Missing data processing method and device, equipment and medium

Abstract

The invention discloses a missing data processing method and device, equipment and a medium, relates to the technical field of artificial intelligence and the technical field of financial science and technology, and can better assist life insurance digital transformation based on a graph neural network model. The technical problem that an existing missing data deletion and missing data filling scheme cannot guarantee the accuracy and integrity of data is solved. The method comprises the steps of obtaining to-be-filled graph structure data according to a task application requirement; according to each node in the graph structure data, generating a vectorization representation of each node by using a graph neural network model; generating a node similarity matrix of each node according to the vectorization representation; and filling missing data in each node according to the node similarity matrix. The method and the device are suitable for missing data filling of various business data in the insurance industry and enable life insurance intelligent business scenes.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence technology and financial technology, and in particular to methods, devices, equipment and media for processing missing data. Background Art

[0002] With the rapid development of big data technology, data loss is one of the common problems in data analysis and mining. In the life insurance scenario, there is also an obvious imbalance in the saturation of customer data. For example, the data saturation of old customers is high, while the data saturation of new customers is generally low, which brings great challenges to the development of new customers and intelligent marketing.

[0003] There are three main traditional methods for dealing with missing data: 1. Direct deletion; 2. Statistical filling methods: such as mean, median, mode, etc.; 3. Interpolation methods: such as linear interpolation, polynomial interpolation, etc. Among the above three processing methods, direct deletion often causes information loss, while statistical methods and interpolation methods are difficult to handle complex data relationships and data distribution situations such as high-dimensional and nonlinear. In addition, existing methods may also introduce bias and errors when dealing with missing data, affecting subsequent data analysis and decision-making. Summary of the invention

[0004] In view of this, the present application provides a method, device, equipment and medium for processing missing data, the main purpose of which is to solve the technical problem that existing solutions for missing data deletion and missing data filling cannot guarantee the accuracy and integrity of data.

[0005] According to one aspect of the present application, a method for processing missing data is provided, the method comprising:

[0006] Obtain the graph structure data to be filled according to the task application requirements;

[0007] According to each node in the graph structure data, a vectorized representation of each node is generated using a graph neural network model;

[0008] Generate a node similarity matrix for each node according to the vectorized representation;

[0009] According to the node similarity matrix, the missing data in each node is filled.

[0010] According to another aspect of the present application, a missing data processing device is provided, the device comprising:

[0011] An acquisition module is used to obtain the graph structure data to be filled according to the task application requirements;

[0012] A first generating module is used to generate a vectorized representation of each node in the graph structure data using a graph neural network model;

[0013] A second generating module, used for generating a node similarity matrix of each node according to the vectorized representation;

[0014] A filling module is used to fill the missing data in each node according to the node similarity matrix.

[0015] According to another aspect of the present application, a computer storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned method for processing missing data is implemented.

[0016] According to another aspect of the present application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor implements the above-mentioned method for processing missing data when executing the program.

[0017] By means of the above technical scheme, the missing data processing method, device, equipment and medium provided by the present application, compared with the existing technical scheme of directly deleting data items and filling data items by statistical methods and interpolation methods, the present application obtains the graph structure data to be filled according to the task application requirements, and generates a vectorized representation of each node according to each node in the graph structure data using the graph neural network model, so as to generate a node similarity matrix for each node according to the vectorized representation, and then fill the missing data in each node according to the node similarity matrix. It can be seen that the feature learning and relationship modeling capabilities of the graph neural network model are used to obtain a low-dimensional dense vectorized representation of each node in the graph structure data, and the node similarity matrix is ​​used to obtain similar nodes with high convergence to fill the missing data, thereby providing more accurate data completion and prediction, thereby effectively improving the accuracy of subsequent data analysis and decision-making.

[0018] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0020] Figure 1 A schematic diagram of a process for processing missing data provided by an embodiment of the present application is shown;

[0021] Figure 2A schematic diagram showing a flow chart of another method for processing missing data provided in an embodiment of the present application;

[0022] Figure 3 A schematic diagram showing the structure of a missing data processing device provided in an embodiment of the present application is shown;

[0023] Figure 4 A schematic diagram of the structure of another missing data processing device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0024] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0025] Directly deleting the existing missing data often results in information loss, while statistical methods and interpolation methods are difficult to handle technical issues such as high-dimensional, nonlinear and other complex data relationships and data distribution. This embodiment provides a method for processing missing data, which can use the feature learning and relationship modeling capabilities of the graph neural network model to mine the implicit relationship between business data and process high-dimensional and nonlinear data relationships, so as to achieve high-quality missing data filling in the missing data scenario, thereby effectively improving the accuracy of subsequent data analysis and decision-making. Figure 1 As shown, the method is applied to a computer device such as a server for illustration, wherein the server can be an independent server or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (CDN: Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms, such as insurance systems, financial systems, digital business platforms, etc. The above method includes the following steps:

[0026] Step 101: Obtain graph structure data to be filled according to task application requirements.

[0027] In this embodiment, there is a corresponding relationship between task application requirements and data analysis and decision-making. In the insurance industry, taking the life insurance scenario as an example, business data includes customer business data and agent business data. When the subsequent data analysis and decision-making are to mine and analyze customer business data alone, the task application requirements can be considered to fill in missing data for customer business data alone, and the graph structure data to be filled is the customer's graph structure data. According to the needs of the actual application scenario, the graph structure data to be filled can also be the agent's graph structure data, or the customer and agent's graph structure data. The business data contained in the graph structure data to be filled is not specifically limited here.

[0028] Step 102: Based on each node in the graph structure data, a vectorized representation of each node is generated using a graph neural network model.

[0029] In this embodiment, the graph neural network model is a graph representation learning, and the GNN-GraphSAGE algorithm based on contrast loss is used for graph representation learning to achieve graph representation learning of customer relationship structure, agent relationship structure, and customer-agent relationship structure, thereby providing an effective dimensionality reduction vectorization representation for subsequent task learning. Specifically, the use of graph representation learning can map the nodes in the graph structure data to be filled obtained in step 101 to a low-dimensional vector space, thereby converting it into a low-dimensional dense vectorization representation, while ensuring that the converted vectorization representation can contain rich high-order structural features and node attribute information.

[0030] Compared with traditional data filling methods, which have the technical problem of difficulty in handling nonlinear and high-dimensional complex data relationships, this embodiment is based on graph neural networks and can effectively model and utilize the associations and dependencies between business data. Through the representation of graph structures and the learning ability of graph neural network models, it can better capture the nonlinear relationships and local dependencies of data, thereby better handling complex data relationships and effectively improving the rationality of data filling.

[0031] Step 103: Generate a node similarity matrix for each node according to the vectorized representation.

[0032] In this embodiment, according to the vectorized representation of different nodes, the node similarity matrix of each node is obtained by calculating the cosine similarity between different nodes. Specifically, for each node, each neighbor node of the node is traversed, the cosine similarity between the node and each neighbor node is calculated respectively, and the node similarity matrix of the node is generated by summarizing.

[0033] Step 104: fill in the missing data in each node according to the node similarity matrix.

[0034] In this embodiment, for each node, the node similarity matrix contains the cosine similarity between the node and each neighboring node. The closer the cosine similarity is to 1, the more similar the two nodes are. Therefore, by setting a similarity threshold, the nodes in the node similarity matrix that are greater than the similarity threshold are determined to achieve high-quality filling of the missing data in the node. It should be noted that the missing data in the filled node is marked so that after the newly added node is obtained, it can be used to verify whether the filled missing data is accurate, so as to further optimize the accuracy of the graph neural network model.

[0035] For this embodiment, the above scheme can be followed to obtain the graph structure data to be filled according to the task application requirements, and based on each node in the graph structure data, the graph neural network model is used to generate a vectorized representation of each node, so as to generate a node similarity matrix for each node based on the vectorized representation, and then fill in the missing data in each node based on the node similarity matrix. Compared with the existing technical solutions of directly deleting data items and filling in data items through statistical methods and interpolation methods, this embodiment can utilize the feature learning and relationship modeling capabilities of the graph neural network model to obtain a low-dimensional dense vectorized representation of each node in the graph structure data, and use the node similarity matrix to obtain similar nodes with high convergence to fill in the missing data, thereby providing more accurate data completion and prediction, thereby effectively improving the accuracy of subsequent data analysis and decision-making.

[0036] Furthermore, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, another method for processing missing data is provided. This embodiment can be applied to various insurance scenarios such as life insurance business, by better capturing the distribution characteristics and implicit rules of business data, avoiding bias and errors introduced by traditional methods as much as possible, thereby filling the missing data in the business data with high quality, so as to improve the accuracy and completeness of missing data filling. Figure 2 As shown, the method includes:

[0037] Step 201: pre-process the original business data to determine the business data to be filled.

[0038] In implementation, the preprocessing includes data priority analysis and data distribution analysis. Specifically, data priority analysis and data distribution analysis are performed on the original business data, and the business data to be filled is obtained through cleaning and preprocessing. According to the needs of the actual application scenario, the original business data of customers and agents are collected and sorted. The original business data of customers includes business data of dimensions such as customer basic information, insurance information, and customer behavior information; the original business data of agents includes business data of dimensions such as agent basic information and sales behavior information. The dimensions of business data are not specifically limited here.

[0039] It can be seen that by analyzing the priority and distribution of business data, data cleaning and preprocessing are completed to ensure the quality and accuracy of business data. For example, unimportant dimensional data are cleaned and preprocessed, and dimensional data with relatively stable data attribute distribution are cleaned and preprocessed. Here, dimensional data with relatively stable data attribute distribution can be user gender, etc. After streamlining the business data to be filled, the business data to be filled is obtained, thereby reducing the subsequent computing pressure to a certain extent.

[0040] Step 202: Acquire a set of business data to be filled that matches the task application requirements.

[0041] During implementation, a combination method of the business data to be filled is determined according to the task application requirements, and a business data set to be filled is determined according to the combination method. Specifically, the task application requirements may include a first task application requirement, a second task application requirement, and a first task application requirement, and accordingly, the business data combination methods to be filled are a first combination method, a second combination method, and a third combination method. For example, when subsequent data analysis and decision-making are conducted on customer business data alone, as the first task application requirement, the business data to be filled is further determined to be the first combination method, that is, the first business data set (customer business data set) is constructed based on the combination method of the customer business data.

[0042] According to the needs of actual application scenarios, the second task application requirement is agent data analysis and decision-making, and the corresponding business data to be filled is the second combination method, that is, the second business data set (agent business data set) is constructed based on the combination method of agent business data; the third task application requirement is customer and agent data analysis and decision-making, and the corresponding business data to be filled is the third combination method, that is, the third business data set (customer and agent business data set) is constructed based on the combination method of customer and agent business data. The combination method of business data is not specifically limited here.

[0043] Step 203: Generate graph structure data to be filled based on the business data set.

[0044] During the implementation, the acquired business data set is converted into a graph structure to obtain the graph structure data to be filled. Taking the customer and agent business data set as an example, each customer and agent can be used as a node Vertex in the graph structure, and the various relationships between customers-customers, agents-agents, and agents-customers form edges in the graph structure, such as customers-customers in the same life cycle, agents-customers in an insurance sales relationship, etc., so as to construct a complete graph structure data of the agent-customer relationship structure.

[0045] Step 204: According to a preset batch size, randomly downsample or randomly upsample the neighbor nodes of each node to obtain sampled neighbor nodes.

[0046] Step 205 : For each node, a vector representation of the relationship features between the node and the sampled neighbor nodes is obtained by performing calculations on the feature vectors of the node and the sampled neighbor nodes.

[0047] In practice, Graph Representation Learning (GRL) can transform graph structure data into a low-dimensional dense vector representation, while ensuring that the transformed vector representation contains rich high-order structural features and node attribute information. Specifically, a minibatch output is constructed to generate data in batches for subsequent graph representation learning. For example, the batchsize of the minibatch is set to 128, which means that 128 neighbor nodes are sampled for each node, and those greater than 128 are randomly downsampled, and those less than 128 are randomly upsampled. Preferably, the GNN-GraphSAGE algorithm based on contrastive loss is used for graph representation learning, and the Mean aggregator aggregation function is used to sample neighbor nodes, that is, When aggregating node v, the feature vectors of node v and its neighborhood are averaged to complete the graph representation learning of the agent-customer relationship structure, providing an effective dimensionality reduction vector representation for subsequent task learning.

[0048] Step 206: Generate a node similarity matrix for each node according to the vectorized representation.

[0049] In the implementation, taking the graph structure data to be filled in by customers and agents as an example, the cosine similarity between customer nodes, agent nodes, and customer and agent nodes is calculated, that is, Among them, A□B represents the dot product of node vector A and node vector B, ||A|| represents the norm of node vector A, and the closer the cosine similarity is to 1, the more similar the two nodes are. Furthermore, the calculated cosine similarities are summarized to generate the node similarity matrices of the customers and agents respectively.

[0050] Step 207: For each node, obtain a set of nodes having a similarity greater than a similarity threshold from the node similarity matrix.

[0051] Step 208: Fill in the missing data in each node according to the node set.

[0052] To illustrate the specific implementation of step 208, as a preferred embodiment, step 208 specifically includes: for each node, obtaining the label type of the missing data in the node, the label type including a numerical label and a category label; if the label type of the missing data in the node is a numerical label, filling the missing data in the node by calculating the mean of the node set; if the label type of the missing data in the node is a category label, filling the missing data in the node by obtaining the mode of the node set.

[0053] In the implementation, a node vector similarity threshold is pre-set, taking 0.8 as an example, and a node set with a cosine similarity greater than 0.8 is screened out in the node similarity matrix. If it is a numerical missing label, it is filled with the mean of the node set. If it is a categorical missing label, it is filled with the mode of the node set. Other statistical algorithms can also be selected here without specific limitation.

[0054] It can be seen that the use of global information and node structure relationships of graph neural networks can effectively improve the accuracy and stability of filling missing data, and then through the missing data filling method of graph neural networks, better application results can be achieved in multiple life insurance intelligent business scenarios compared with traditional missing data processing methods. Therefore, for subsequent task applications, based on the filled data samples, it can be applied to subsequent data analysis or data mining tasks, such as new customer development, customer marketing, agent management, risk assessment, etc.

[0055] Step 209: Obtain newly added graph structure data.

[0056] Step 210: Iterate the graph neural network model by verifying whether the filled missing data is consistent with the newly added graph structure data.

[0057] In order to illustrate the specific implementation method of step 210, as a preferred embodiment, step 210 specifically includes: if the filled missing data is consistent with the newly added graph structure data, using the graph neural network model as the target graph neural network model; if the filled missing data is inconsistent with the newly added graph structure data, continue to iterate the graph neural network model until the target graph neural network model is obtained.

[0058] In implementation, the acquired newly added graph structure data can also be used to update the attribute information in the business data to be filled.

[0059] During implementation, with the advancement of life insurance business, new customer or agent business data will be continuously generated. If the new business data contains the missing attributes of existing nodes, the new missing attributes will be filled into the existing business data by adding or replacing them. If the new business data does not contain the missing attributes of existing nodes, the new business data will be added to the iterative process of missing data filling of the graph neural network to realize iterative data filling, thereby continuously improving the accuracy of node missing value filling.

[0060] For optimizing data analysis and decision-making, filling missing data also has an important impact on subsequent data analysis and mining. The missing data filling method based on graph neural network can provide a more complete and reliable data set, and provide more accurate input for data mining, machine learning and decision-making models, which helps to optimize data analysis results, improve prediction accuracy and decision-making effects, thereby better empowering life insurance intelligent business scenarios and facilitating the digital transformation of life insurance.

[0061] By applying the technical solution of this embodiment, a method and framework for filling missing data of life insurance based on a graph neural network GNN is provided, and the feature learning and relationship modeling capabilities of the graph neural network are utilized to calculate the cosine similarity of nodes between customers or between agents or between customers and agents. Taking customer nodes as an example, the higher the similarity between nodes, the higher the similarity between the two customers' basic information, interests, hobbies, business behaviors, etc. Therefore, missing data can be filled by aggregating attributes of high-similarity node groups, thereby providing more accurate data completion and prediction. Compared with the existing technical solutions of directly deleting data items and filling data items through statistical methods and interpolation methods, this embodiment is based on graph-based data representation and adaptive feature propagation, which can better capture the distribution characteristics and implicit laws of business data, avoid introducing bias and errors as much as possible, and provide high-quality missing data filling effects.

[0062] Further, as Figure 1 The specific implementation of the method, the embodiment of the present application provides a missing data processing device, such as Figure 3 As shown, the device includes: an acquisition module 31, a first generation module 32, a second generation module 33, and a filling module 34.

[0063] The acquisition module 31 is used to acquire the graph structure data to be filled according to the task application requirements.

[0064] The first generation module 32 is used to generate a vectorized representation of each node in the graph structure data using a graph neural network model.

[0065] The second generating module 33 is used to generate a node similarity matrix of each node according to the vectorized representation.

[0066] The filling module 34 is used to fill the missing data in each node according to the node similarity matrix.

[0067] In specific application scenarios, such as Figure 4 As shown, the acquisition module 31 includes: a preprocessing unit 311 , a first acquisition unit 312 , and a generation unit 313 .

[0068] The preprocessing unit 311 is used to preprocess the original business data to determine the business data to be filled, wherein the preprocessing includes data priority analysis and data distribution analysis.

[0069] The first acquisition unit 312 is used to acquire a set of to-be-filled business data that matches the task application requirement.

[0070] The generating unit 313 is used to generate graph structure data to be filled according to the business data set.

[0071] In a specific application scenario, the first generating module 32 includes: a sampling unit 321 and a computing unit 322 .

[0072] The sampling unit 321 is used to randomly downsample or randomly upsample the neighbor nodes of each node according to a preset batch size to obtain sampled neighbor nodes.

[0073] The operation unit 322 is used to obtain a vectorized representation of the relationship features between each node and the sampled neighbor nodes by performing operation processing on the feature vectors of the node and the sampled neighbor nodes.

[0074] In a specific application scenario, the filling module 34 includes: a second obtaining unit 341 and a filling unit 342 .

[0075] The second acquisition unit 341 is used to acquire, for each node, a set of nodes having a similarity greater than a similarity threshold from the node similarity matrix.

[0076] The filling unit 342 is used to fill the missing data in each node according to the node set.

[0077] In a specific application scenario, the filling unit 332 is specifically used to obtain, for each node, the label type of the missing data in the node, where the label type includes a numerical label and a category label; if the label type of the missing data in the node is a numerical label, the missing data in the node is filled by calculating the mean of the node set; if the label type of the missing data in the node is a category label, the missing data in the node is filled by obtaining the mode of the node set.

[0078] In a specific application scenario, the device also includes an updating module 35 and a verification module 36 .

[0079] The updating module 35 is used to update the attribute information in the business data to be filled according to the acquired newly added graph structure data.

[0080] The verification module 36 is used to obtain the newly added graph structure data; and iterate the graph neural network model by verifying whether the filled missing data is consistent with the newly added graph structure data; wherein, the step of iterating the graph neural network model by verifying whether the filled missing data is consistent with the newly added graph structure data specifically includes: if the filled missing data is consistent with the newly added graph structure data, taking the graph neural network model as the target graph neural network model; if the filled missing data is inconsistent with the newly added graph structure data, continuing to iterate the graph neural network model until the target graph neural network model is obtained.

[0081] It should be noted that for other corresponding descriptions of the functional units involved in the missing data processing device provided in the embodiment of the present application, reference can be made to Figure 1 and Figure 2 The corresponding description in will not be repeated here.

[0082] Based on the above Figure 1 and Figure 2 The method shown in the embodiment of the present application accordingly provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned Figure 1 and Figure 2 The missing data are handled as shown.

[0083] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0084] Based on the above Figure 1 , Figure 2 The method shown, and Figure 3 In order to achieve the above-mentioned purpose, the embodiment of the present application also provides a computer device, which can be a personal computer, a server, a network device, etc. The physical device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1 and Figure 2 The missing data are handled as shown.

[0085] Optionally, the computer device may further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.

[0086] Those skilled in the art will appreciate that the computer device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0087] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of a computer device and supports the operation of information processing programs and other software and / or programs. The network communication module is used to achieve communication between components within the storage medium and with other hardware and software in the physical device.

[0088] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus the necessary general hardware platform, or by hardware. By applying the technical solution of the present application, compared with the existing technical solutions of directly deleting data items and filling data items through statistical methods and interpolation methods, the present embodiment can utilize the feature learning and relationship modeling capabilities of the graph neural network model to obtain a low-dimensional dense vectorized representation of each node in the graph structure data, and utilize the node similarity matrix to obtain similar nodes with high convergence to fill in the missing data, thereby providing more accurate data completion and prediction, and effectively improving the accuracy of subsequent data analysis and decision-making.

[0089] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0090] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A method for processing missing data, characterized in that: include: Obtain the graph structure data to be filled according to the task application requirements; According to each node in the graph structure data, a vectorized representation of each node is generated using a graph neural network model; Generate a node similarity matrix for each node according to the vectorized representation; According to the node similarity matrix, the missing data in each node is filled.

2. The method according to claim 1, characterized in that The step of obtaining the graph structure data to be filled according to the task application requirements includes: Preprocess the original business data to determine the business data to be filled; Acquire a set of business data to be filled that matches the task application requirements; Generate graph structure data to be filled according to the business data set; Wherein, the preprocessing includes data priority analysis and data distribution analysis.

3. The method according to claim 1, characterized in that: The step of generating a vectorized representation of each node in the graph structure data using a graph representation learning model comprises: According to the preset batch size, the neighbor nodes of each node are randomly downsampled or randomly upsampled to obtain sampled neighbor nodes; For each node, the feature vectors of the node and the sampled neighbor nodes are processed to obtain a vectorized representation of the relationship features between the node and the sampled neighbor nodes.

4. The method according to claim 1 or 3, characterized in that: The step of filling the missing data in each node according to the node similarity matrix comprises: For each node, obtaining a node set having a similarity greater than a threshold value from the node similarity matrix; Based on the set of nodes, missing data in each node is filled.

5. The method according to claim 4, characterized in that The step of filling the missing data in each node according to the node set specifically includes: For each node, obtain the label type of the missing data in the node, where the label type includes a numerical label and a category label; If the label type of the missing data in the node is a numerical label, the missing data in the node is filled by calculating the mean of the node set; If the label type of the missing data in the node is a category label, the missing data in the node is filled by obtaining the mode of the node set.

6. The method according to claim 2, characterized in that The method further comprises: According to the acquired newly added graph structure data, the attribute information in the business data to be filled is updated.

7. The method according to claim 1, characterized in that The method further comprises: Get the newly added graph structure data; Iterating the graph neural network model by verifying whether the filled missing data is consistent with the newly added graph structure data; The step of iterating the graph neural network model by verifying whether the filled missing data is consistent with the newly added graph structure data specifically includes: If the filled missing data is consistent with the newly added graph structure data, the graph neural network model is used as the target graph neural network model; If the filled missing data is inconsistent with the newly added graph structure data, continue to iterate the graph neural network model until the target graph neural network model is obtained.

8. A missing data processing device, characterized in that: include: An acquisition module is used to obtain the graph structure data to be filled according to the task application requirements; A first generating module is used to generate a vectorized representation of each node in the graph structure data using a graph neural network model; A second generating module, used for generating a node similarity matrix of each node according to the vectorized representation; A filling module is used to fill the missing data in each node according to the node similarity matrix.

9. A computer storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the missing data processing method described in any one of claims 1 to 7 is implemented.

10. A computer device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the program, the missing data processing method according to any one of claims 1 to 7 is implemented.

Images