Data classification method, device and computer program product based on graph neural network

Abstract

In order to solve the problem of low data classification accuracy in the prior art, the present invention discloses a data classification method, device and computer program product based on graph neural network. The present invention introduces a data association graph based on graph structure to characterize data association relationship, and before each classification, uses each header in the data table of this classification to update the association graph; then, the header feature vector can be extracted based on the updated association graph; in this way, the feature vector determined based on the graph contains the association relationship between each header in the data table of this time and between each header of the historical classification; finally, the header feature vector is input into the data classification model, and the classification result of each header can be obtained; therefore, when classifying, the present invention not only relies on the characteristics of the data itself, but also combines the data association graph to retain the association relationship between the data, based on which, the accuracy of data classification and grading can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of information security technology, and specifically relates to a data classification method, device and computer program product based on graph neural network. Background Art

[0002] With the development of big data and artificial intelligence technology, data security and privacy protection have become increasingly important. In the process of data sharing and data mining, the leakage of sensitive information may lead to serious privacy issues and security risks. Therefore, data privacy protection has received more and more attention. At present, data desensitization is often used to achieve data privacy protection. Among them, data sensitivity classification and grading is an important step before data desensitization. Especially in the case that existing data audits and data security compliance checks have become standard in various industries, data classification and grading often bring huge manual review workload.

[0003] In actual application, although there are some data classification and grading schemes based on pattern recognition, text similarity, NER algorithm (Named Entity Recognition), convolutional neural network and other technologies; however, most of the aforementioned existing technologies break the relationship between data, the calculation process is relatively independent, and the correlation between data is discarded; this will reduce the accuracy of data classification and grading, thereby affecting the reliability of desensitization; therefore, based on the aforementioned deficiencies, how to provide a data grading method with high accuracy has become a problem to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a data classification method, device and computer program product based on graph neural network, so as to solve the problem that the prior art discards the correlation between data when performing data classification, resulting in low data classification accuracy, thereby affecting the reliability of desensitization.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, a data classification method based on a graph neural network is provided, comprising:

[0007] Acquire a target data table, wherein the target data table includes a plurality of data items, each of which is used to represent a table header in the target data table, and the plurality of data items are used to represent personnel information data;

[0008] Obtaining a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and an edge is used to establish an association relationship between the nodes;

[0009] Using each data item in the target data table, updating the data association graph to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table;

[0010] Based on the updated data association graph, determining a header feature vector of the target data table;

[0011] Obtaining a data classification model based on a graph neural network, wherein the data classification model is trained by taking the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output;

[0012] Inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table;

[0013] According to the classification probability of each table header in the target data table, the classification result of each table header in the target data table is obtained, so as to perform desensitization processing on the data in the target data table based on the classification result of each table header in the target data table.

[0014] Based on the above disclosed content, the present invention pre-constructs a data association graph based on a graph structure, wherein each node in the data association graph corresponds to a historical data item (i.e., each header in the historical data table), and an association relationship is established between each node using an edge; in this way, it is equivalent to using a graph structure to represent the data association between each header; based on this, in actual use, after obtaining the target data table, the present invention can update the aforementioned data association graph according to each data item in the target data table, so as to insert each data item in the target data table into the data association graph, thereby establishing a data association between the headers corresponding to each data item; then, the header feature vector of the target data table can be determined based on the updated data association graph; thus, the feature vector determined based on the updated data association graph contains the association relationship between each header in the target data table and between the historical headers; finally, the header feature vector is input into the data classification model to obtain the classification probability of each header in the target data table, thereby determining the classification result of each header in the target data table.

[0015] Through the above design, the present invention introduces a data association graph based on a graph structure to characterize the association relationship between each header in the data table, and before each grading, uses each header in the data table of this grading to update the data association graph, so as to establish the data association between each header in the data table of this grading; then, the header feature vector can be extracted based on the updated data association graph; in this way, the feature vector determined based on the updated data association graph contains the association relationship between each header in this data table and between each header of the historical grading; finally, the header feature vector is input into the data grading model, and the grading result of each header in this data table can be obtained; therefore, when grading data, the present invention not only relies on the characteristics of the data itself, but also combines the data association graph to retain the association relationship between the data. Based on this, the accuracy of data classification and grading can be improved, thereby ensuring the reliability of data desensitization, and therefore, it is very suitable for large-scale application and promotion.

[0016] In a possible design, the data association graph is updated using each data item in the target data table to obtain an updated data association graph, including:

[0017] Acquire semantic information of each data item in the target data table, wherein the semantic information of any data item is in vector form;

[0018] According to the semantic information of each data item, each data item is inserted into the data association graph as a new insertion node, so as to obtain the updated data association graph after the insertion is completed.

[0019] In a possible design, any node in the data association graph has a vertex identifier, wherein the vertex identifier of any node is used to uniquely represent the historical data item corresponding to the any node, and the vertex identifier and the edge are used to establish an association relationship between the nodes;

[0020] According to the semantic information of each data item, each data item is inserted into the data association graph as a new insertion node, so as to obtain the updated data association graph after the insertion is completed, including:

[0021] For the i-th data item in the target data table, obtain metadata of the i-th data item, and construct a vertex identifier of the i-th data item according to the metadata;

[0022] Calculating the similarity between the semantic information of the i-th data item and the semantic information of the historical data items corresponding to each node in the data association graph, so as to obtain multiple similarities after the calculation is completed;

[0023] Determine whether there is a similarity greater than or equal to a first threshold among the multiple similarities;

[0024] If yes, the first target node is used to replace the newly inserted node of the i-th data item, and an association relationship is established between the vertex identifier of the first target node and the vertex identifier corresponding to the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established, wherein the first target node is a node corresponding to a similarity greater than or equal to the first threshold among multiple similarities;

[0025] Increment i by 1 and reacquire the metadata of the i-th data item until i is equal to n, complete the insertion of all data items in the target data table, and obtain the updated data association graph, where the initial value of i is 1 and n is the total number of data items in the target data table.

[0026] In a possible design, if there is no similarity greater than or equal to the first threshold among the multiple similarities, the method further includes:

[0027] Determine whether there is a similarity among the multiple similarities that is greater than or equal to a second threshold and less than the first threshold;

[0028] If so, the i-th data item is inserted into the data association graph as a newly inserted node, and the newly inserted node corresponding to the i-th data item is connected to the second target node by an edge, wherein the second target node is a node corresponding to a similarity greater than or equal to the second threshold and less than the first threshold among multiple similarities;

[0029] An association relationship is established between the vertex identifier of the second target node and the vertex identifier of the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established.

[0030] In a possible design, if there is no similarity among the multiple similarities that is greater than or equal to the second threshold and less than the first threshold, the method further includes:

[0031] The i-th data item is taken as a new insertion node and inserted into the data association graph to complete the insertion of the i-th data item.

[0032] In a possible design, determining the header feature vector of the target data table based on the updated data association graph includes:

[0033] Performing sampling processing on the updated data association graph to obtain a plurality of sampling nodes;

[0034] Obtaining adjacent nodes of each sampling node, wherein the adjacent nodes of any sampling node are nodes in the updated data association graph that have an edge connection relationship with the any sampling node;

[0035] Performing feature aggregation processing on feature vectors of adjacent nodes of each sampling node, so as to obtain feature vectors of each sampling node after the feature aggregation processing;

[0036] The feature vector of each sampling node is used as the header feature vector of the target data table.

[0037] In a possible design, the updated data association graph includes a plurality of nodes, each node corresponds to a data item, any node has a vertex identifier, the vertex identifier is used to uniquely represent the data item corresponding to any node, and the nodes are associated with each other using vertex identifiers and edges;

[0038] The step of inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table includes:

[0039] Inputting the feature vector of each sampling node into the data classification model to obtain the classification probability of each sampling node;

[0040] According to the vertex identifiers and grading probabilities of the sampling nodes, the grading probabilities of the corresponding headers of the data items in the target data table are determined, so as to obtain the grading results of the corresponding headers of the data items in the target data table according to the grading probabilities of the corresponding headers of the data items in the target data table, and based on the grading results of the corresponding headers of the data items in the target data table, the data corresponding to the headers in the target data table are desensitized.

[0041] In a possible design, the classification result of the corresponding table header of each data item in the target data table is determined according to the vertex identifier and the classification probability of each sampling node, including:

[0042] For any sampling node, according to the vertex identifier of any sampling node, from the updated data association graph, determine a vertex identifier associated with the vertex identifier of any sampling node as a target identifier;

[0043] Determine whether there is a vertex identifier identical to the target identifier among the vertex identifiers corresponding to the data items in the target data table;

[0044] If so, the grading probability corresponding to any sampling node is used as the grading probability of the table header corresponding to the specified data item, so as to obtain the grading result of the table header corresponding to the specified data item based on the grading probability of the table header corresponding to the specified data item, wherein the specified data item is the data item in the target data table corresponding to the vertex identifier that is the same as the target identifier.

[0045] In a second aspect, a data classification device based on a graph neural network is provided, comprising:

[0046] An acquisition unit, used for acquiring a target data table, wherein the target data table includes a plurality of data items, each of which is used to represent a table header in the target data table, and the plurality of data items are used to represent personnel information data;

[0047] A graph forming unit is used to obtain a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and an edge is used to establish an association relationship between the nodes;

[0048] The mapping unit is further used to update the data association graph using each data item in the target data table to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table;

[0049] A feature extraction unit, used to determine a header feature vector of the target data table based on the updated data association graph;

[0050] A data classification unit, used to obtain a data classification model based on a graph neural network, wherein the data classification model is trained by taking the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output;

[0051] A data classification unit, used for inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table;

[0052] The data grading unit is also used to obtain the grading results of each header in the target data table according to the grading probability of each header in the target data table, so as to desensitize the data in the target data table based on the grading results of each header in the target data table.

[0053] In the third aspect, another data classification device based on graph neural network is provided. Taking the device as an electronic device as an example, the device includes a memory, a processor and a transceiver which are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program to execute the data classification method based on graph neural network as in the first aspect or any possible design in the first aspect.

[0054] In a fourth aspect, a storage medium is provided, on which instructions are stored. When the instructions are executed on a computer, the data classification method based on the graph neural network as described in the first aspect or any possible design of the first aspect is executed.

[0055] In a fifth aspect, a computer program product comprising instructions is provided, which, when executed on a computer, causes the computer to execute the data classification method based on a graph neural network as described in the first aspect or any possible design of the first aspect.

[0056] Beneficial effects:

[0057] (1) The present invention introduces a data association graph based on a graph structure to characterize the association relationship between each header in the data table, and before each classification, uses each header in the data table of the current classification to update the data association graph, thereby establishing the data association between each header in the data table of the current classification; then, the header feature vector can be extracted based on the updated data association graph; thus, the feature vector determined based on the updated data association graph contains the association relationship between each header in the current data table and between each header in the historical classification; finally, the header feature vector is input into the data classification model, and the classification result of each header in the current data table can be obtained; therefore, when classifying data, the present invention not only relies on the characteristics of the data itself, but also combines the data association graph to retain the association relationship between the data. Based on this, not only the accuracy of data classification and classification can be improved, and the reliability of data desensitization can be guaranteed, but also the audit workload can be reduced compared with manual audit, so that it is very suitable for large-scale application and promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic diagram of the steps of a data classification method based on a graph neural network provided in an embodiment of the present invention;

[0059] Figure 2 A schematic diagram of the structure of a data classification device based on a graph neural network provided in an embodiment of the present invention;

[0060] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0062] It should be understood that although the terms first, second, etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another unit. For example, a first unit can be referred to as a second unit, and similarly, a second unit can be referred to as a first unit without departing from the scope of the exemplary embodiments of the present invention.

[0063] It should be understood that the term "and / or" that may appear in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" that may appear in this article describes another type of association object relationship, indicating that two relationships may exist. For example, A / and B can represent two situations: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this article generally indicates that the previous and next associated objects are in an "or" relationship.

[0064] Example:

[0065] See also Figure 1As shown, the data classification method based on graph neural network provided by this embodiment introduces a data association graph based on a graph structure to characterize the association relationship between each header in the data table, and extracts the header feature vector corresponding to the data table based on the data association graph; then, the header feature vector is input into a pre-trained data classification model, and the classification probability of each header in the data table can be obtained. Finally, based on this, the classification result of each header (that is, whether it belongs to sensitive data or non-sensitive data) can be obtained; in this way, when classifying data, this method not only relies on the characteristics of the data itself, but also combines data association. Figure, to retain the correlation between data, based on this, not only can the accuracy of data classification and grading be improved, thereby ensuring the reliability of data desensitization, but also compared with manual review, the review workload is reduced. Therefore, it is very suitable for large-scale application and promotion in the field of data classification and grading; among them, for example, the method can be but not limited to running on the data grading end side, optionally, the data grading end can be but not limited to a personal computer, a tablet computer or a smart phone. It can be understood that the aforementioned execution subject does not constitute a limitation on the embodiments of the present application. Accordingly, the operation steps of the method can be but not limited to the following steps S1 to S7.

[0066] S1. Obtain a target data table, wherein the target data table includes multiple data items, each data item is used to represent a table header in the target data table, and multiple data items are used to represent personnel information data; in this embodiment, the target data table is a data table to be desensitized, and each data item in the target data table is used to represent a table header; if the target data table is a personnel information table, then the table header in the table may include but is not limited to: name, age, gender, ID number, salary, work number, etc., and the aforementioned name, age and other information are personnel information data; of course, the table headers in different data tables are different, which will not be repeated here.

[0067] After obtaining the target data table, each data item in the target data table can be inserted into the data association diagram to establish an association relationship between the corresponding table headers of each data item in the target data table; wherein, the insertion process of each data item in the target data table can be but is not limited to as shown in the following steps S2 and S3.

[0068] S2. Obtain a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and edges are used to establish association relationships between the nodes; in specific implementation, the data association graph is constructed during the last data classification and grading (if it is the first grading, the data association graph can be constructed using the data items in the target data table), and the graph is essentially an undirected graph, each node in the graph represents a data item (i.e., a header in the historical data table), and if there is an association relationship between two nodes, the two nodes are connected by an edge; at the same time, if there is no association relationship, there is no edge between the two nodes; of course, the association relationship between the nodes is measured by semantic similarity, that is, the distance between the vectors of the node corresponding to the header.

[0069] At the same time, any node in the data association graph is taken as an example to have a vertex identifier, wherein the vertex identifier of any node is used to uniquely represent the historical data item corresponding to the any node, and the vertex identifiers and edges are used to establish an association relationship between the nodes; in this embodiment, assuming that nodes A and B in the data association graph have an association relationship, then the vertex identifier of node A (i.e., vid, assumed to be A1) has an association relationship with the vertex identifier of node B (assumed to be B1), such as A1-B1, and is recorded in the information of nodes A and B; of course, the above examples are for illustration only and are not limited to this.

[0070] Furthermore, the reason for introducing the association relationship between data through the data association graph is that in actual use, isolated data does not generate data value, such as isolated mobile phone numbers, names, addresses, etc. do not generate value. Only when the data are connected, do they have data value; for example, department staff salary information, where a single name has no value, only when it is connected with employee_id (identity proof) and salary, it is necessary to perform security audits on this data; therefore, traditional technology classifies each data independently without considering the association between data, which will lead to the problem of misclassification, thereby reducing the accuracy of classification; thus, this embodiment uses a data association graph to retain the association relationship between data during classification, thereby improving the accuracy of data classification.

[0071] Optionally, the process of updating the data association graph based on each data item in the target data table may be, but is not limited to, as shown in the following step S3.

[0072] S3. Using each data item in the target data table, update the data association graph to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table; in specific applications, this embodiment measures the similarity between each data item in the target data table and the historical data item corresponding to each node through the semantic information of each data item in the target data table and the semantic information of the historical data item corresponding to each node in the data association graph; then, based on the similarity, insert each data item in the target data table into the data association graph as a newly inserted node; wherein the aforementioned process may be but is not limited to as shown in the following steps S31 and S32.

[0073] S31. Obtain the semantic information of each data item in the target data table, wherein the semantic information of any data item is in vector form; in specific implementation, it is equivalent to converting each header into a vector, so that the vector corresponding to each header is used as the semantic information of the data item corresponding to each header; optionally, each header in the target data table can be embedded vectorized to obtain an embedded vector (Embedding vector), so that the embedded vector of each header can be used as the semantic information of the data item corresponding to each header; of course, data Embedding vectorization is a common technology for data vector representation, and its principle will not be repeated here.

[0074] Furthermore, if there is no annotation information for each header in the target data table (for example, the header is an ID, and the meaning of the ID is not explained, that is, whether it is an identity ID or a work number ID), then for the header without annotation information, the regular expression pattern matching technology is used to sample the data in the column where the header without annotation information is located, so as to identify the meaning of the header through the sampled data, and then generate the corresponding embedding vector; at the same time, the metadata of each header in the target data table can also be generated (which is data used to describe the attributes, source, content and other information of the data itself), so that based on the metadata of each header, when updating the data association graph, the vertex identifiers of the data items corresponding to each header are generated, so as to facilitate the establishment of the association relationship between each data item in the target data table and the original node in the data association graph based on the vertex identifier when the subsequent data items are inserted.

[0075] After the semantic information of each data item in the target data table is obtained, each data item in the target data table can be inserted based on the semantic information, and the process can be but not limited to the one shown in the following step S32.

[0076] S32. Based on the semantic information of each data item, each data item is inserted into the data association graph as a new insertion node, so as to obtain the updated data association graph after the insertion is completed; in specific applications, each data item is inserted based on the similarity between the semantic information of each data item in the target data table and the semantic information of the historical data items corresponding to each historical node in the data association graph, wherein, for example, but not limited to, the following steps S32a to S32e can be used to complete the insertion of each data item in the target data table.

[0077] S32a. For the ith data item in the target data table, obtain the metadata of the ith data item, and construct a vertex identifier of the ith data item based on the metadata; in this embodiment, for example, but not limited to, the metadata of the ith data item is directly used as the vertex identifier of the ith data item; in this way, after obtaining the vertex identifier of the ith data item, the semantic similarity can be calculated, and the process can be but not limited to as shown in the following step S32b.

[0078] S32b. Calculate the similarity between the semantic information of the ith data item and the semantic information of the historical data items corresponding to each node in the data association graph, so as to obtain multiple similarities after the calculation is completed; in the specific implementation, it has been explained above that the semantic information of the ith data item is actually a vector, so the semantic information of the historical data items corresponding to each node in the data association graph is also a vector; based on this, by calculating the distance between the vectors (such as cosine distance or Euclidean distance), the similarity between the ith data item and each node can be obtained; then, the ith data item can be inserted based on the similarity, and the process can be but not limited to as shown in the following steps S32c and S32d.

[0079] S32c. Determine whether there is a similarity greater than or equal to a first threshold value among multiple similarities; in a specific implementation, the smaller the distance between the semantic information corresponding to the i-th data item and the semantic information of the historical data item corresponding to any node, the higher the similarity between the two. Therefore, the maximum similarity can be screened out from the multiple similarities (i.e. determined according to the minimum distance value), and then, determine whether the maximum similarity is greater than or equal to a first threshold value (which can be set in advance); wherein, if it is greater than or equal to the first threshold value, it means that there is a node in the data association graph that is most similar to the i-th data item. At this time, there is no need to establish a new node, and the node that is most similar to it can be used instead. The process can be but is not limited to the following step S32d.

[0080] S32d. If yes, the first target node is used to replace the newly inserted node of the i-th data item, and an association relationship is established between the vertex identifier of the first target node and the vertex identifier corresponding to the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established, wherein the first target node is a node corresponding to a similarity greater than or equal to the first threshold among multiple similarities; in a specific implementation, this step is explained with an example, assuming that the node corresponding to the similarity greater than or equal to the first threshold is node A, that is, the first target node is node A, and its corresponding vertex identifier is A1, and at the same time, the vertex identifier corresponding to the i-th data item is D1, then, at this time, node A is used to replace the newly inserted node corresponding to the i-th data item (that is, there is no need to insert a new node in the data association graph), and then, A1 is associated with D1 (which can be expressed as A1-D1), and then the association relationship is stored in the information of node A, and at this time, node A stores its associated data items.

[0081] Similarly, if there is no similarity greater than or equal to the first threshold among the aforementioned multiple similarities, then it is necessary to continue to judge and perform the following steps, namely: (1) judge whether there is a similarity greater than or equal to the second threshold and less than the first threshold among the multiple similarities; (2) if so, insert the i-th data item into the data association graph as a newly inserted node, and connect the newly inserted node corresponding to the i-th data item with the second target node using an edge, wherein the second target node is the node corresponding to the similarity greater than or equal to the second threshold and less than the first threshold among the multiple similarities; (3) establish an association relationship between the vertex identifier of the second target node and the vertex identifier of the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established.

[0082] Optionally, an example is used to illustrate. Assume that the node corresponding to the similarity among multiple similarities that is greater than or equal to the second threshold and less than the first threshold is node B, and its corresponding node identifier is B1; at this time, the i-th data item needs to be inserted into the data association graph as a newly inserted node (assuming it is D), that is, a new node D will be added to the data association graph; then, the newly added node D and node B are connected with an edge to establish an association relationship between the two; at the same time, the vertex identifiers of the two are also associated, that is, B1 is associated with D1 (expressed as B1-D1); finally, the association relationship is stored in node B and node D; in this way, the insertion of the i-th data item can be completed.

[0083] In addition, when there is no similarity greater than or equal to the second threshold among the multiple similarities, that is, when the maximum similarity is less than the second threshold, at this time, the i-th data item is directly inserted into the data association graph as a newly inserted node to complete the insertion of the i-th data item; as explained on the basis of the above example, that is, a new node D is added to the data association graph, and there is no edge connection between the node D and the original node in the data association graph. It is an isolated node, indicating that it has no association relationship with the remaining nodes.

[0084] Based on the above explanation, the insertion of the i-th data item can be completed, and then, based on the same principle, the remaining data items in the target data table can be inserted into the data association diagram, thereby completing the update of the data association diagram; wherein, the insertion process of the remaining data items can be but is not limited to as shown in the following step S32e.

[0085] S32e. Increment i by 1, and reacquire the metadata of the i-th data item until i is equal to n, and complete the insertion of all data items in the target data table to obtain the updated data association graph, wherein the initial value of i is 1, and n is the total number of data items in the target data table; in this embodiment, the updated data association graph includes each data item in the target data table, that is, it includes multiple nodes, each node corresponds to a data item, and any node has a vertex identifier, which is used to uniquely represent the data item corresponding to any node, and the nodes are still associated with each other using vertex identifiers and edges.

[0086] In addition, after obtaining the updated data association graph, in order to prevent the scale of the graph from being excessively expanded, for example, but not limited to, pruning operations may be performed on the updated data association graph to reduce the scale of the graph, thereby reducing the complexity of subsequent sampling and aggregation; of course, pruning operations are a commonly used technique for reducing data scale, and the principles thereof will not be elaborated on.

[0087] Therefore, through the aforementioned steps S31 and S32, and the sub-steps corresponding to step S32, each data item in the target data table can be inserted into the data association graph, thereby establishing the association relationship between the various table headers in the target data table, as well as the association relationship with the historical table headers; in this way, by introducing a data association graph based on a graph structure, the semantic consistency problem of the data can be effectively handled, and a data graph that is both accurate and efficient can be constructed, thereby facilitating the maintenance of the association relationship between data in the subsequent data classification process.

[0088] After completing the insertion of each data item in the target data table and obtaining the updated data association graph, the header feature vector of the target data table can be generated based on the updated data association graph, so as to subsequently perform the classification of each header in the target data table based on this; wherein, the extraction process of the header feature vector can be but is not limited to as shown in the following step S4.

[0089] S4. Based on the updated data association graph, determine the header feature vector of the target data table. In a specific implementation, for example, but not limited to, the following steps S41 to S44 may be used to determine the header feature vector of the target data table.

[0090] S41. Sampling the updated data association graph to obtain multiple sampling nodes. In a specific implementation, multiple nodes in the updated data association graph can be randomly selected as sampling nodes. Then, the feature vector of each sampling node can be determined based on the adjacent nodes of the sampling node. The process can be but not limited to the following steps S42 to S44.

[0091] S42. Obtain the adjacent nodes of each sampling node, wherein the adjacent nodes of any sampling node are nodes in the updated data association graph that have an edge connection relationship with the any sampling node; in a specific implementation, for example, the adjacent nodes of any sampling node may include but are not limited to a first neighbor node and a second neighbor node, wherein the first neighbor node is a node that has an edge connection relationship with the any sampling node (that is, the distance with the any sampling node is 1), and the second neighbor node is a node that has a distance of 2 from the any sampling node (that is, the path from any second neighbor node to the any sampling node contains two edges), that is, assuming that node A is directly connected to node B and node C, and node B is connected to nodes F and T, and node F is connected to node Y, then the first neighbor nodes of node A are node B and node C, and the second neighbor nodes are nodes F and T.

[0092] Furthermore, when obtaining the adjacent nodes of each sampling node, since each sampling node has multiple adjacent nodes, a random sampling method can be used to perform sampling processing from the adjacent nodes of each sampling node, so as to obtain the adjacent nodes finally used; for example, assuming that the first neighbor nodes corresponding to node A have 10, at this time, 3 can be randomly sampled as the adjacent nodes finally used; similarly, its corresponding second neighbor nodes have 20, then 6 can be randomly sampled as the adjacent nodes finally used; at this time, the actual number of adjacent nodes obtained is 9; of course, the number of randomly sampled adjacent nodes can be specifically set according to actual use, and is not specifically limited here.

[0093] In this way, after the adjacent nodes of each sampling node are obtained, the feature vector of each sampling node can be constructed according to the feature vectors of each adjacent node. The process can be but not limited to the following step S43.

[0094] S43. Perform feature aggregation processing on the feature vectors of the neighboring nodes of each sampling node to obtain the feature vector of each sampling node after the feature aggregation processing; in specific implementation, for example, but not limited to, first perform feature aggregation processing on the feature vectors of the second neighboring nodes of each sampling node to obtain a first aggregated feature vector; then, perform feature aggregation processing on the feature vectors of the first neighboring nodes of each sampling node and the corresponding first aggregated feature vector to obtain the feature vector of each sampling node after the feature aggregation processing; of course, the feature vector of any neighboring node is its corresponding semantic information, and feature aggregation is a commonly used technology in the feature extraction process, and its principle will not be repeated.

[0095] After the feature vectors of each sampling node are obtained, the feature vectors of each sampling node can be used as the header feature vectors of the target data table. The process can be but is not limited to the following step S44.

[0096] S44. Using the feature vector of each sampling node as the header feature vector of the target data table.

[0097] In this way, through the aforementioned steps S41 to S44, the feature vectors of the multi-hop neighbors of each sampling node are used to generate the embedded representation of the sampling node itself, so that the feature vector of the sampling node can contain the association relationship of each node associated with it, thereby retaining the association relationship between the data; at the same time, through this hierarchical feature aggregation method, the local structural information of the node in the graph can be effectively captured, thereby providing accurate node embedding for subsequent node classification tasks.

[0098] After obtaining the header feature vector of the target data table, each header in the target data table can be graded based on it; wherein, the present embodiment pre-trains a data grading model based on a graph neural network, so that in actual use, the header feature vector of the aforementioned target data table is input into the model, and the corresponding grading result can be obtained; optionally, the process of obtaining the data grading model can be but is not limited to as shown in the following step S5.

[0099] S5. Obtain a data classification model based on a graph neural network, wherein the data classification model is trained with the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output; in specific implementation, the construction process of the header feature vector of each sample data table is the same as that of the header feature vector of the aforementioned target data table, which will not be repeated here; at the same time, when training the model, the data association graph can be updated first based on each sample data table, and after the update, the original data is obtained from the data association graph, that is, it contains a number of sample sampling nodes and corresponding adjacent nodes, and then the obtained data is divided into a training set and a test set; then, the feature vectors of the sample sampling nodes corresponding to the adjacent nodes in the training set are used to aggregate the feature vectors of each sample sampling node in the training set, thereby obtaining the header feature vector of the sample data table corresponding to the training set; then, the feature vectors of the sampling nodes in the training set (that is, the header feature vectors corresponding to the sample data table) are used to train the graph neural network; finally, the test set is used to verify the accuracy of the trained graph neural network, and when the model indicators meet the preset conditions, the data classification model is obtained.

[0100] Thus, after obtaining the data classification model, the header feature vector of the aforementioned target data table is input into the data classification model to obtain the classification result of the corresponding header, and the process may be but is not limited to the following steps S6 and S7.

[0101] S6. Input the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table; in specific implementation, input the feature vector of each sampling node into the data classification model to obtain the classification probability of each sampling node; then, according to the vertex identifier and the classification probability of each sampling node, determine the classification probability of the header corresponding to each data item in the target data table, so as to obtain the classification result of the header corresponding to each data item in the target data table according to the classification probability of the header corresponding to each data item in the target data table.

[0102] Specifically, there are multiple classification probabilities for any sampling node. In this embodiment, the largest classification probability is taken as the actual classification probability of any sampling node. At the same time, as mentioned above, the association relationship between each node is established using the vertex identifier. Therefore, the classification results of each header in the target data table can be determined based on the vertex identifier of each sampling node and the association relationship between the vertex identifiers of the remaining nodes. The process is shown in the following step S7.

[0103] S7. According to the classification probabilities of the various table headers in the target data table, the classification results of the various table headers in the target data table are obtained, so that the data in the target data table are desensitized based on the classification results of the various table headers in the target data table. In the specific implementation, any sampling node is taken as an example for explanation, and the process may be but not limited to: (1) for any sampling node, according to the vertex identifier of any sampling node, a vertex identifier associated with the vertex identifier of any sampling node is determined from the updated data association graph as the target identifier; (2) it is determined whether there is a vertex identifier identical to the target identifier among the vertex identifiers corresponding to each data item in the target data table; (3) if so, the classification probability corresponding to any sampling node is used as the classification probability of the table header corresponding to the specified data item, so that the classification result of the table header corresponding to the specified data item is obtained based on the classification probability of the table header corresponding to the specified data item, wherein the specified data item is a data item in the target data table corresponding to the vertex identifier identical to the target identifier.

[0104] Furthermore, the classification result of any header is that the any header is a sensitive header or a non-sensitive header. When any header is a sensitive header, the data in the column where any header is located in the target data table is sensitive data; in this way, when desensitizing processing is performed, the data corresponding to the sensitive header in the target data table is desensitized, while the data corresponding to the non-sensitive header does not need to be desensitized.

[0105] In the specific implementation, the aforementioned step S7 is explained by an example. Assuming that any sampling node is node A, the association relationship of its vertex identifier A1 is: A1-D1, A1-B1; A1-F1; wherein D1 is the vertex identifier of node D, B1 is the vertex identifier of node B, and F1 is the vertex identifier of node F. At this time, D1, B1 and F1 are used as target identifiers.

[0106] Next, determine whether there is a vertex identifier that is the same as the target identifier among the vertex identifiers corresponding to the data items in the target data table; assuming that the data items in the target data table are the first data item, the second data item and the third data item, and the vertex identifiers corresponding to the three are F1, G1 and H1 respectively, then it can be determined that the vertex identifier of the first data item is the same as the target identifier F1; therefore, the actual grading probability of node A can be used as the grading probability of the table header corresponding to the first data item; then, the grading result of the table header corresponding to the first data item can be obtained based on the grading probability of the table header corresponding to the first data item.

[0107] Furthermore, this embodiment is provided with a classification threshold. When the classification probability of the header corresponding to any data item in the target data table is greater than or equal to the classification threshold, the header corresponding to the any data item is classified as a sensitive header. At this time, if the first data item is represented as an ID card number, then the data corresponding to the ID card number column in the target data table is classified as sensitive data.

[0108] Of course, if it is determined that there is no vertex identifier identical to the target identifier among the vertex identifiers corresponding to the data items in the target data table; at this time, the vertex identifier associated with the vertex identifier of the next sampling node is determined from the updated data association graph; then steps (1) to (3) in the aforementioned step S7 are repeated until all sampling nodes are polled, and the classification results of all table headers in the target data table can be obtained.

[0109] Among them, if the vertex identifiers associated with the vertex identifiers of all sampling nodes do not contain the vertex identifiers of each data item in the target data table, then it means that all headers in the target data table are not sensitive headers; at the same time, if the vertex identifiers associated with all sampling nodes contain the vertex identifiers of some data items in the target data table, then the headers corresponding to the remaining data items in the target data table are non-sensitive headers.

[0110] As explained on the basis of the above example, the vertex identifiers corresponding to each data item in the target data table are F1, G1 and H1. After polling the identifiers associated with the vertex identifiers of all sampling nodes, only F1 and G1 are determined to exist in the vertex identifiers with the associated relationship, and the corresponding classification probabilities are greater than or equal to the classification threshold. Then, the headers represented by the data items corresponding to F1 and G1 (that is, the headers corresponding to the first data item and the second data item) are sensitive headers, and the headers represented by the data items corresponding to H1 are non-sensitive headers; then, when desensitizing, the data in the column where the headers represented by the first and second data items are located are desensitized (assuming that they are the ID number and name, then the content corresponding to the ID number and name in the table is desensitized as sensitive data); of course, the above example is only for illustration, and this embodiment is not limited to this.

[0111] Therefore, through the data classification method based on graph neural network described in detail in the aforementioned steps S1 to S7, the present invention introduces a data association graph based on a graph structure to characterize the association relationship between each header in the data table, and extracts the header feature vector corresponding to the data table based on the data association graph; then, the header feature vector is input into the pre-trained data classification model, and the classification probability of each header in the data table can be obtained. Finally, based on this, the classification result of each header (that is, whether it is sensitive data or non-sensitive data) can be obtained; in this way, when classifying data, the present invention not only relies on the characteristics of the data itself, but also combines the data association graph to retain the association relationship between the data. Based on this, it can not only improve the accuracy of data classification and grading, thereby ensuring the reliability of data desensitization, but also reduce the audit workload compared to manual review. Therefore, it is very suitable for large-scale application and promotion in the field of data classification and grading.

[0112] like Figure 2 As shown, the second aspect of this embodiment provides a hardware device for implementing the data classification method based on graph neural network described in the first aspect of the embodiment, including:

[0113] The acquisition unit is used to acquire a target data table, wherein the target data table contains multiple data items, each data item is used to represent a table header in the target data table, and multiple data items are used to represent personnel information data.

[0114] The graph forming unit is used to obtain a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and edges are used to establish association relationships between the nodes.

[0115] The mapping unit is further used to update the data association graph using each data item in the target data table to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table.

[0116] The feature extraction unit is used to determine the header feature vector of the target data table based on the updated data association graph.

[0117] A data classification unit is used to obtain a data classification model based on a graph neural network, wherein the data classification model is trained with the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output.

[0118] The data classification unit is used to input the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table.

[0119] The data grading unit is also used to obtain the grading results of each header in the target data table according to the grading probability of each header in the target data table, so as to desensitize the data in the target data table based on the grading results of each header in the target data table.

[0120] The working process, working details and technical effects of the device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0121] like Figure 3 As shown, the third aspect of this embodiment provides another data classification device based on graph neural network. Taking the device as an electronic device as an example, it includes: a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program to execute the data classification method based on graph neural network as described in the first aspect of the embodiment.

[0122] For example, the memory may include, but is not limited to, random access memory (RAM), read only memory (ROM), flash memory, first input first output (FIFO) and / or first in last out (FILO), etc. Specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). At the same time, the processor may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state.

[0123] 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. For example, the processor may not be limited to a microprocessor of the STM32F105 series, a reduced instruction set computer (RISC) microprocessor, an X86 architecture processor, or a processor with an integrated embedded neural network processing unit (NPU); the transceiver may be, but is not limited to, a wireless fidelity (WIFI) wireless transceiver, a Bluetooth wireless transceiver, a General Packet Radio Service (GPRS) wireless transceiver, a ZigBee protocol (a low-power LAN protocol based on the IEEE802.15.4 standard, ZigBee) wireless transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. In addition, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0124] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0125] The fourth aspect of this embodiment provides a storage medium storing instructions for the data classification method based on graph neural network as described in the first aspect of the embodiment, that is, the storage medium stores instructions, and when the instructions are run on a computer, the data classification method based on graph neural network as described in the first aspect of the embodiment is executed.

[0126] The storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, a CD, a hard disk, a flash memory, a USB flash drive and / or a memory stick, and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0127] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0128] The fifth aspect of this embodiment provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the data classification method based on graph neural network as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0129] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. 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 classification method based on graph neural network, characterized in that: include: Acquire a target data table, wherein the target data table includes a plurality of data items, each of which is used to represent a table header in the target data table, and the plurality of data items are used to represent personnel information data; Obtaining a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and an edge is used to establish an association relationship between the nodes; Using each data item in the target data table, updating the data association graph to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table; Based on the updated data association graph, determining a header feature vector of the target data table; Obtaining a data classification model based on a graph neural network, wherein the data classification model is trained by taking the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output; Inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table; According to the classification probability of each table header in the target data table, a classification result of each table header in the target data table is obtained, so as to perform desensitization processing on the data in the target data table based on the classification result of each table header in the target data table; Determining a header feature vector of the target data table based on the updated data association graph includes: Performing sampling processing on the updated data association graph to obtain a plurality of sampling nodes; Obtaining adjacent nodes of each sampling node, wherein the adjacent nodes of any sampling node are nodes in the updated data association graph that have an edge connection relationship with the any sampling node; Performing feature aggregation processing on feature vectors of adjacent nodes of each sampling node, so as to obtain feature vectors of each sampling node after the feature aggregation processing; Using the feature vector of each sampling node as the header feature vector of the target data table; The updated data association graph includes a plurality of nodes, each node corresponds to a data item, any node has a vertex identifier, the vertex identifier is used to uniquely represent the data item corresponding to any node, and the nodes are associated with each other using vertex identifiers and edges; The step of inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table includes: Inputting the feature vector of each sampling node into the data classification model to obtain the classification probability of each sampling node; According to the vertex identifiers and classification probabilities of the sampling nodes, the classification probabilities of the corresponding headers of the data items in the target data table are determined, so as to obtain the classification results of the corresponding headers of the data items in the target data table according to the classification probabilities of the corresponding headers of the data items in the target data table, and perform desensitization processing on the data corresponding to the headers of the target data table based on the classification results of the corresponding headers of the data items in the target data table; According to the vertex identifiers and classification probabilities of the sampling nodes, the classification results of the corresponding table headers of the data items in the target data table are determined, including: For any sampling node, according to the vertex identifier of any sampling node, from the updated data association graph, determine a vertex identifier associated with the vertex identifier of any sampling node as a target identifier; Determine whether there is a vertex identifier identical to the target identifier among the vertex identifiers corresponding to the data items in the target data table; If so, the grading probability corresponding to any sampling node is used as the grading probability of the table header corresponding to the specified data item, so as to obtain the grading result of the table header corresponding to the specified data item based on the grading probability of the table header corresponding to the specified data item, wherein the specified data item is the data item in the target data table corresponding to the vertex identifier that is the same as the target identifier.

2. The method according to claim 1, characterized in that Using each data item in the target data table, updating the data association graph to obtain an updated data association graph includes: Acquire semantic information of each data item in the target data table, wherein the semantic information of any data item is in vector form; According to the semantic information of each data item, each data item is inserted into the data association graph as a new insertion node, so as to obtain the updated data association graph after the insertion is completed.

3. The method according to claim 2, characterized in that Any node in the data association graph has a vertex identifier, wherein the vertex identifier of any node is used to uniquely represent the historical data item corresponding to the node, and the vertex identifiers and edges are used to establish associations between the nodes; According to the semantic information of each data item, each data item is inserted into the data association graph as a new insertion node, so as to obtain the updated data association graph after the insertion is completed, including: For the i-th data item in the target data table, obtain metadata of the i-th data item, and construct a vertex identifier of the i-th data item according to the metadata; Calculating the similarity between the semantic information of the i-th data item and the semantic information of the historical data items corresponding to each node in the data association graph, so as to obtain multiple similarities after the calculation is completed; Determine whether there is a similarity greater than or equal to a first threshold among the multiple similarities; If yes, the first target node is used to replace the newly inserted node of the i-th data item, and an association relationship is established between the vertex identifier of the first target node and the vertex identifier corresponding to the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established, wherein the first target node is a node corresponding to a similarity greater than or equal to the first threshold among multiple similarities; Increment i by 1 and reacquire the metadata of the i-th data item until i is equal to n, complete the insertion of all data items in the target data table, and obtain the updated data association graph, where the initial value of i is 1 and n is the total number of data items in the target data table.

4. The method according to claim 3, characterized in that If there is no similarity greater than or equal to the first threshold among the multiple similarities, the method further includes: Determine whether there is a similarity among the multiple similarities that is greater than or equal to a second threshold and less than the first threshold; If so, the i-th data item is inserted into the data association graph as a newly inserted node, and the newly inserted node corresponding to the i-th data item is connected to the second target node by an edge, wherein the second target node is a node corresponding to a similarity greater than or equal to the second threshold and less than the first threshold among multiple similarities; An association relationship is established between the vertex identifier of the second target node and the vertex identifier of the i-th data item, so as to complete the insertion of the i-th data item after the association relationship is established.

5. The method according to claim 4, characterized in that If there is no similarity among the multiple similarities that is greater than or equal to the second threshold and less than the first threshold, the method further includes: The i-th data item is taken as a new insertion node and inserted into the data association graph to complete the insertion of the i-th data item.

6. A data classification device based on graph neural network, characterized in that: include: An acquisition unit, used for acquiring a target data table, wherein the target data table includes a plurality of data items, each of which is used to represent a table header in the target data table, and the plurality of data items are used to represent personnel information data; A graph forming unit is used to obtain a data association graph, wherein the data structure of the data association graph is a graph structure, the data association graph includes a plurality of nodes, each node corresponds to a historical data item, and an edge is used to establish an association relationship between the nodes; The mapping unit is further used to update the data association graph using each data item in the target data table to obtain an updated data association graph, wherein the updated data association graph includes all data items in the target data table; A feature extraction unit, used to determine a header feature vector of the target data table based on the updated data association graph; A data classification unit, used to obtain a data classification model based on a graph neural network, wherein the data classification model is trained by taking the header feature vector of each sample data table as input and the classification probability of each header in each sample data table as output; A data classification unit, used for inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table; The data classification unit is further used to obtain the classification results of each header in the target data table according to the classification probability of each header in the target data table, so as to perform desensitization processing on the data in the target data table based on the classification results of each header in the target data table; Determining a header feature vector of the target data table based on the updated data association graph includes: Performing sampling processing on the updated data association graph to obtain a plurality of sampling nodes; Obtaining adjacent nodes of each sampling node, wherein the adjacent nodes of any sampling node are nodes in the updated data association graph that have an edge connection relationship with the any sampling node; Performing feature aggregation processing on feature vectors of adjacent nodes of each sampling node, so as to obtain feature vectors of each sampling node after the feature aggregation processing; Using the feature vector of each sampling node as the header feature vector of the target data table; The updated data association graph includes a plurality of nodes, each node corresponds to a data item, any node has a vertex identifier, the vertex identifier is used to uniquely represent the data item corresponding to any node, and the nodes are associated with each other using vertex identifiers and edges; The step of inputting the header feature vector of the target data table into the data classification model to obtain the classification probability of each header in the target data table includes: Inputting the feature vector of each sampling node into the data classification model to obtain the classification probability of each sampling node; According to the vertex identifiers and classification probabilities of the sampling nodes, the classification probabilities of the corresponding headers of the data items in the target data table are determined, so as to obtain the classification results of the corresponding headers of the data items in the target data table according to the classification probabilities of the corresponding headers of the data items in the target data table, and perform desensitization processing on the data corresponding to the headers of the target data table based on the classification results of the corresponding headers of the data items in the target data table; According to the vertex identifiers and classification probabilities of the sampling nodes, the classification results of the corresponding table headers of the data items in the target data table are determined, including: For any sampling node, according to the vertex identifier of any sampling node, from the updated data association graph, determine a vertex identifier associated with the vertex identifier of any sampling node as a target identifier; Determine whether there is a vertex identifier identical to the target identifier among the vertex identifiers corresponding to the data items in the target data table; If so, the grading probability corresponding to any sampling node is used as the grading probability of the table header corresponding to the specified data item, so as to obtain the grading result of the table header corresponding to the specified data item based on the grading probability of the table header corresponding to the specified data item, wherein the specified data item is the data item in the target data table corresponding to the vertex identifier that is the same as the target identifier.

7. A computer program product comprising instructions, characterized in that When the instructions are executed on a computer, the computer is caused to execute the data classification method based on graph neural network as described in any one of claims 1 to 5.

Images