Data completion method and system based on generative adversarial network and variational auto-encoder

Abstract

The invention provides a data completion method and system based on a generative adversarial network and a variational auto-encoder, and relates to the technical field of data completion, and the method comprises the steps: obtaining to-be-completed data; calculating mutual information among different attributes in the to-be-complemented data, and dividing the to-be-complemented data into different attribute groups according to the mutual information; preprocessing the to-be-complemented data to obtain preprocessed data, and generating a corresponding mark matrix; and inputting the preprocessed data and the mark matrix into a data filling model, and complementing the data. According to the invention, the accuracy and stability of the completion of the user-changing associated data can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data completion, and in particular to a data completion method and system based on a generative adversarial network and a variational autoencoder. Background Art

[0002] Due to various factors such as sensor failure and communication interruption, the substation-to-user association data between the substation and the users it supplies will inevitably be missing to varying degrees, thus affecting the integrity and accuracy of the substation-to-user association data. Effective filling of missing data is of vital importance for subsequent data analysis, data prediction, and the stable operation of the power system. In dealing with the problem of missing data, traditional statistical methods have gradually failed to meet the high-dimensional and large-scale characteristics of modern data. Therefore, missing data filling technology based on machine learning has gradually become the focus of research. At present, machine learning in the field of missing data filling is mainly divided into two directions: the first direction is to regard the missing data filling problem as a regression estimation problem, and estimate the missing values ​​through model training. Typical discriminant models include K-Nearest Neighbor (KNN) and random forests; the second direction is to regard missing data filling as a data generation task, and generate complete data similar to the missing samples through training models. Typical generative models include variational autoencoders (VAE) and generative adversarial networks (GAN).

[0003] Generative adversarial networks are widely used in the field of gap filling because they can generate high-quality data. However, their adversarial training process is difficult to achieve Nash equilibrium and is prone to training collapse, causing the generator to "cheat" the discriminator and generate unreasonable data. The loss function of variational autoencoder training is measured by variance, which results in the generated data not being able to preserve data features well. Summary of the invention

[0004] In order to address the deficiencies of the prior art, the object of the present invention is to provide a data completion method and system based on a generative adversarial network and a variational autoencoder, which combines the advantages of a variational autoencoder and a generative adversarial network to improve the accuracy and stability of data completion.

[0005] In order to achieve the above object, according to some embodiments, a first aspect of the present invention provides a data completion method based on a generative adversarial network and a variational autoencoder, comprising:

[0006] Get the data to be completed;

[0007] Calculate the mutual information between different attributes in the data to be completed, and divide the data to be completed into different attribute groups according to the mutual information;

[0008] Preprocess the data to be completed to obtain the preprocessed data and generate the corresponding label matrix;

[0009] The preprocessed data and the label matrix are input into the data filling model to complete the data;

[0010] Among them, the data completion model includes a generative network and a discriminator. The generative network takes the preprocessed data and the labeling matrix as input, outputs estimated data, and then obtains the completed data. The input layer and hidden layer in the generative network are grouped according to the attribute grouping of the data to be completed. The nodes in the same group are fully connected, and the nodes between different groups have no neural connection. The discriminator takes the completed data as input and improves the data completion quality of the generative network by adversarial training.

[0011] Preferably, the data to be completed are divided into different attribute groups according to mutual information, including, based on the Fast-Unfolding algorithm, dividing attributes with high mutual information into the same attribute group and dividing attributes with low mutual information into different attribute groups.

[0012] Preferably, the data to be completed is preprocessed, including converting the attribute values ​​in the data using one-hot encoding;

[0013] and / or, replace missing values ​​in the data with the column mean;

[0014] and / or, normalizing the data;

[0015] According to a further technical solution, if the data in the preprocessed data X needs to be completed, the corresponding position in the marking matrix is ​​0; if the data in the preprocessed data X does not need to be completed, the corresponding position in the marking matrix is ​​1.

[0016] Preferably, the completed data is expressed as follows:

[0017]

[0018] Among them, X′ represents the completed data, X represents the preprocessed data, M represents the label matrix, Represents the estimated data of the generated network output, and ⊙ represents the multiplication of the elements at the corresponding positions in the matrix.

[0019] Preferably, the number of nodes in the input layer of the generated network corresponds to the number of attributes of the data to be completed;

[0020] And / or, in the hidden layer of the generated network, the number of nodes in each group is proportional to the number of nodes in each group of the input layer.

[0021] A second aspect of the present invention provides a data completion system based on a generative adversarial network and a variational autoencoder, comprising:

[0022] A data acquisition module is configured to acquire data to be completed;

[0023] An attribute grouping module is configured to calculate mutual information between different attributes in the data to be completed, and divide the data to be completed into different attribute groups according to the mutual information;

[0024] A preprocessing module is configured to preprocess the data to be completed, obtain preprocessed data, and generate a corresponding labeling matrix;

[0025] The data completion module is configured to input the preprocessed data and the label matrix into the data completion model to complete the data; wherein, the data completion model includes a generator network and a discriminator, the generator network takes the preprocessed data and the label matrix as input, outputs estimated data, and then obtains the completed data, and the input layer and hidden layer in the generator network are grouped according to the attribute grouping of the data to be completed, the nodes in the same group are fully connected, and the nodes between different groups have no neural connection; the discriminator takes the completed data as input, and improves the data completion quality of the generator network by adversarial training.

[0026] According to a third aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to complete the steps of the above-mentioned data completion method based on a generative adversarial network and a variational autoencoder.

[0027] In a fourth aspect of the present invention, a computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned data completion method based on a generative adversarial network and a variational autoencoder.

[0028] In a fifth aspect of the present invention, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned data completion method based on a generative adversarial network and a variational autoencoder.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention provides a data completion method and system based on a generative adversarial network and a variational autoencoder. The variational autoencoder and the generative adversarial network are combined to complete the transformer-user association data between the transformer substation and the users it supplies power to. The mutual information between the data is calculated, the attribute groups are divided according to the mutual information, and the data in each attribute group is grouped and completed, so as to make full use of the correlation between the transformer-user association data and improve the accuracy of data completion. The generative network in the data completion model is improved based on the generator of the variational autoencoder and the generative adversarial network, and the original fully connected network structure is replaced by a sparse network structure. The input layer and the hidden layer are grouped according to different attribute groups, and the neural connections between the network layer nodes of different groups are deleted, thereby reducing unnecessary calculations and neural connections, thereby improving the computational efficiency of the model training and completion process, so that the model can process large-scale data sets faster, thereby optimizing the computational efficiency.

[0031] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0033] Figure 1 Schematic diagram of the method in Embodiment 1 of the present invention;

[0034] Figure 2 This is a schematic diagram of the MI-VAEGAN model;

[0035] Figure 3 This is a schematic diagram of the MI-GVAEGAN model;

[0036] Figure 4 A schematic diagram of the generated network structure. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] Embodiment 1

[0039] Embodiment 1 of the present invention provides a data completion method based on a generative adversarial network and a variational autoencoder, such as Figure 1-Figure 4 As shown, including:

[0040] S1. Obtain the data to be completed;

[0041] S2, calculating the mutual information between different attributes in the data to be completed, and dividing the data to be completed into different attribute groups according to the mutual information;

[0042] S3, preprocessing the data to be completed to obtain preprocessed data and generating a corresponding labeling matrix;

[0043] S4. Input the preprocessed data and the label matrix into the data filling model to complete the data. The data filling model includes a generator network and a discriminator. The generator network takes the preprocessed data and the label matrix as input, outputs estimated data, and then obtains the completed data. The input layer and hidden layer in the generator network are grouped according to the attribute grouping of the data to be completed. The nodes in the same group are fully connected, and the nodes between different groups have no neural connection. The discriminator takes the estimated data after filling as input, and improves the data completion quality of the generator network by adversarial training.

[0044] In order to solve the problems of insufficient accuracy and low generation quality in the existing missing data filling process, this embodiment combines the advantages of two machine learning models, variational autoencoder VAE and generative adversarial network GAN, in generating data. First, a missing data imputation model based on variational autoencoder generative adversarial network (Mi-VAEGAN) is constructed. Then, considering the correlation of variant user-related data, MI-VAEGAN is further improved, and a grouped missing data imputation model based on attribute mutual information (Missing data Imputation by Grouped Variational Autoencoder Generative Adversarial Network, MI-GVAEGAN) is proposed. The model optimizes the variational autoencoder through adversarial training, improves the element-level judgment ability of generated samples, and deeply analyzes the attribute correlation of incomplete data. The attributes are grouped accordingly, and a sparse network structure is used to replace the fully connected generative network in the original model based on the grouping situation. The missing data can be inferred from the observed data, thereby increasing the influence between strongly correlated attributes and reducing the interference of redundant information, so as to improve the accuracy of data completion while reducing the network size.

[0045] Data completion is performed based on the missing value filling model MI-VAEGAN generated by the variational autoencoder adversarial network, which mainly includes three steps: data preprocessing, missing value estimation and adversarial training. Data preprocessing is used to sort and organize the associated data of the variation household, missing value estimation is used to improve the integrity of the variation household data, and adversarial training is used to improve the quality of variation household data generation.

[0046] For the data preprocessing process, in the missing data filling task, the input sample of the model is an incomplete data sample with missing data. Before the data is input into the model for filling, the incomplete data sample is first preprocessed. The attribute values ​​of the variable user-related data are converted using one-hot encoding, and the missing data, that is, the Nan-labeled data, is replaced with the column mean. In order to avoid the influence of the data scale of different attribute values ​​on the model training, the data is normalized by min-max:

[0047]

[0048] Among them, x scale represents the normalized data, x ij represents the data before normalization, max(d j ) and min(d j ) represents the maximum and minimum values ​​of the j-th dimension attribute, and the normalized data value x scale In the interval [0,1].

[0049] In the preprocessed data, the Nan label is replaced by the column mean to obtain the preprocessed data X. At this time, it is impossible to distinguish between missing values ​​and observed values ​​in the data, so the label matrix M corresponding to the preprocessed data X is further generated. Another function of data preprocessing is to generate a label matrix M from incomplete data for model training. The label matrix M uses 0 and 1 states to mark whether the data in the preprocessed data X is a missing value. If the data in the preprocessed data X needs to be completed, the corresponding position in the label matrix is ​​0. If the data in the preprocessed data X does not need to be completed, the corresponding position in the label matrix is ​​1.

[0050] After preprocessing, missing value estimation is performed. The core work of missing value filling is the estimation of missing values, which is completed by the generative network in the model. The generative network of the MI-VAEGAN model includes an encoder and a generator, where the encoder is the encoder in the variational autoencoder, and the generator can be regarded as the decoder of the variational autoencoder or the generator of the generative adversarial network. The encoder takes the preprocessed data X and the label matrix M as input. Therefore, for data with d attribute columns, its input layer has 2d nodes, and the output layer of the generator has d nodes. The output is the estimated data of the reconstructed d attribute values. The estimated data output by the generator is expressed as The incomplete data filling result can be represented by the labeling matrix M, the preprocessed data X and the estimated data Calculate using the following formula:

[0051]

[0052] Where X′=(x i ' j ) n×d represents the completed data. The ⊙ operation is the multiplication of the elements at the corresponding positions in the matrix. The completed data consists of the observed data (preprocessed data X) and the estimated data (the estimated data output by the generator). ), retaining the observed data and filling the missing values ​​with the estimated data corresponding to the missing positions.

[0053] The model improves the data generation quality of the generative network through adversarial training. The generative network must minimize two loss functions during training: hidden loss and reconstruction loss. latent The calculation method used in conventional VAE is:

[0054]

[0055] Among them, D KL represents KL divergence, N(μ,σ) represents the latent variable distribution. Normal distribution is commonly used in VAE, and N(θ,l) represents the standard normal distribution.

[0056] During the training process, since there are no complete data samples, the reconstruction loss is calculated using the non-missing values ​​in the data. The reconstruction loss L rec Using the MSE function, it can be expressed as:

[0057]

[0058] Among them, x ij Indicates the probability of whether the jth attribute value of the i-th sample is the real data, m ij is the element in the missing marker matrix M.

[0059] The MI-VAEGAN model introduces a discriminator to improve the quality of generated samples through adversarial training. The discriminator focuses on judging the various components of the input samples. In MI-VAEGAN, the input sample of the discriminator is the completed data X'. The discriminator judges each element in X'. If it is a true value, it outputs 1, and if it is a filled value, it outputs 0. The output of the discriminator is a data matrix M' of the same size as X'. Obviously, the most ideal result of M' is the labeling matrix M. The loss function L of the discriminator can also be calculated without a real complete data sample. dis :

[0060]

[0061] Among them, m' ij is the corresponding element in the missing label matrix M' output by the discriminator.

[0062] Minimize the loss function of the discriminator so that the output of the discriminator can distinguish the filled data from the real data as much as possible. To achieve the purpose of adversarial training, it is also necessary to add a generation loss function L to the generative network. gen , whose optimization goal is to deceive the discriminator as much as possible:

[0063]

[0064] Minimize the loss function so that the discriminator can ij The position corresponding to m' is 0 ij Close to 1, that is, deceiving the discriminator to judge the data filled in by the generating network as real data.

[0065] The optimization goal of MI-VAEGAN is to minimize four loss functions. The overall optimization goal of the model is:

[0066] minLoss=α×L rec +L latent +L dis +L gen

[0067] Among them, α is a hyperparameter used to adjust the proportion of reconstruction loss in the overall loss function Loss. The reconstruction error in missing filling is an important indicator for evaluating the filling effect. Therefore, the value of α in the MI-VAEGAN model is usually greater than 1.

[0068] All network layers in the MI-VAEGAN model use fully connected networks, with 2 hidden layers in the generator network and 2 hidden layers in the discriminator network. The number of nodes in each layer is determined by the number of attributes of the incomplete data sample, and the number of nodes decreases as the number of layers increases, because the deep layer focuses on higher-level features. The weight parameters and bias parameters in the network are initialized with Xavier. All input layers and hidden layers in the network are activated with the LeakRelu function, and the output layer is activated with the Sigmoid function, so that the output data is in the same interval as the input data. The hyperparameter α has been experimentally tested and set to 100 to ensure that the generated samples are more realistic, while not restricting the training and optimization of the discriminator network.

[0069] The above is to build a missing filling model MI-VAEGAN based on the variational autoencoder generative adversarial network. Considering the correlation of the variational user-related data, this embodiment further improves MI-VAEGAN and proposes a grouped missing filling model MI-GVAEGAN based on attribute mutual information to further improve the data completion quality and computational efficiency.

[0070] Since the user-related data involves the relationship between multiple devices and users, these related data are an important basis for the detection of power grid operation status. If a device fails, resulting in a missing or erroneous attribute value of the user-related data, the MI-GVAEGAN model can use the correlation between the various attributes of the data to infer the missing data through the observed data, thus completing the missing or erroneous data and simplifying the network structure of the model.

[0071] Mutual information is a measure of the mutual dependence between two random variables and can be used to measure nonlinear relationships. Mutual information is defined by information entropy. For a random variable X = {x1, x2, ..., x n}, its information entropy H(X) is defined as:

[0072]

[0073] Among them, p(x i ) indicates that the random variable value is x i The information entropy represents the uncertainty of random variables. The smaller the entropy value, the smaller the uncertainty.

[0074] The mutual information of two variables X and Y can be expressed as:

[0075] I(X;Y)=H(X)-H(X|Y)

[0076] It can be seen that the meaning of mutual information is the reduction of uncertainty in variable X when variable Y is determined. Analyzing the data filling problem from the perspective of mutual information, each attribute in the incomplete data set is regarded as a random variable. The smaller the uncertainty of the missing variable, the higher the accuracy of the filling. This means that in all the observed data, the attribute variable with higher mutual information with the missing variable is more helpful for the calculation of missing values.

[0077] By analyzing the mutual information between data attributes, the attribute grouping method is adopted to group attributes with high mutual information into the same attribute group, and attributes with low mutual information into different attribute groups, and the model is trained in each attribute group for data filling. In order to make the correlation within each attribute group high and the correlation between groups low after division, this embodiment draws on the idea of ​​community division, and the division of attribute groups uses the Fast-Unfolding algorithm commonly used in community division.

[0078] The Fast-Unfolding algorithm is an algorithm for iteratively dividing communities, using modularity as the optimization goal. The modularity here is an indicator that reflects the tightness of community division in the network, and is defined as follows:

[0079] in

[0080] In the formula, A ij represents the edge weight from node i to node j, m represents the number of edges in the network, c i represents the number of the community where the i-node belongs, and the function δ(c i ,c j ) determines whether the communities i and j belong to are the same. If they are the same, the value is 1, otherwise the value is 0. From the definition of modularity, it can be seen that modularity means the ratio of the edge weights in the community to all edge weights in a certain community partition, minus the ratio of the edge weights within the community to all edge weights under random partition.

[0081] The Fast-Unfolding algorithm adopts a greedy approach, aims to improve modularity, and tries different community divisions.

[0082] (1) Modularity optimization. Each node in the network is treated as an independent community, and the modularity of the network is calculated. For each community, all its neighboring nodes are traversed, and the modularity gain after incorporating the node into the community is calculated. The community partition with the largest modularity gain is retained.

[0083] (2) Community cohesion. Each community in the network is cohesively aggregated into a network node. After cohesion, if any edges overlap, the weight of the edge is set to the sum of the weights of the overlapping edges.

[0084] The Fast-Unfolding algorithm iteratively repeats the above two steps until the number of communities reaches the expected target or the modularity gain is lower than a preset threshold.

[0085] After the attribute grouping is completed, the data is preprocessed similar to the process of data completion based on MI-VAEGAN, and then sent to the trained MI-GVAEGAN for data completion. The overall structure of MI-GVAEGAN is similar to MI-VAEGAN, except that the generated network adopts a network structure based on grouping sparsification. Figure 4 As shown in the figure, different filling patterns are used to distinguish the input layer and hidden layer nodes according to different attribute groups. According to the attribute grouping, the hidden layer network and the input layer network are grouped accordingly, and the number of nodes in each group in the hidden layer is proportional to the number of nodes in each group in the input layer. The network deletes the neural connections between network layers corresponding to different groups, and the networks belonging to the same group adopt a fully connected structure, thereby increasing the influence between strongly correlated attributes, reducing unnecessary calculations and neural connections, and improving the computational efficiency of model training and intelligent filling processes, so that the model can process large-scale data sets faster and optimize computational efficiency.

[0086] Compared with MI-VAEGAN, MI-GVAEGAN only further optimizes the network structure. After the network structure is simplified, the training process and the formula form of the loss function remain unchanged, so they will not be repeated here.

[0087] Embodiment 2

[0088] This embodiment provides a data completion system based on a generative adversarial network and a variational autoencoder, including:

[0089] A data acquisition module is configured to acquire data to be completed;

[0090] An attribute grouping module is configured to calculate mutual information between different attributes in the data to be completed, and divide the data to be completed into different attribute groups according to the mutual information;

[0091] A preprocessing module is configured to preprocess the data to be completed, obtain preprocessed data, and generate a corresponding labeling matrix;

[0092] The data completion module is configured to input the preprocessed data and the label matrix into the data completion model to complete the data; wherein, the data completion model includes a generator network and a discriminator, the generator network takes the preprocessed data and the label matrix as input, outputs estimated data, and then obtains the completed data, and the input layer and hidden layer in the generator network are grouped according to the attribute grouping of the data to be completed, the nodes in the same group are fully connected, and the nodes between different groups have no neural connection; the discriminator takes the estimated data after filling as input, and improves the data completion quality of the generator network by adversarial training.

[0093] It should be noted here that the various modules in this embodiment correspond one-to-one to the steps of the method in Example 1, and the specific implementation process is the same, which will not be repeated here.

[0094] Embodiment 3

[0095] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to complete the steps of the method in the first embodiment.

[0096] The processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, and the like. The processor may be implemented in at least one of the following hardware forms: DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and LA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0097] The memory may include one or more computer-readable media, which may be non-transitory. The memory may also include high-speed random access memory, and non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable medium in the memory is used to store at least one computer program, which is used to be executed by the processor to implement a data completion method based on a generative adversarial network and a variational autoencoder provided in an embodiment of the present disclosure.

[0098] Those skilled in the art can understand that the electronic device provided in this embodiment can include more or fewer components, or combine certain components, or adopt different component arrangements.

[0099] Embodiment 4

[0100] This embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the method in the first embodiment are completed.

[0101] Embodiment 5

[0102] This embodiment provides a computer program product, including a computer program / instruction, which implements the steps of the method in Embodiment 1 when executed by a processor.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A data completion method based on generative adversarial network and variational autoencoder, characterized in that: include: Get the data to be completed; Calculate the mutual information between different attributes in the data to be completed, and divide the data to be completed into different attribute groups according to the mutual information; Preprocess the data to be completed to obtain preprocessed data and generate a corresponding labeling matrix; The preprocessed data and the label matrix are input into the data filling model to complete the data; Among them, the data completion model includes a generative network and a discriminator. The generative network takes the preprocessed data and the labeling matrix as input, outputs estimated data, and then obtains the completed data. The input layer and hidden layer in the generative network are grouped according to the attribute grouping of the data to be completed. The nodes in the same group are fully connected, and the nodes between different groups have no neural connection. The discriminator takes the completed data as input and improves the data completion quality of the generative network by adversarial training.

2. The data completion method based on generative adversarial network and variational autoencoder according to claim 1, characterized in that: During the training process of the data filling model, the goal is to minimize the overall loss function, including reconstruction loss, hidden loss, discriminator loss function and generation loss function.

3. The data completion method based on generative adversarial network and variational autoencoder according to claim 1, characterized in that: The data to be completed is divided into different attribute groups according to the mutual information, including, based on the Fast-Unfolding algorithm, dividing the attributes with high mutual information into the same attribute group and dividing the attributes with low mutual information into different attribute groups.

4. The data completion method based on generative adversarial network and variational autoencoder according to claim 1, characterized in that: Preprocess the data to be completed, including converting the attribute values ​​in the data using one-hot encoding; and / or, replace missing values ​​in the data with the column mean; and / or, normalizing the data; Alternatively, if the data in the preprocessed data X needs to be completed, the corresponding position in the marking matrix is ​​0, and if the data in the preprocessed data X does not need to be completed, the corresponding position in the marking matrix is ​​1.

5. The data completion method based on generative adversarial network and variational autoencoder according to claim 1, characterized in that: The expression of the completed data is: Among them, X′ represents the completed data, X represents the preprocessed data, M represents the label matrix, Represents the estimated data of the generated network output, and ⊙ represents the multiplication of the elements at the corresponding positions in the matrix.

6. The data completion method based on generative adversarial network and variational autoencoder according to claim 1, characterized in that: The number of nodes in the input layer of the generated network corresponds to the number of attributes of the data to be completed; And / or, in the hidden layer of the generated network, the number of nodes in each group is proportional to the number of nodes in each group of the input layer.

7. A data completion system based on generative adversarial network and variational autoencoder, characterized in that: include: A data acquisition module is configured to acquire data to be completed; An attribute grouping module is configured to calculate mutual information between different attributes in the data to be completed, and divide the data to be completed into different attribute groups according to the mutual information; A preprocessing module is configured to preprocess the data to be completed, obtain preprocessed data, and generate a corresponding labeling matrix; The data completion module is configured to input the preprocessed data and the label matrix into the data completion model to complete the data; wherein, the data completion model includes a generator network and a discriminator, the generator network takes the preprocessed data and the label matrix as input, outputs estimated data, and then obtains the completed data, and the input layer and hidden layer in the generator network are grouped according to the attribute grouping of the data to be completed, the nodes in the same group are fully connected, and the nodes between different groups have no neural connection; the discriminator takes the completed data as input, and improves the data completion quality of the generator network by adversarial training.

8. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to complete the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images