A deep learning network and its method and device for predicting missing data

Abstract

The present invention relates to the field of learning network technologies, and specifically provides a deep learning network and a method and device for predicting missing data thereof, aiming to solve the problem of low accuracy of the missing data predicted by the prior art. For this purpose, the deep learning network of the present invention includes a vector extraction layer, a data encoding layer, a feature fusion layer, a feature decoding layer, and a data prediction layer. Through the above deep learning network, the missing data corresponding to the input data can be obtained, the accuracy of the obtained missing data is improved, support for the missing ratio of any input data is achieved, it is applicable to the prediction of any missing data, and the actual needs of users are met.

Description

Technical Field

[0001] The present invention relates to the technical field of learning networks, and specifically provides a deep learning network and a method and device for predicting missing data thereof. Background Art

[0002] Currently, many data in the real world are often incomplete and have missing values. Traditional non-neural network-based methods are often too weak in capabilities and cannot discover the patterns in the data, making it difficult to apply when the quantity is large. Deep learning-based methods often need to design special network structures for specific data, and their input sizes are fixed. Deep learning methods based on Transformer will remove a fixed proportion of the input to train the network, while in the task of data filling, the proportion of missing values in the input data is often not fixed. In addition, the missing value algorithm often has a unique filling value for the missing value, which cannot reflect the possible variation of the missing value. Therefore, the accuracy of the missing data predicted by the prior art is relatively low and it is difficult to meet the actual needs.

[0003] Correspondingly, there is a need in the art for a new prediction scheme for missing data to solve the above problems. Summary of the Invention

[0004] In order to overcome the above defects, the present invention is proposed to provide a solution to solve or at least partially solve the technical problem that the accuracy of the missing data predicted by the prior art is relatively low. The present invention provides a deep learning network and a method and device for predicting missing data thereof.

[0005] In a first aspect, the present invention provides a deep learning network for predicting missing data corresponding to input data. The deep learning network includes: a vector extraction layer for extracting a relationship vector between any two elements in the input data; a data encoding layer for encoding each element in the input data to obtain a feature vector corresponding to each element; a feature fusion layer for performing feature fusion according to the feature vector corresponding to each element and the relationship vector to obtain a first feature vector corresponding to each coordinate of the input data; a feature decoding layer for decoding according to the first feature vector and the relationship vector to obtain a second feature vector corresponding to each coordinate of the input data; and a data prediction layer for predicting the missing data corresponding to the input data according to the second feature vector corresponding to each coordinate of the input data to obtain a predicted value of the missing data.

[0006] In an embodiment, the vector extraction layer includes: a coordinate encoding layer for encoding the coordinates corresponding to each element in the input data to obtain a coordinate vector corresponding to the coordinates of each element; and a relationship calculation layer for obtaining a relationship vector between any two elements in the input data according to the coordinate vector.

[0007] In one embodiment, the feature fusion layer includes: a perceptron, configured to obtain a third feature vector according to the feature vector and the relationship vector corresponding to each element; a feature update layer, configured to determine a first feature vector corresponding to each coordinate of the input data according to the third feature vector.

[0008] In one embodiment, the feature update layer includes: an obtaining unit, configured to obtain the maximum value of each dimension of the third feature vector to obtain an update vector, and the calculation formula of the update vector is:

[0009] f(i)′=element_wise_max l F i,l

[0010] where f(i)′ is the update vector, and element_wise_max represents taking the maximum value of each dimension of the third feature vector F i,l ;

[0011] a determining unit, configured to determine a first feature vector corresponding to each coordinate of the input data based on the update vector.

[0012] In one embodiment, the data prediction layer includes: a first prediction layer, configured to perform reparameterization based on the second feature vector corresponding to each coordinate of the input data to obtain the mean and variance corresponding to each coordinate of the missing data when the input data is a continuous variable; a second prediction layer, configured to determine a predicted value of the missing data according to the mean and variance corresponding to each coordinate of the missing data.

[0013] In one embodiment, the number of the feature fusion layer and the feature decoding layer is at least one.

[0014] In a second aspect, the present invention provides a method for predicting missing data, including the following steps: obtaining a historical data set and data to be predicted; constructing a deep learning network; training the deep learning network based on the historical data set to obtain a trained deep learning network; inputting the data to be predicted into the trained deep learning network to obtain a predicted value of the missing data corresponding to the data to be predicted.

[0015] In a third aspect, the present invention provides a device for predicting missing data, including: an input module, configured to obtain a historical data set and data to be predicted; a construction module, configured to construct the deep learning network according to any one of claims 1 to 6; a training module, configured to train the deep learning network based on the historical data set to obtain a trained deep learning network; a prediction module, configured to input the data to be predicted into the trained deep learning network to obtain a predicted value of the missing data corresponding to the data to be predicted.

[0016] In a fourth aspect, an electronic device is provided, which includes a processor and a storage device. The storage device is adapted to store multiple program codes, and the program codes are adapted to be loaded and run by the processor to execute the method for predicting missing data described above.

[0017] In a fifth aspect, a computer-readable storage medium is provided, in which multiple program codes are stored, and the program codes are adapted to be loaded and run by a processor to execute the method for predicting missing data described above.

[0018] One or more of the above technical solutions of the present invention have at least one or more of the following beneficial effects:

[0019] Through the above deep learning network in the present invention, the missing data corresponding to the input data can be obtained, the accuracy of the obtained missing data is improved, the support for the missing ratio of any input data is realized, it is applicable to the prediction of any missing data, and the actual needs of users are met.

[0020] The feature fusion layer in the present application can fuse the information of elements corresponding to different coordinates. Among them, the element_wise_max operation allows the same network to support different numbers of input coordinates, thereby realizing the support for the missing ratio of any input data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Referring to the accompanying drawings, the disclosure of the present invention will become more understandable. It is easy for those skilled in the art to understand that these drawings are only for illustrative purposes and are not intended to limit the protection scope of the present invention. In addition, similar numbers in the figures are used to represent similar components, where:

[0022] Figure 1 is a schematic diagram of the main structure of a deep learning network according to an embodiment of the present invention;

[0023] Figure 2 is a schematic diagram of the structure of a deep learning network according to an embodiment of the present invention;

[0024] Figure 3 is a schematic diagram of the structure of a perceptron according to an embodiment of the present invention;

[0025] Figure 4 is a schematic diagram of the main step flow of a method for predicting missing data according to an embodiment of the present invention;

[0026] Figure 5 is a schematic diagram of the main structural block diagram of a device for predicting missing data according to an embodiment of the present invention.

[0027] List of Reference Signs :

[0028] 11: Input module; 12: Construction module; 13: Training module; 14: Prediction module. Specific implementation manners

[0029] Some implementation manners of the present invention will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation manners are only used to explain the technical principle of the present invention and are not intended to limit the protection scope of the present invention.

[0030] In the description of the present invention, "module" and "processor" may include hardware, software, or a combination of both. A module may include a hardware circuit, various suitable sensors, communication ports, memories, and may also include a software part, such as program code, or may be a combination of software and hardware. The processor may be a central processing unit, a microprocessor, an image processor, a digital signal processor, or any other suitable processor. The processor has data and / or signal processing functions. The processor may be implemented in software, in hardware, or in a combination of both. The non-transitory computer-readable storage medium includes any suitable medium for storing program code, such as magnetic disks, hard disks, optical disks, flash memories, read-only memories, random access memories, and the like. The term "A and / or B" represents all possible combinations of A and B, such as only A, only B, or A and B. The term "at least one A or B" or "at least one of A and B" has the same meaning as "A and / or B" and may include only A, only B, or A and B. The singular terms "a" and "this" may also include the plural form.

[0031] Currently, the accuracy of the missing data predicted by the traditional technology is relatively low and it is difficult to meet the actual needs. For this reason, the present application proposes a deep learning network and a method and device for predicting missing data. Through the deep learning network, the missing data corresponding to the input data can be obtained, the accuracy of the obtained missing data is improved, the support for the missing data ratio of any input data is realized, it is applicable to the prediction of any missing data, and the actual needs of users are met.

[0032] Refer to the attached Figure 1 , Figure 1 is a schematic diagram of the main structure of a deep learning network according to an embodiment of the present invention. As Figure 1As shown in the figure, the deep learning network in the embodiments of the present invention mainly includes a vector extraction layer 100, a data encoding layer 200, a feature fusion layer 300, a feature decoding layer 400, and a data prediction layer 500. Among them, the vector extraction layer 100 can be used to extract the relationship vectors between any two elements in the input data. The data encoding layer 200 can be used to encode each element in the input data to obtain the feature vector corresponding to each element. The feature fusion layer 300 can be used to perform feature fusion according to the feature vector and the relationship vector corresponding to each element to obtain the first feature vector corresponding to each coordinate of the input data. The feature decoding layer 400 can perform decoding according to the first feature vector and the relationship vector to obtain the second feature vector corresponding to each coordinate of the input data. The data prediction layer 500 can be used to predict the missing data corresponding to the input data according to the second feature vector corresponding to each coordinate of the input data to obtain the predicted value of the missing data. In one embodiment, the number of the above-mentioned feature fusion layer and feature decoding layer is at least one. Specifically, the following of this application will take the number of the feature fusion layer and the feature decoding layer as multiple for detailed description, which will not be elaborated here.

[0033] Through the above deep learning network, the missing data corresponding to the input data can be obtained, the accuracy of the obtained missing data is improved, the support for the missing data ratio of any input data is realized, it is applicable to the prediction of any missing data, and the actual needs of users are met.

[0034] In one embodiment, the vector extraction layer includes a coordinate encoding layer and a relationship calculation layer. Among them, the coordinate encoding layer can be used to encode the coordinates corresponding to each element in the input data to obtain the coordinate vector of the coordinates corresponding to each element, and the relationship calculation layer can be used to obtain the relationship vectors between any two elements in the input data according to the coordinate vectors.

[0035] In one embodiment, the feature fusion layer includes a perceptron and a feature update layer. Among them, the perceptron can be used to obtain the third feature vector according to the feature vector and the relationship vector corresponding to each element, and the feature update layer can be used to determine the first feature vector corresponding to each coordinate of the input data according to the third feature vector.

[0036] In one embodiment, the feature update layer includes an acquisition unit and a determination unit. Among them, the acquisition unit can be used to obtain the maximum value of each dimension of the third feature vector to obtain an update vector, and the determination unit can be used to determine the first feature vector corresponding to each coordinate of the input data based on the update vector. The calculation formula of the update vector is:

[0037] f(i)′=element_wise_max l F i，l

[0038] where f(i)' is the updated vector, and element_wise_max represents taking the maximum value for each dimension of the third feature vector F i，l The feature fusion layer can fuse the information of elements corresponding to different coordinates. Among them, the element_wise_max operation allows the same network to support different numbers of input coordinates, thereby achieving support for any input data missing ratio.

[0039] In one embodiment, the data prediction layer includes a first prediction layer and a second prediction layer. The first prediction layer can be used to reparameterize based on the second feature vector corresponding to each coordinate of the input data when the input data is a continuous variable, to obtain the mean and variance corresponding to each coordinate of the missing data; the second prediction layer can be used to determine the predicted value of the missing data according to the mean and variance corresponding to each coordinate of the missing data.

[0040] Specifically, the number of feature fusion layers and feature decoding layers in the deep learning network in this embodiment can be multiple, as Figure 2 shown. The output of the first feature fusion layer can be used as the input of the next feature fusion layer. Similarly, the output of the first feature decoding layer can be used as the input of the next feature decoding layer, and the output of the last feature fusion layer can be used as the input of the first feature decoding layer. Input the input data into the deep learning network, and output the predicted value corresponding to the missing data of the input data, which will be illustrated in detail through the following embodiments.

[0041] For data x, assume that the elements in this data are x1, x2,..., x n , where some elements (x a , a ∈ M(x)) are missing. Now it is necessary to fill in the missing elements according to the original data in data x and represents the complement set of M(x). Note that here the set M is a function of x, that is, the missing sets corresponding to different x are different.

[0042] First, the original data in data x As input data, it is input into a deep learning network. Each element of the input data has its corresponding coordinate. For example, the coordinates corresponding to an image are two-dimensional coordinates, and the coordinates corresponding to text and sound are one-dimensional coordinates. Most of the data in machine learning also corresponds to one-dimensional coordinates. For any coordinate i in the input data, the position encoding layer in the above deep learning network is used to perform position encoding on it to generate coordinate vectors corresponding to the coordinates of each element, such as e(i), e(j), etc., where i and j can be the coordinates corresponding to any two elements in the input data. Generally speaking, different data types correspond to different position encoding processing schemes. On the one hand, one-hot encoding is adopted. For example, each number from 1 to n is encoded using one-hot encoding. Specifically, for position i, a vector e(i) of length n is used to represent it, where only the i-th position is 1 and the other positions are 0. This scheme is applicable to input data where there is no relationship between coordinates. For example, the user profile in Internet data contains many features of a user, such as age, gender, purchase records, etc. Each feature corresponds to a coordinate, and there is no relationship between these coordinates. On the other hand, the cosine scheme is used to implement position encoding. Specifically, for any coordinate i in the input data, a vector e(i) of dimension 2*A is generated. The even dimensions of this vector are defined as e 2*j = sin(i / 10000 2i / A ), and the odd dimensions are defined as: e 2*j+1 = cos(i / 10000 2i / A ), where A is a preset dimension. The scheme of using the cosine scheme to implement position encoding is applicable to input data where there is a certain periodicity between different coordinates, such as image, text, sound data, etc. The position encoding in this application is not limited to the scheme described in the above embodiments, and any encoding that can reflect the differences between different coordinates can be used as a position encoding scheme.

[0043] After generating the vectors e(i) and e(j) by performing position encoding on the input data, a relationship vector R(i, j) can be calculated. This calculation can be implemented by a three-layer perceptron (MLP1), but it is not limited to this. MLP1 can also be replaced by other network structures as long as it meets the requirements of taking two coordinate vectors as input and outputting a relationship vector. The specific structure of the three-layer perceptron is as Figure 3 shown. After inputting the vectors e(i) and e(j) into the perceptron (MLP1), the output is the relationship vector R(i, j).

[0044] Encoding each element in the input data to obtain the feature vector corresponding to each element. For example, any two elements x i and x jBy encoding them separately, we can obtain f(i) and f(j). For each element in the input data, it may belong to different data types, such as discrete variables, continuous variables, images, etc. Therefore, different types of data are encoded in different ways. Specifically, when the input data is a continuous variable, each element of the input data needs to be normalized, that is, the mean of each element dimension is subtracted from each element and then divided by the variance to achieve encoding. For discrete values, one-hot encoding is used, and for sound, Mel-Frequency Cepstrum encoding is used.

[0045] Next, the feature fusion layer of the deep learning network can be used to fuse the feature vectors f(i) and f(j) corresponding to each element and the relationship vector R(i,j). Specifically, the feature vectors f(i) and f(j) corresponding to each element and the relationship vector R(i,j) are first input into the first feature fusion layer. A three-layer perceptron (MLP2) in the feature fusion layer outputs the third feature vector F(i,j). Then, the first feature vector f(i)′ is calculated based on the following formula:

[0046] f(i)′=element_wise_max l F i,l

[0047] Among them, f(i)′ is the update vector, element_wise_max represents the third eigenvector F i,l Take the maximum value for each dimension, F i,l That is, F(i,j). In this embodiment, when , f(i) and f(j) are the feature vectors f(i) and f(j) output by the data encoding layer. Then, the output of the first feature fusion layer can be used as the input of the second feature fusion layer, until all the feature fusion layers are traversed, and the updated first feature vectors f(i)′ and f(j)′ can be obtained. The MLP2 used in each application of the feature fusion layer has its own independent parameters. Since the input data is missing, the encoding part only calculates the coordinates of the part that is not missing. The MLP2 in this application can be replaced by other network structures, not just limited to MLP2. In addition, element_wise_max can also be replaced by other operations, as long as the order of input does not affect the structure of this operation, such as finding the mean of each feature dimension, etc., but not limited to this.

[0048] Secondly, the output of the last feature fusion layer is used as the input of the first feature decoding layer for decoding. After traversing all the feature fusion layers, the second feature vectors corresponding to each coordinate of the decoded input data can be obtained. Specifically, the structure of each feature decoding layer is similar to that of the feature fusion layer. For any i, j ∈ N, the perceptron MLP2 in the feature decoding layer calculates F(i,j)′ according to the inputs f(i)′, f(j)′, R(i,j), and then calculates the updated value f(i)″ of f(i)′ = element_wise_max l F i,l , where element_wise_max means taking the maximum value for each dimension of the feature. The input of the first feature decoding layer is the output of the last feature fusion layer. After traversing all the feature decoding layers, the updated second feature vectors f(i)″, f(j)″ can be obtained. In this step, if i ∈ M, then set f(i)′ = 0 and f(j)′ = 0.

[0049] Finally, the second feature vectors f(i)″, f(j)″ are input into the data prediction layer of the deep learning network, and the predicted values of the missing data can be obtained. Specifically, for continuous variables, the "reparameterization" scheme is adopted. For each coordinate, the prediction layer outputs two variables, representing the mean μ and variance δ of the predicted value of the element corresponding to the current coordinate respectively. Each time a random number r is sampled from the standard normal distribution, and finally the predicted value of the element corresponding to each coordinate in the missing data can be calculated by the formula r * δ + μ. For discrete variables, softmax can be used to classify the output results of the prediction layer, so as to obtain the probability distribution of the element corresponding to each coordinate in the missing data. Finally, the value corresponding to the maximum probability in the probability distribution can be used as the predicted value of the element corresponding to each coordinate in the missing data.

[0050] When training the deep learning network model, for an input data x = x1, x2,..., x n , each time a part of i ∈ M1 is randomly selected from as the input, and then the values of the remaining part of the coordinates i ∈ M2 are predicted through the network, where M1 ∩ M2 = φ. Specifically, for different types of input data, different types of objective functions can be selected for training. In this application, for continuous variables, the objective function is the mean squared error, and for discrete variables, the objective function is the negative log-likelihood function. During the model training process, the specific training method adopts the Stochastic Gradient Descent (SGD) algorithm with momentum, the learning rate is 0.01, and the momentum is 0.99. To accelerate the training, the input size of each batch is the same, but the input sizes of different batches can be different.

[0051] Furthermore, the present invention also provides a method for missing data, as Figure 4 shown, the prediction method for missing data in the embodiments of the present invention is mainly implemented based on the following steps S101 - step S104.

[0052] Step S101: Obtain a historical data set and data to be predicted.

[0053] Step S102: Construct a deep learning network. The deep learning network constructed in this embodiment can be the deep learning network detailed in the foregoing examples. For the specific deep learning network, please refer to the foregoing embodiments and will not be elaborated here.

[0054] Step S103: Train the deep learning network based on the historical data set to obtain a trained deep learning network. Specifically, during the training process, the Stochastic Gradient Descent (SGD) algorithm with momentum is used for network training, where the learning rate is 0.01 and the momentum is 0.99. When the input data is a continuous variable, the objective function is the mean squared error; when the input data is a discrete variable, the objective function is the negative log-likelihood function.

[0055] Step S104: Input the data to be predicted into the trained deep learning network to obtain the predicted value of the missing data corresponding to the data to be predicted. After obtaining the trained deep learning network based on the foregoing step S103, input the data that needs to be predicted into the trained deep learning network to obtain the predicted value of the missing data corresponding to the data to be predicted, thereby realizing data filling. Specifically, for the entire complete data set x = x1, x2,..., x n , input the data x i , i ∈ M, and then through network sampling, the predicted value of the missing data can be output

[0056] Based on the above steps S101 - step S104, the predicted value of the missing data corresponding to the input data can be obtained, the accuracy of the obtained missing data is improved, support for any missing data ratio of the input data is realized, it is applicable to the prediction of any missing data, and the actual needs of users are met.

[0057] It should be noted that although the above steps are described in a specific order in the above embodiments, those skilled in the art can understand that in order to achieve the effects of the present invention, different steps do not necessarily need to be executed in such an order, and they can be executed simultaneously (in parallel) or in other orders, and these variations are all within the protection scope of the present invention.

[0058] Furthermore, the present invention also provides a prediction device for missing data.

[0059] Refer to the appendix Figure 5 ,Figure 5 It is a main structural block diagram of a prediction device for missing data according to an embodiment of the present invention. As Figure 5 shown, the prediction device for missing data in the embodiment of the present invention mainly includes an input module 11, a construction module 12, a training module 13, and a prediction module 14. In some embodiments, one or more of the input module 11, the construction module 12, the training module 13, and the prediction module 14 may be combined into one module. In some embodiments, the input module 11 may be configured to obtain a historical data set and data to be predicted. The construction module 12 may be configured to construct a deep learning network. The training module 13 may be configured to train the deep learning network based on the historical data set to obtain a trained deep learning network. The prediction module 14 may be configured to input the data to be predicted into the trained deep learning network to obtain a predicted value of the missing data corresponding to the data to be predicted. In one implementation manner, the description of the specific implementation functions can refer to those described in steps S101 - S104.

[0060] The above - mentioned prediction device for missing data is used to execute Figure 4 the prediction method embodiment for missing data shown. The technical principles, the technical problems solved, and the technical effects produced by both are similar. Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working process and relevant descriptions of the prediction device for missing data can refer to the content described in the embodiment of the prediction method for missing data, which will not be elaborated here.

[0061] Those skilled in the art can understand that all or part of the processes in the method of implementing the above - mentioned embodiment of the present invention can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer - readable storage medium. When the computer program is executed by a processor, the steps of the above - mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer - readable storage medium can include: any entity or device, medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read - only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer - readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer - readable storage medium does not include electrical carrier signals and telecommunication signals.

[0062] Furthermore, the present invention also provides an electronic device. In an embodiment of the electronic device according to the present invention, the electronic device includes a processor and a storage device. The storage device can be configured to store a program for executing the method for predicting missing data in the above method embodiment. The processor can be configured to execute the program in the storage device, and the program includes, but is not limited to, the program for executing the method for predicting missing data in the above method embodiment. For the sake of convenience of description, only the parts related to the embodiments of the present invention are shown. For the specific technical details not disclosed, please refer to the method part of the embodiments of the present invention.

[0063] Furthermore, the present invention also provides a computer-readable storage medium. In an embodiment of the computer-readable storage medium according to the present invention, the computer-readable storage medium can be configured to store a program for executing the method for predicting missing data in the above method embodiment. The program can be loaded and run by a processor to implement the above method for predicting missing data. For the sake of convenience of description, only the parts related to the embodiments of the present invention are shown. For the specific technical details not disclosed, please refer to the method part of the embodiments of the present invention. The computer-readable storage medium can be a storage device formed by various electronic devices. Optionally, the computer-readable storage medium in the embodiments of the present invention is a non-transitory computer-readable storage medium.

[0064] Furthermore, it should be understood that since the setting of each module is only for illustrating the functional units of the device of the present invention, the corresponding physical devices of these modules can be the processor itself, or a part of the software in the processor, a part of the hardware, or a part of the combination of software and hardware. Therefore, the number of each module in the figure is only illustrative.

[0065] Those skilled in the art can understand that the various modules in the device can be adaptively split or combined. Such splitting or combination of the specific modules will not cause the technical solution to deviate from the principle of the present invention. Therefore, the technical solutions after splitting or combination will all fall within the protection scope of the present invention.

[0066] So far, the technical solution of the present invention has been described in conjunction with the preferred embodiments shown in the drawings. However, those skilled in the art can easily understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

Claims

1. A method for constructing a deep learning network for predicting missing data corresponding to input data, characterized in that, The deep learning network includes: A vector extraction layer, configured to extract a relationship vector between any two elements in the input data, where the input data is an image, text, or sound. The coordinates corresponding to the image are two-dimensional coordinates, and the coordinates corresponding to the text and sound are one-dimensional coordinates. The vector extraction layer includes: A coordinate encoding layer, configured to encode the coordinates corresponding to each element in the input data to obtain a coordinate vector for the coordinates corresponding to each element; A relationship calculation layer, configured to obtain a relationship vector between any two elements in the input data according to the coordinate vectors; A data encoding layer, configured to encode each element in the input data to obtain a feature vector corresponding to each element; A feature fusion layer, configured to perform feature fusion according to the feature vector and the relationship vector corresponding to each element to obtain a first feature vector corresponding to each coordinate of the input data. The feature fusion layer includes: A perceptron, configured to obtain a third feature vector according to the feature vector and the relationship vector corresponding to each element; An acquisition unit, configured to obtain the maximum value of each dimension of the third feature vector to obtain an update vector. The calculation formula of the update vector is: f(i)' = element_wise_max l F i,l Among them, f(i)′ is the update vector, and element_wise_max represents taking the maximum value for each dimension of the third feature vector F i,l ; A determination unit, configured to determine a first feature vector corresponding to each coordinate of the input data based on the update vector; A feature decoding layer, configured to perform decoding according to the first feature vector and the relationship vector to obtain a second feature vector corresponding to each coordinate of the input data; A data prediction layer, configured to predict the missing data corresponding to the input data according to the second feature vector corresponding to each coordinate of the input data to obtain a predicted value of the missing data.

2. The method for constructing a deep learning network according to claim 1, wherein The data prediction layer includes: A first prediction layer, configured to perform reparameterization based on the second feature vector corresponding to each coordinate of the input data when the input data is a continuous variable to obtain the mean and variance corresponding to each coordinate of the missing data; A second prediction layer, configured to determine the predicted value of the missing data according to the mean and variance corresponding to each coordinate of the missing data.

3. The method for constructing a deep learning network according to claim 1, characterized in that, The number of the feature fusion layer and the feature decoding layer is at least one.

4. A method for predicting missing data, characterized in that, Including the following steps: Obtain a historical data set and data to be predicted; Construct a deep learning network according to the method described in any one of claims 1 to 3; Train the deep learning network based on the historical data set to obtain a trained deep learning network; Input the data to be predicted into the trained deep learning network to obtain a predicted value of the missing data corresponding to the data to be predicted.

5. A prediction device for missing data, characterized in that, Including: An input module, configured to obtain a historical data set and data to be predicted; A construction module, configured to construct a deep learning network according to the method described in any one of claims 1 to 3; A training module, configured to train the deep learning network based on the historical data set to obtain a trained deep learning network; A prediction module, configured to input the data to be predicted into the trained deep learning network to obtain a predicted value of the missing data corresponding to the data to be predicted.

6. An electronic device, comprising a processor and a storage device, the storage device being adapted to store a plurality of program codes, characterized in that The program code is adapted to be loaded and run by the processor to execute the prediction method for missing data described in claim 4.

7. A computer-readable storage medium storing multiple program codes, characterized in that, The described program code is adapted to be loaded and run by a processor to execute the method for predicting missing data according to claim 4.

Images