator to also complete missing data for multiple different time-track data points, thus solving the problem of missing data completion for multi-track data.

[0009] In addition, the target discriminator of the node device is generated and constructed from the local sample time series data of the node device, which can improve the accuracy of missing completion of the subsequent missing completion model, thereby providing accurate data support for subsequent prediction and other functions. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an embodiment of the missing completion method provided in this application;

[0011] Figure 2 This is a schematic diagram of the interaction between the various node devices provided in this application;

[0012] Figure 3 yes Figure 1 The flowchart of step S13 shown is a schematic diagram of one embodiment.

[0013] Figure 4 This is a schematic diagram of an embodiment of the missing completion device provided in this application;

[0014] Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application;

[0015] Figure 6 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0016] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0017] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0018] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this document means two or more. Moreover, the term "at least one" in this document means any combination of at least two of any one or more of a plurality of objects. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0019] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the missing completion method provided in this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily reflect that outcome. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, this embodiment includes:

[0020] Step S11: Perform longitudinal federated learning based on the local sample time series data of the node device and the federated specified model to obtain the federated target model.

[0021] The method in this embodiment is used for missing value completion in time series data. Specifically, it fills in missing values ​​in time series data with high accuracy. Furthermore, this method can fill in missing values ​​for multiple different time trajectories. The fields of time series data mentioned herein include, but are not limited to, the medical field, the economic field, and the earth science field. Additionally, the time series data mentioned herein includes, but is not limited to, irregularly sampled time series (ISTS) data and ISMTS data, etc., without specific limitations.

[0022] In this embodiment, longitudinal federated learning is performed based on the local sample time-series data of the node device and the federated specified model to obtain the federated target model; wherein, the sample time-series data is marked with whether or not a value is masked. In other words, the node device uses the local sample time-series data to train the federated specified model to obtain the federated target model.

[0023] It should be noted that federated learning aims to ensure information security during big data exchange, protect terminal data and personal data privacy, and guarantee the legality and compliance of the process, while enabling efficient, secure, and reliable construction of machine learning models among multiple participants or computer nodes. The node devices mentioned in this paper refer to the aforementioned participants or computer nodes, and each participant or computer node must execute the missing data completion method provided in this application as a node device. Furthermore, vertical federated learning is generally suitable for federated learning scenarios involving participants whose datasets share the same sample space but have different feature spaces.

[0024] In one implementation, such as Figure 2 As shown, Figure 2 This is a schematic diagram of the interaction between the node devices provided in this application. Figure 2 Participant A) specifies the federated model structure, model hyperparameters, and loss function to generate the federated specified model; then, the node devices ( Figure 2 Participant A in the process sends the message to other node devices ( Figure 2 In this model, participant B is one of the node devices. This allows all node devices to jointly construct a federated target model. Specifically, after receiving the federated specified model, other node devices perform longitudinal federated learning based on their local sample time-series data and the federated specified model. This results in the subsequent fusion of the models trained by each node device to obtain the federated target model. Specifically, during the longitudinal federated learning process based on the node device's local sample time-series data and the federated specified model to obtain the federated target model, a loss function is calculated to optimize the federated specified model and ultimately arrive at the federated target model.

[0025] In one specific implementation, the federated designation model is an M-RNN model structure, and the federated designation model is optimized using the loss function shown below. The specific loss function is as follows:

[0026]

[0027] Where, x t Indicates specific characteristics; m represents x t Whether it is a missing value; t represents the time series; δ represents the time interval between two features; L represents the loss function; Φ * ,Ψ * The optimal solution for Φ and Ψ is represented by the top-level Φ and Ψ created jointly by all node devices. It consists of Φ1, Φ2, Ψ1, and Ψ2 from each party. During the calculation, L1, L2, or L is transmitted through methods such as homomorphic encryption.

[0028] Since vertical federated learning requires training with overlapping samples, the aforementioned sample time-series data is identical for each node device. However, the local time-series data of each node device may differ. Therefore, in one implementation, before each node device performs vertical federated learning based on its local sample time-series data and the specified federated model, the node device transmits data IDs to other node devices to perform data intersection and sample alignment, thereby determining the sample time-series data from its large local time-series data.

[0029] Step S12: Perform federated imputation based on the federated target model of the node device in the federated learning system to obtain the initial generator of the node device for missing data completion of time series data.

[0030] In this embodiment, federated imputation is performed based on the federated target model of the node device in the federated learning system to obtain an initial generator for the node device to perform missing data completion on time series data. That is, the node device constructs an initial generator for missing data completion on time series data using the trained federated target model. Since the initial generator is constructed based on a federated specified model through multi-party, multi-track federated training—that is, multi-party, multi-track federated training based on a federated specified model, using local sample time series data to generate an adversarial initial generator—the initial generator can complete missing data for multiple different time trajectories. This allows the subsequent missing data completion model obtained based on the initial generator to complete missing data for multiple different time trajectories, thus solving the problem of missing data completion for multi-track data.

[0031] In one implementation, after the node device obtains the federated target model by performing longitudinal federated learning based on local sample time series data and the federated specified model, the obtained federated target model is directly used as the initial generator to complete the missing data of the time series data.

[0032] Step S13: Based on the initial generator and sample time series data, perform generative adversarial training on the missing completion task to obtain the target discriminator and target generator of the node device.

[0033] In this embodiment, generative adversarial training is performed on the missing completion task based on the initial generator and sample time-series data to obtain the target discriminator and target generator for the node devices. The target discriminator of the node devices is generated and constructed from the sample time-series data locally on the node devices, which can improve the accuracy of missing completion in the subsequently obtained missing completion model, thereby providing accurate data support for subsequent prediction and other functions. Furthermore, it should be noted that each node device has its own corresponding target discriminator and target generator.

[0034] In one implementation, such as Figure 3 As shown, Figure 3 yes Figure 1 The flowchart shown in step S13 is a schematic diagram of an embodiment. Based on the initial generator and sample time-series data, generative adversarial training is performed on the missing data completion task to obtain the target discriminator and target generator of the node device. Specifically, it includes the following sub-steps:

[0035] Step S31: Obtain the initial discriminator, and based on the sample time position of the masked value in the sample time series data, generate sample prompt data, and based on the annotation information of the sample time series data, obtain the sample time series data after the value is masked.

[0036] In this embodiment, an initial discriminator is obtained, and numerically masked sample time series data is obtained based on the annotation information of the sample time series data. The numerically masked sample time series data is obtained based on the annotation information of the sample time series data, that is, the characters at one or more sample time positions in the sample time series data are masked to obtain numerically masked sample time series data.

[0037] The specific sample time and location of the masked values ​​are not limited; they can be set according to actual usage needs.

[0038] Furthermore, in this embodiment, sample hint data is generated based on the sample time positions where masked values ​​are marked in the sample time series data. The sample hint data includes the probability value of the masked value at the sample time position, which can be used to subsequently determine the authenticity of the sample completion data obtained by the initial generator for missing data completion. Since the sample hint data is generated based on the sample time series data, it is equivalent to the sample hint data being locally generated and constructed, which can greatly improve the accuracy of missing data completion in the subsequent missing data completion model, thereby providing accurate data support for subsequent prediction and other functions.

[0039] In one embodiment, sample prompt data is generated based on the sample time positions marked with masked values ​​in the sample time series data, specifically including the following steps: Step 1: Select at least one sample time position marked with masked values ​​in the sample time series data as the first sample position, and select at least one sample time position without masked values ​​in the sample time series data as the second sample position; then, fill the sample time positions marked with masked values ​​outside the first sample position with the first value, fill the sample positions marked with unmasked values ​​outside the second sample position with the second value, and fill the first sample position and the second sample position with the third value; wherein, the third value is located between the first value and the second value.

[0040] Step S32: Based on the sample noise data, use the initial generator to complete the missing data of the sample time series data after numerical masking to obtain the sample completed data.

[0041] In this embodiment, based on the sample noise data, an initial generator is used to impute missing data in the numerically masked sample time series data, resulting in completed sample data. Since the initial generator is used to impute missing data in the time series data, it is possible to use the initial generator to impute missing data in the numerically masked sample time series data, thus obtaining completed sample data.

[0042] Step S33: Based on the sample prompt data, use the initial discriminator to discriminate the sample completion data to obtain the second discrimination result.

[0043] In this embodiment, based on the sample hint data, an initial discriminator is used to discriminate the sample completion data to obtain a second discrimination result. The second discrimination result includes at least a second probability value that the completed value at the sample time position marked with the masking value is determined to be false by the initial discriminator. Since the sample hint data contains the probability value of the masked value at the sample time position, the initial discriminator can determine the authenticity of the sample completion data obtained by missing data completion using the initial generator based on the sample hint data. Furthermore, the use of the sample hint data enhances the discrimination capability of the initial discriminator.

[0044] Step S34: Based on the differences between the second discrimination result and the annotation information of the sample time series data, as well as the differences between the sample completion data and the sample time series data, adjust the initial generator and the initial discriminator to obtain the target generator and the target discriminator.

[0045] In this embodiment, the initial generator and initial discriminator are adjusted based on the differences between the second discrimination result and the labeled information of the sample time-series data, as well as the differences between the sample completion data and the sample time-series data, to obtain the target generator and target discriminator. Adjusting the initial generator and initial discriminator based on these differences minimizes the discriminant information between the second discrimination result and the labeled information of the sample time-series data, and minimizes the discriminant information between the sample completion data and the sample time-series data. This ensures that the sample completion data obtained using the initial generator approximates the sample time-series data, and that the second discrimination result obtained by the initial discriminator approximates the labeled information of the sample time-series data. This drives the initial generator's missing completion and the initial discriminator's discrimination to be as accurate as possible, thereby improving the missing completion accuracy of the subsequently obtained missing completion model.

[0046] In one embodiment, the initial generator G and the initial discriminator D are adjusted based on the differences between the second discrimination result and the labeled information of the sample time series data, as well as the differences between the sample completion data and the sample time series data, using the following formulas. The specific formulas are as follows:

[0047] min G max D (E Z,M,H (MlogD(G(z,x,m),H)+(1-M)log(1-D(G(z,x,m),H))))

[0048] Step S14: Based on the target generator, perform missing completion on the test time series data to obtain test completed data, and based on the first discrimination result of the target discriminator on the test completed data, determine whether to use the target generator as the missing completion model of the node device.

[0049] In this embodiment, missing data is imputed using the target generator to obtain test-completed data. Based on the first discrimination result of the target discriminator on the test-completed data, it is determined whether to use the target generator as the missing data imputation model for the node device. In other words, after obtaining the target generator through generative adversarial training, the missing data imputation capability of the target generator is tested. Based on the test results, it is determined whether to use the currently obtained target generator as the missing data imputation model for the node device for subsequent data imputation.

[0050] In one embodiment, the test time sequence data is marked with whether a masking value is displayed at the test time position. The first discrimination result includes at least a first probability value that the test completion data at the test time position marked with the masking value is determined to be false by the discriminator. At this time, in response to the sum of the first probability values ​​being greater than a preset threshold, it is determined that the target generator cannot be used as a missing completion model; in response to the sum of the first probability values ​​not being greater than the preset threshold, the target generator is determined to be used as a missing completion model. When the sum of the first probability values ​​is greater than the preset threshold, it indicates that the accuracy of the test completion data obtained by the target generator from missing completion of the test time sequence data is abnormal, and the target generator is not used as a missing completion model. When the sum of the first probability values ​​is not greater than the preset threshold, it indicates that the accuracy of the test completion data obtained by the target generator from missing completion of the test time sequence data is high, and the target generator can be used for missing completion of subsequent data, that is, the target generator is used as a missing completion model.

[0051] In one implementation, if it is determined that the target generator cannot be used as the missing completion model, after determining whether to use the target generator as the missing completion model for the node device based on the first discrimination result of the target discriminator on the test completion data, the federated target model is used as the new federated designated model. Based on the new federated designated model, the step of vertical federated learning based on the local sample time-series data of the node device and the federated designated model is re-executed to obtain the federated target model, and subsequent steps, until it is determined that the most recently trained target generator is used as the missing completion model. That is, if it is determined that the missing completion effect of the currently obtained target generator is not significant based on the first discrimination result of the target discriminator on the test completion data, the federated target model obtained in the previous training is retrained until it is determined that the most recently trained target generator is used as the missing completion model.

[0052] To improve the target discriminator's ability to distinguish between true and false test completion data, after the target generator performs missing completion on the test time series data to obtain the test completion data, and before determining whether to use the target generator as the missing completion model for the node device based on the first discrimination result of the target discriminator on the test completion data, test prompt data is generated based on the test time position marked with masking values ​​in the test time series data; among them, the test prompt data indicates the probability value of the masking value at the test time position.

[0053] At this point, the steps for the target discriminator to obtain the first discrimination result of the test completion data include: based on the test prompt data, using the target discriminator to discriminate the test completion data to obtain the first discrimination result. Since the test prompt data indicates the probability value of the masked value at the test time position, based on the test prompt data, the target discriminator can determine the authenticity of the test completion data obtained by missing data completion using the target generator. Furthermore, the use of the test prompt data enhances the discrimination capability of the target discriminator.

[0054] In one specific implementation, the step of generating test prompt data based on the test time positions marked with masking values ​​in the test timing data includes: selecting at least one test time position marked with masking values ​​in the test timing data as a first test position, and selecting at least one test time position in the test timing data without masking values ​​as a second test position; filling the test time positions marked with masking values ​​outside the first test position with a first value, and filling the test positions marked with unmasked values ​​outside the second test position with a second value; and filling the first test position and the second test position with a third value; wherein the third value is located between the first value and the second value.

[0055] There are no restrictions on the values ​​of the first, second, and third values; they can be set according to actual usage needs. For example, the first value can be 0, the second value can be 1, and the third value can be 0.5.

[0056] In the above implementation, since the initial generator is constructed based on a federated model with multi-party, multi-track federated training—that is, based on a federated model with multi-party, multi-track federated training—and uses local sample time-series data to generate an adversarial initial generator, the initial generator can perform missing completion for multiple different time-track data. This enables the subsequent missing completion model based on the initial generator to perform missing completion for multiple different time-track data, thus solving the problem of missing completion for multi-track data.

[0057] In addition, the target discriminator of the node device is generated and constructed from the local sample time series data of the node device, which can improve the accuracy of missing completion of the subsequent missing completion model, thereby providing accurate data support for subsequent prediction and other functions.

[0058] Please see Figure 4 , Figure 4This is a schematic diagram of an embodiment of the missing data completion device provided in this application. The missing data completion device 40 includes a federated learning module 41, a federated filling module 42, a generative adversarial module 43, and a completion testing module 44. The federated learning module 41 is used to perform longitudinal federated learning based on the sample time series data and the federated specified model on the node device to obtain a federated target model; wherein, the sample time series data is marked with whether the sample time position is masked; the federated filling module 42 is used to perform federated filling based on the federated target model of the node device in the federated learning system to obtain an initial generator for the node device to perform missing data completion on the time series data; the generative adversarial module 43 is used to perform generative adversarial training on the missing data completion task based on the initial generator and the sample time series data to obtain a target discriminator and a target generator for the node device; the completion testing module 44 is used to perform missing data completion on the test time series data based on the target generator to obtain test completed data, and based on the first discrimination result of the target discriminator on the test completed data, to determine whether to use the target generator as the missing data completion model of the node device.

[0059] The federated learning module 41 is used to determine whether to use the target generator as the missing completion model for the node device after determining whether the target generator can be used as the missing completion model based on the first discrimination result of the target discriminator on the test completion data. Specifically, it includes: using the federated target model as the new federated designated model; re-executing the step of longitudinal federated learning based on the sample time series data and the federated designated model on the node device based on the new federated designated model to obtain the federated target model, and subsequent steps, until it is determined that the latest trained target generator is used as the missing completion model.

[0060] In the aforementioned test time series data, the test time positions are marked with whether or not a value is masked. The first discrimination result includes at least a first probability value that the completed value of the test completion data at the test time position marked with a masked value is determined to be false by the discriminator. The completion test module 44 is used to determine whether to use the target generator as the missing completion model of the node device based on the first discrimination result of the target discriminator on the test completion data, including at least one of the following: in response to the sum of the first probability values ​​being greater than a preset threshold, it is determined that the target generator cannot be used as the missing completion model; in response to the sum of the first probability values ​​not being greater than the preset threshold, it is determined that the target generator is used as the missing completion model.

[0061] The missing data completion device 40 further includes a prompt test module 45. The prompt test module 45 is used after the test time series data is missing data completed based on the target generator, and before determining whether to use the target generator as the missing data completion model for the node device based on the first discrimination result of the test completion data from the target discriminator. Specifically, it includes: generating test prompt data based on the test time position where the masked value is marked in the test time series data; wherein the test prompt data indicates the probability value of the masked value at the test time position; the step of obtaining the first discrimination result of the test completion data from the target discriminator includes: using the target discriminator to discriminate the test completion data based on the test prompt data to obtain the first discrimination result.

[0062] The prompt test module 45 is used to generate test prompt data based on the test time positions marked with masking values ​​in the test timing data. Specifically, it includes: selecting at least one test time position marked with masking values ​​in the test timing data as the first test position, and selecting at least one test time position without masking values ​​in the test timing data as the second test position; filling the test time positions marked with masking values ​​outside the first test position with the first value, and filling the test positions marked with unmasked values ​​outside the second test position with the second value; and filling the first test position and the second test position with the third value; wherein the third value is located between the first value and the second value.

[0063] The generative adversarial module 43 is used to perform generative adversarial training on the missing completion task based on the initial generator and sample time series data to obtain the target discriminator and target generator of the node device. Specifically, it includes: obtaining the initial discriminator, generating sample prompt data based on the sample time position of the masked value in the sample time series data, and obtaining sample time series data after numerical masking based on the annotation information of the sample time series data; wherein the sample prompt data prompts the probability value of the masked value at the sample time position; using the initial generator to perform missing completion on the sample time series data after numerical masking based on the sample noise data to obtain sample completion data; using the initial discriminator to judge the sample completion data based on the sample prompt data to obtain a second judgment result; wherein the second judgment result includes at least a second probability value that the completed value of the sample completion data at the sample time position of the masked value is judged as false by the initial discriminator; adjusting the initial generator and initial discriminator based on the difference between the second judgment result and the annotation information of the sample time series data, and the difference between the sample completion data and the sample time series data, to obtain the target generator and target discriminator.

[0064] The adversarial generation module 43 is used to generate sample prompt data based on the sample time positions marked with masked values ​​in the sample time series data. Specifically, it includes: selecting at least one sample time position marked with masked values ​​in the sample time series data as the first sample position, and selecting at least one sample time position without masked values ​​in the sample time series data as the second sample position; filling the sample time positions marked with masked values ​​outside the first sample position with the first value, and filling the sample positions marked with unmasked values ​​outside the second sample position with the second value; and filling the first sample position and the second sample position with the third value; wherein the third value is located between the first value and the second value.

[0065] Please see Figure 5 , Figure 5 This is a schematic diagram of an embodiment of the electronic device provided in this application. The electronic device 50 includes a memory 51 and a processor 52 coupled to each other. The processor 52 is used to execute program instructions stored in the memory 51 to implement the steps of any of the above-described missing completion method embodiments. In a specific implementation scenario, the electronic device 50 may include, but is not limited to, a microcomputer or a server. In addition, the electronic device 50 may also include mobile devices such as laptops and tablets, which are not limited here.

[0066] Specifically, processor 52 controls itself and memory 51 to implement the steps of any of the above-described missing completion method embodiments. Processor 52 may also be referred to as a CPU (Central Processing Unit). Processor 52 may be an integrated circuit chip with signal processing capabilities. Processor 52 may also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor. Furthermore, processor 52 may be implemented using integrated circuit chips.

[0067] Please see Figure 6 , Figure 6This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 60 of this application embodiment stores program instructions 61. When executed, these program instructions 61 implement the methods provided by any embodiment of the missing information completion method and any non-conflicting combination thereof. The program instructions 61 can form a program file and be stored in the aforementioned computer-readable storage medium 60 in the form of a software product, so that a computer device (which may be a personal computer, server, or network device, etc.) can execute all or part of the steps of the methods of various embodiments of this application. The aforementioned computer-readable storage medium 60 includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0068] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A missing completion method, characterized in that, include: Node devices send sample time series data IDs to other node devices; Based on the local sample time-series data of the node devices and the federated specified model, longitudinal federated learning is performed to obtain the federated target model; wherein, the sample time-series data is marked with whether the sample time position is masked, and the federated target model is obtained by fusing the model trained by each of the node devices using the local sample time-series data and the federated specified model. Based on the federated target model of the node device in the federated learning system, federated imputation is performed to obtain the initial generator of the node device for missing data completion of time series data. Based on the initial generator and the sample time series data, generative adversarial training is performed on the missing completion task to obtain the target discriminator and target generator of the node device; The target generator performs missing completion on the test time series data to obtain test completed data. Based on the first discrimination result of the target discriminator on the test completed data, it is determined whether to use the target generator as the missing completion model of the node device.

2. The method according to claim 1, characterized in that, If it is determined that the target generator cannot be used as the missing completion model, after determining whether to use the target generator as the missing completion model for the node device based on the first discrimination result of the target discriminator on the test completion data, the method further includes: The aforementioned federal target model will be used as the new federal designation model. Based on the new federated designation model, the step of performing longitudinal federated learning based on the sample time-series data local to the node device and the federated designation model to obtain the federated target model and subsequent steps are re-executed until the latest trained target generator is determined as the missing completion model.

3. The method according to claim 1, characterized in that, The test time sequence data contains test time positions marked with whether or not a value is masked. The first discrimination result includes at least a first probability value that the imputed value of the test completion data at the test time position marked with a masked value is determined to be false by the discriminator. The determination of whether to use the target generator as the missing completion model of the node device based on the first discrimination result of the target discriminator on the test completion data includes at least one of the following: In response to the sum of the first probability values ​​being greater than a preset threshold, it is determined that the target generator cannot be used as the missing completion model; In response to the fact that the sum of the first probability values ​​is not greater than the preset threshold, the target generator is determined to be the missing completion model.

4. The method according to claim 1 or 3, characterized in that, After performing missing completion on the test time series data based on the target generator to obtain test completed data, and before determining whether to use the target generator as the missing completion model for the node device based on the first discrimination result of the test completed data by the target discriminator, the method further includes: Based on the test time position marked with the masking value in the test time sequence data, test prompt data is generated; wherein, the test prompt data indicates the probability value of the masking value at the test time position; The step of obtaining the first discrimination result of the target discriminator on the test completion data includes: Based on the test prompt data, the target discriminator is used to discriminate the test completion data to obtain the first discrimination result.

5. The method according to claim 4, characterized in that, The step of generating test prompt data based on the test time position marked with the masking value in the test timing data includes: Select at least one test time position in the test timing data that is marked with a masking value as the first test position, and select at least one test time position in the test timing data that is not marked with a masking value as the second test position. Fill the test time positions marked with shielded values ​​outside the first test position with the first value, fill the test positions marked with unshielded values ​​outside the second test position with the second value, and fill the test positions marked with unshielded values ​​with the third value at the first test position and the second test position; The third value is located between the first value and the second value.

6. The method according to claim 1, characterized in that, The step of performing generative adversarial training on a missing completion task based on the initial generator and the sample time-series data to obtain the target discriminator and target generator of the node device includes: An initial discriminator is obtained, and sample prompt data is generated based on the sample time position of the masked value in the sample time series data. Based on the annotation information of the sample time series data, sample time series data after numerical masking is obtained; wherein, the sample prompt data indicates the probability value of the masked value at the sample time position. Based on the sample noise data, the initial generator is used to fill in the missing data of the sample time series data after numerical masking to obtain the sample completed data. Based on the sample prompt data, the initial discriminator is used to discriminate the sample completion data to obtain a second discrimination result; wherein, the second discrimination result includes at least a second probability value that the completion value of the sample completion data at the sample time position marked with the masking value is determined to be false by the initial discriminator; Based on the difference between the second discrimination result and the annotation information of the sample time series data, as well as the difference between the sample completion data and the sample time series data, the initial generator and the initial discriminator are adjusted to obtain the target generator and the target discriminator.

7. The method according to claim 6, characterized in that, The step of generating sample prompt data based on the sample time position marked with masking values ​​in the sample time series data includes: At least one sample time position in the sample time series data marked with a masking value is selected as the first sample position, and at least one sample time position in the sample time series data without a masking value is selected as the second sample position. Fill the sample time positions marked with masked values ​​outside the first sample position with the first value, fill the sample positions marked with unmasked values ​​outside the second sample position with the second value, and fill the first sample position and the second sample position with the third value. The third value is located between the first value and the second value.

8. A missing information completion device, characterized in that, include: The federated learning module is used by node devices to send sample time series data IDs to other node devices; Based on the local sample time-series data of the node devices and the federated specified model, longitudinal federated learning is performed to obtain the federated target model; wherein, the sample time-series data is marked with whether the sample time position is masked, and the federated target model is obtained by fusing the model trained by each of the node devices using the local sample time-series data and the federated specified model. The federated filling module is used to perform federated filling based on the federated target model of the node device in the federated learning system, so as to obtain the initial generator of the node device for missing data completion of time series data. A generative adversarial module is used to perform generative adversarial training on a missing completion task based on the initial generator and the sample time series data, so as to obtain the target discriminator and target generator of the node device. The completion test module is used to complete the missing data of the test time series based on the target generator to obtain the test completed data, and to determine whether to use the target generator as the missing completion model of the node device based on the first discrimination result of the test completed data by the target discriminator.

9. A node device, characterized in that, The method includes a communication circuit, a memory, and a processor, wherein the memory stores program instructions, and the processor executes the program instructions to implement the missing completion method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The system stores program instructions that can be executed by a processor, the program instructions being used to implement the missing completion method according to any one of claims 1 to 7.

Images