A method and device for extracting important factors of taxpayer enterprise classification

Abstract

The application discloses a kind of important factor extraction method and device for taxpayer enterprise classification, comprising: first, data cleaning is carried out to original data, and incomplete data is eliminated, and is constructed into heterogeneous graph data form, for subsequent step;Second, construct the heterogeneous graph neural network based on attention mechanism, and utilize semi-supervised algorithm to optimize network, realize high-precision rate abnormal enterprise detection result;Third, design the heterogeneous graph interpreter based on meta-path disturbance, realize the extraction of important factor influencing abnormal data detection;Finally, combine abnormal classification result with important factor, in the form of visualization, deliver to professional tax inspection personnel for subsequent use.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning explainability, and particularly relates to an important factor extraction method and device for taxpayer enterprise classification. BACKGROUND

[0002] Tax revenue is an important source of finance, and the tax evasion behavior of taxpaying enterprises will have a huge impact on finance. At present, using a computer machine learning has become a method that can efficiently and accurately judge the abnormal behavior of enterprises.

[0003] In related technologies, a large amount of data is used to realize abnormal data detection, and a neural network method is used as a whole. The core idea is to use part of the tax data with labels to train a neural network model, and to use it to predict unlabeled enterprises to determine whether they are abnormal. However, due to the black box nature of the neural network, its judgment reasoning process is unknowable to people, and people cannot fully trust the output results of the network. Therefore, in sensitive industries such as medical treatment, finance, and tax, the abnormal determination of data needs rigorous explanation. After relevant personnel obtain the abnormal data detection result, they cannot directly qualify the data and need to further explore the important factors affecting the data discrimination.

[0004] Therefore, how to extract important factors affecting the discrimination of abnormal data has become a key problem for those skilled in the art.

[0005] The following documents provide referenceable important factor extraction techniques that affect the discrimination result:

[0006] [1] Mu Chang. Click rate important factor extraction method, system, terminal device and medium [P]. Guangdong Province: CN114462499A, 2022-05-10.

[0007] [2] Liao Jingpeng. A landscape experience influencing factor and its emotional inclination and importance analysis method [P]. Hunan Province: CN114565011A, 2022-05-31.

[0008] Document 1 proposes a method for extracting important factors affecting click rate. Based on the support and confidence under different threshold settings, the geodesic line filtering is performed on the filtered candidate set to obtain the target item set, and the click rate important factor is further extracted from the target item set.

[0009] Document 2 provides a method for extracting important factors affecting landscape experience. Based on the maximum entropy model, the P value of the hypothesis test result of the parameter in the normal distribution is used to filter out landscape experience influencing factors that are not statistically significant. The positive and negative values of the weight are used to judge the positive and negative inclination of the landscape experience influencing factors, and the t value is used to screen significant influencing factors and sort the influence degree.

[0010] The method proposed in document 1 has certain reference significance, that is, by setting a threshold function, it is judged whether the factor will affect the final result. However, in a real scene, it is difficult to set the threshold function, lack of expert knowledge, and an inappropriate threshold function may cause a large error in the result, which is unacceptable for data-sensitive industries. The method in document 2 is only applicable to a certain specific model, and its method has poor portability, and cannot be accurately adapted to different industries. In addition, using the positive and negative values of the weight to judge the influence factor tendency is only applicable to the case where the data are independent of each other, and cannot handle the scene where some data are closely dependent, such as the tax scene.

[0011] Currently, scholars have made relevant researches on the important factor extraction method in the tax scene, but these methods generally focus on simple models, such as designing a decision tree to judge whether a taxpaying enterprise is abnormal, and further extracting important factors affecting the classification result. However, when facing more complex and multi-dimensional data, the above method is difficult to meet the requirements. SUMMARY

[0012] The purpose of the present application is to provide an important factor extraction method and device for taxpayer enterprise classification, by constructing a transaction heterogeneous graph in the tax scene, using an explainable method based on meta-path perturbation, obtaining important factors affecting the abnormal data discrimination, and improving the credibility of the abnormal enterprise data detection result in the tax field.

[0013] The present application is realized by adopting the following technical solutions:

[0014] An important factor extraction method for taxpayer enterprise classification, comprising the following steps:

[0015] Step one, the data preprocessing module performs heterogeneous graph processing on the original data;

[0016] Step two, constructing a heterogeneous graph neural network model based on an attention mechanism;

[0017] Step three, using a semi-supervised classification method to optimize the heterogeneous graph neural network model;

[0018] Step four, designing a heterogeneous graph interpreter based on meta-path perturbation to realize the extraction of important factors affecting abnormal data detection.

[0019] The further improvement of the present application is that in step one, the original data includes invoice data generated from the enterprise operation process and data that can be queried on the Internet.

[0020] The further improvement of the present application is that in step one, the specific implementation method is as follows:

[0021] (1) data cleaning is performed on original data, and data with incomplete content is removed;

[0022] (2) based on data categories, data of different categories are defined as nodes of different categories, meanwhile, different categories of edges are constructed according to the association information between data; the same dimension feature vector is defined for each node;

[0023] (3) a heterogeneous graph data is constructed according to the defined nodes and edges.

[0024] The further improvement of the application is that in step two, the embedding representation of the data node under the heterogeneous graph is learned through the hierarchical attention mechanism, which is used for subsequent data detection and important factor extraction; the hierarchical attention includes two stages: node-level attention learning and semantic-level attention learning.

[0025] The further improvement of the application is that in step two, the specific implementation method is as follows:

[0026] (1) a heterogeneous graph neural network based on node-level attention is constructed, the neighborhood node features obtained from the above heterogeneous graph construction are converged through layer node-level attention, the neighbor node weights based on the meta path are learned by defining different meta paths, and the embedding representation of the node with specific semantics is obtained by aggregating them;

[0027] (2) a heterogeneous graph neural network layer based on semantic-level attention is constructed, the weight of each meta path is learned by semantic-level attention, and the semantic-specific node embedding representation obtained in the front is fused through semantic-level attention;

[0028] (3) the different neural network layers are combined to form a heterogeneous graph neural network based on the hierarchical attention mechanism.

[0029] The further improvement of the application is that in step three, a self-training semi-supervised algorithm is used in combination with the heterogeneous graph neural network of step two to realize end-to-end model optimization and realize the data anomaly detection task under the background of lacking labeled data.

[0030] The further improvement of the application is that in step three, the specific implementation method is as follows:

[0031] (1) the node embedding representation obtained based on the hierarchical attention mechanism is used to calculate the loss with the existing labeled data, and the end-to-end optimization of the constructed heterogeneous graph neural network model is performed;

[0032] (2) the optimized model is used to realize the semi-supervised node classification detection task under a small amount of labeled data, so as to detect normal data and abnormal data.

[0033] The further improvement of the present application is that in step four, the method is implemented as follows:

[0034] (1) define the meta-path mask matrix, which is used to obtain a new masked heterogeneous subgraph by multiplying the original heterogeneous graph with the mask matrix;

[0035] (2) randomly initialize the meta-path mask matrix, input the new heterogeneous subgraph into the optimized heterogeneous graph neural network to obtain a new output result, and compare it with the output of the original heterogeneous subgraph;

[0036] (3) define the loss function between the subgraph output and the original graph output, and optimize the mask matrix to minimize the loss;

[0037] (4) multiply the finally optimized mask matrix with the original heterogeneous graph to obtain the masked heterogeneous subgraph, wherein the remaining neighbor nodes are the important factors affecting the node classification result.

[0038] The further improvement of the present application is that it further comprises step five, visualization of abnormal data and its important factors;

[0039] Multiply the optimized mask matrix with the original heterogeneous graph to obtain the final heterogeneous subgraph to be displayed, and send the abnormal company node and its important factors in the form of a graph to the tax inspection personnel for further checking.

[0040] An important factor extraction device for taxpayer enterprise classification, characterized in that it comprises:

[0041] A data preprocessing module for preprocessing the original data based on the data format to obtain a heterogeneous graph data format;

[0042] A heterogeneous neural network construction module for extracting representation vectors of the heterogeneous graph data using a hierarchical attention mechanism heterogeneous graph neural network model to obtain the representation vectors corresponding to each data node;

[0043] A semi-supervised optimization module for semi-supervised optimization of the meta-path mask module of the heterogeneous graph neural network model in the case of lack of labeled data, for extracting important factors affecting the judgment of the data node.

[0044] Compared with the prior art, the present application has at least the following beneficial technical effects:

[0045] 1. The data processing method based on graph data makes the connection between data more close. Compared with traditional data, the extracted important factors affecting enterprise classification have better explainability due to the correlation between graph data, thereby improving the credibility of the present method compared with previous methods.

[0046] 2、The neural network model using multi-level attention mechanism is more excellent in performance compared with the existing message passing model, improves the accuracy of abnormal data detection, and further improves the confidence of the tax enterprise classification task, thereby providing accuracy guarantee for the extraction of subsequent important factors.

[0047] 3、The important factor extraction method based on the heterogeneous graph provided by the application can be used in a more complex and real heterogeneous data scene, and is not limited to a homogeneous data scene, which provides guarantee for the expansion and generalization of the subsequent applicable scene. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0049] Figure 1 A flowchart of an important factor extraction method for abnormal detection based on a heterogeneous graph provided by the embodiments of the present application;

[0050] Figure 2 A structural schematic diagram of an important factor extraction method for abnormal detection based on a heterogeneous graph provided by the embodiments of the present application. DETAILED DESCRIPTION

[0051] The core of the present application is to provide an important factor extraction method, an important factor extraction device, a server and a computer readable storage medium based on a heterogeneous graph for abnormal detection, so as to enhance the credibility of abnormal data detection.

[0052] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] The following will illustrate an important factor extraction method for taxpayer enterprise classification provided by the present application through a specific embodiment.

[0054] Please refer to Figure 1 , Figure 1 A flowchart of an important factor extraction method for taxpayer enterprise classification provided by the embodiments of the present application.

[0055] In this embodiment, the method can include:

[0056] S101 constructing a heterogeneous graph of original data based on a data format

[0057] Invoice information generated in the transaction process and enterprise registration information are the only data sources for studying transactions and paying taxes. The enterprise information is usually sorted and stored in a database. This data item contains all the registration information and transaction information of the enterprise, including the enterprise taxpayer identification number, enterprise registered legal person identification number, transaction commodity number, transaction amount, and other important information. However, there is a lot of redundant information, which is of little help in identifying abnormal enterprises. Moreover, the data item is mainly stored in the form of a data table, and cannot explicitly represent the association between enterprises, so the data needs to be converted to graph data for processing.

[0058] The various types of information of each enterprise obtained from the tax invoices in a certain region are sorted and cleaned. The enterprise and its legal person information, registration information, main business, and transactions with other enterprises are presented in the form of a heterogeneous graph. The node types include: enterprise, legal person, event, and commodity; the edge types include: transaction edge (enterprise→enterprise), information change edge (enterprise→event), sales edge (enterprise→commodity), purchase edge (commodity→enterprise), and legal person edge (enterprise→legal person). At the same time, node features of the same dimension are constructed to facilitate subsequent processing;

[0059] The specific implementation of S101 is described below.

[0060] First, in the database, the enterprise table is queried by SQL (Structured Query Language) statement operation to obtain the information of each enterprise, including the enterprise taxpayer identification number and legal person information. Each enterprise is represented as a node in the graph, with a node category of 0, representing an enterprise node. The node ID starts from 0 and goes up to N1-1, representing a total of N1 enterprises;

[0061] Similarly, the registered legal person of the enterprise is also represented as a node in the graph, with a node category of 1, representing an enterprise legal person. The node ID starts from N1 and goes up to N2-1, representing a total of N2-N1 legal persons;

[0062] The transaction database is queried using the enterprise taxpayer identification number as the key to obtain all the transaction information related to the enterprise. The transaction information includes purchase invoices with the enterprise as the buyer and sales invoices with the enterprise as the seller. The commodity information is extracted and displayed in the form of a node in the graph. The node category is 2, the node ID starts from N2 and goes up to N3-1, representing a total of N3-N2 commodities;

[0063] With the enterprise taxpayer identification number as the key, the enterprise event database is queried to obtain event information about the enterprise. The event information includes information such as the monitoring record, investor equity change, and legal representative change, which has an impact on the operation of the enterprise, and these information is extracted to the graph to be displayed in the form of a node. The node category is 3, the node ID starts from N3 and ends at N4-1, representing a total of N4-N3 events.

[0064] The above node construction process constructs a total of N=N1+N2+N3+N4 nodes. In order to facilitate subsequent feature extraction, a feature vector of the same dimension is constructed for all nodes. At this point, the nodes in the graph data have been constructed. Next, the edge construction process of the graph data is introduced:

[0065] The heterogeneous graph has a total of 5 types of edges, including transaction edges (enterprise→enterprise), information change edges (enterprise→event), sales edges (enterprise→product), purchase edges (product→enterprise), and legal person edges (enterprise→legal person), which are stored in the form of directed edges (src, dst) in the graph, where src is the starting node of the edge and dst is the ending node of the edge. The type of the edge includes five types from 0 to 4, according to the different starting and ending node types, to determine the type of the edge. Transaction edge (0, 0), category 0; sales edge (0, 2), category 1; purchase edge (2, 0), category 2; information change edge (0, 3), category 3; legal person edge (0, 1), category 4. At this point, the nodes and edges of the heterogeneous graph have been constructed, and the construction of the table data into the graph data is complete.

[0066] Using the above method, the transaction information of 44396 enterprises in the region is obtained, including the names and taxpayer identification numbers of 44396 enterprises, 8789 legal person information, 4260 products, and 54567 event information. According to the above method, the heterogeneous graph data is constructed to obtain a total of 112012 nodes, 198903 edges, and a heterogeneous graph with a node feature vector dimension of 300.

[0067] S102 Constructing a Heterogeneous Neural Network Based on Hierarchical Attention Mechanism

[0068] A hierarchical attention heterogeneous neural network model is constructed to learn the embedding representation of each enterprise node from the above-constructed heterogeneous graph for downstream classification tasks. The model has two stages: node-level attention learning and semantic-level attention learning, which are introduced as follows:

[0069] First, the weight of the neighborhood based on the meta-path is learned by node-level attention, and the semantic-specific node embedding is obtained by aggregating it. The meta-path is a semantic path connection defined according to expert knowledge. In this embodiment, the meta-path has the following types: Company-Item-Company (CIC), Company-Person-Company (CPC), and Company-Event-Company (CEC). After defining the meta-path, the representation vector of 44369 company nodes is learned.

[0070] Given a meta-path, node-level attention is used to learn the node representation under a specific semantic. Self-attention mechanism is used to learn the representation vectors between the node and its neighbors. The specific formula is as follows:

[0071]

[0072] where Φ is the given meta-path, h i and h j are the feature vectors of nodes i and j, respectively, is the importance of node j to node i

[0073] Through the softmax layer, the attention weight of node j to node i is obtained:

[0074]

[0075] Finally, the node representation under a specific semantic is learned by the node-level aggregation operation:

[0076]

[0077] Given a meta-path set {Φ0, Φ1,..., Φ P}, after node-level attention, P groups of semantic-specific node embedding representations {Z0, Z1,..., Z P} can be obtained.

[0078] Then, the semantic-level attention is used to distinguish different meta-paths, so as to obtain the optimal weighted combination of semantic-specific node embedding for a specific task.

[0079] The formal description of semantic-level attention is as follows:

[0080]

[0081] where, is the attention weight of each meta-path. The importance of each semantic (meta-path) is learned by a single-layer neural network and a semantic-level attention vector, and is normalized by softmax.

[0082] The final node embedding is calculated as follows

[0083]

[0084] S103 realizes the semi-supervised node classification task by using the existing labels:

[0085] In the tax scenario, the number of labeled abnormal companies is much less than the total number of companies. In the data of this region, the number of company nodes with labels is 1770, and the number of company nodes without labels is 42626. The number of labeled companies is much less than the total number of companies. Therefore, under the premise of lacking sufficient labels, a semi-supervised algorithm of self-training is used to optimize the heterogeneous graph neural network model constructed above. For normal enterprises, the label is defined as 0, and for abnormal enterprises, the label is defined as 1. The model loss function is defined as follows:

[0086]

[0087] where C represents the parameters of the classifier, y represents the index of the labeled node, Y represents the label of the labeled node, and Z represents the embedding representation of the labeled node. After multiple iterations, the optimized heterogeneous graph neural network model H is obtained.

[0088] S104 extracts the important factor of the influence node classification result based on the meta-path:

[0089] The mask matrix based on the meta-path is composed of multi-dimensional adjacency matrices, each dimension representing an adjacency mask matrix, M i×j×k = 1, representing that the edge between the i-th company node and the j-th company node in the k-th meta-path can be retained, and M i×j×k = 0 means that the edge between the i-th company node and the j-th company node in the k-th meta-path needs to be deleted.

[0090] First, the meta-path mask matrix M 112012×112012×3 = {0 or 1} is randomly initialized, the original heterogeneous graph G is multiplied by the mask matrix to obtain the heterogeneous sub-graph G c

[0091] G c = G·M c , c∈{1,…,3}

[0092] Select the enterprise predicted as an abnormal category in S103 as a node that needs to further extract important factors affecting the judgment result, and use a back propagation algorithm to optimize the mask matrix M, wherein the loss function is defined as follows:

[0093]

[0094] Wherein The formula sums the conditional entropy under different meta-paths Φ by Weighted sum of weight coefficients, which ensures the accuracy of the model and takes into account the importance of different meta-paths.

[0095] Finally, the mask matrix M is subjected to fast gradient descent according to the loss function L to optimize M. After reaching the maximum iteration number K or the loss function L is less than the set threshold l0, the final result M and the masked heterogeneous subgraph G k are output. Through experimental analysis, here l0=4.0 and K=10.

[0096] Visualization of abnormal nodes and important factors thereof in S105

[0097] Multiply the mask matrix optimized in S104 by the original heterogeneous graph to obtain the final heterogeneous subgraph to be displayed, and send the abnormal company nodes and their important factors in the form of a graph to the tax inspection personnel for further verification. This method provides a basis for abnormal enterprise identification and provides a reliable solution.

[0098] In summary, the specific embodiment first processes the original data to obtain heterogeneous graph data, and then uses a heterogeneous graph neural network based on an attention mechanism to obtain an embedding representation of each data node. This representation can be used for downstream node classification tasks. Then, a self-training semi-supervised algorithm is used to optimize the heterogeneous graph neural network, improving the accuracy of data anomaly detection. Finally, by optimizing the mask matrix based on the meta-path, important factors affecting data anomaly detection are extracted, improving the credibility of the overall process.

[0099] The important factor extraction device provided in the embodiments of the present application will be described below. The important factor extraction device described below can be correspondingly referred to the important factor extraction method described above.

[0100] Please refer to Figure 2 , Figure 2 for a structural schematic diagram of an important factor extraction device for taxpayer enterprise classification provided in the embodiments of the present application.

[0101] In this embodiment, the device can include:

[0102] The data preprocessing module 100 is configured to preprocess the original data based on the data format to obtain a heterogeneous graph data format.

[0103] The heterogeneous neural network construction module 200 is configured to adopt a hierarchical attention mechanism heterogeneous graph neural network model to extract a representation vector of the heterogeneous graph data, and obtain a representation vector corresponding to each data node.

[0104] The semi-supervised optimization module 300 is configured to semi-supervise the optimization of the heterogeneous graph neural network model in the case of lacking label data.

[0105] The visual display module 500 is configured to visually display the abnormal node and the extracted important factor.

[0106] Optionally, the data preprocessing module 100 is specifically configured to clean the original data to obtain to-be-processed data, and perform a table data to graph data conversion process on the to-be-processed data based on data formats and data attributes.

[0107] The above describes in detail the important factor extraction method for anomaly detection based on a heterogeneous graph provided in the present application. The principles and implementation manners of the present application are described by using specific examples, and the above description of the embodiments is only used to help understand the method and the core idea of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A method for extracting important factors for taxpayer enterprise classification, characterized by, Comprise the following steps: Step one, the data preprocessing module of the original data of the heterogeneous graph processing; The original data includes invoice data generated from the enterprise operation and the data that can be queried on the Internet from the enterprise registration information; Step two, construct a heterogeneous graph neural network model based on attention mechanism; Through hierarchical attention mechanism, learn the embedding representation of the data node under the heterogeneous graph, which is used for subsequent data detection and important factor extraction; The hierarchical attention includes two stages: learning based on node-level attention and learning based on semantic-level attention; The specific implementation method is as follows: (1) Construct a heterogeneous graph neural network based on node-level attention, which converges the neighborhood node features obtained from the above heterogeneous graph construction, learns the neighbor node weight based on the meta path by defining different meta paths, and aggregates the embedding representation of the node with specific semantics; (2) Construct a heterogeneous graph neural network layer based on semantic-level attention, which jointly learns the weight of each meta path and fuses the semantic-specific node embedding representation obtained in the previous step through semantic-level attention; (3) Combine the above different neural network layers to form a heterogeneous graph neural network based on hierarchical attention mechanism; Step three, use semi-supervised classification method to realize the optimization of the heterogeneous graph neural network model; Step four, design a heterogeneous graph interpreter based on meta path perturbation to realize the extraction of important factors affecting abnormal data detection; The specific implementation method is as follows: (1) Define the meta path mask matrix, which is used to obtain a new masked heterogeneous subgraph by multiplying the original heterogeneous graph with the mask matrix; (2) Randomly initialize the meta path mask matrix, input the new heterogeneous subgraph into the above optimized heterogeneous graph neural network, obtain the new output result, and compare it with the output of the original heterogeneous subgraph; (3) Define the loss function between the subgraph output and the original graph output, and optimize the mask matrix to minimize the loss; (4) Multiply the final optimized mask matrix with the original heterogeneous graph to obtain the masked heterogeneous subgraph, wherein the remaining neighbor nodes are the important factors affecting the node classification result.

2. The taxpayer-oriented enterprise classification important factor extraction method according to claim 1, characterized in that, In step one, the specific implementation method is as follows: (1) Data cleaning is performed on the original data to eliminate incomplete data; (2) Based on the data category, define different categories of data as different categories of nodes, and construct different categories of edges according to the association information between data; Define the same dimension feature vector for each node; (3) Construct the heterogeneous graph data according to the defined nodes and edges.

3. The taxpayer business classification oriented important factor extraction method according to claim 1, characterized in that, In step three, the self-training semi-supervised algorithm is used in combination with the heterogeneous graph neural network of step two to realize end-to-end model optimization in the absence of labeled data and realize the data anomaly detection task.

4. The taxpayer-oriented enterprise classification important factor extraction method according to claim 3, characterized in that, In step three, the specific implementation method is as follows: (1) Calculate the loss between the node embedding representation obtained based on the hierarchical attention mechanism and the existing label data, and perform end-to-end optimization on the above constructed heterogeneous graph neural network model; (2) Using the optimized model, realize semi-supervised node classification detection task under small amount of labeled data, so as to detect normal data and abnormal data.

5. The taxpayer business classification oriented important factor extraction method according to claim 1, characterized in that, It also includes step five, visualization of abnormal data and its important factors; Multiply the optimized mask matrix with the original heterogeneous graph to obtain the final heterogeneous subgraph to be displayed, and send the abnormal company node and its important factors to the tax inspection personnel in the form of a graph for further checking.

6. A taxpayer enterprise classification-oriented important factor extraction apparatus characterized by comprising: It includes: A data preprocessing module for preprocessing the original data into a heterogeneous graph data format based on the data format; The original data includes invoice data generated from the enterprise operation process and enterprise registration information that can be queried on the Internet; A heterogeneous neural network construction module for extracting representation vectors of heterogeneous graph data using a hierarchical attention mechanism heterogeneous graph neural network model to obtain the representation vector corresponding to each data node; Through the hierarchical attention mechanism, the embedding representation of the data node under the heterogeneous graph is learned, which is used for subsequent data detection and important factor extraction; The hierarchical attention includes two stages: node-level attention-based learning and semantic-level attention-based learning; The specific implementation method is as follows: (1) Construct a heterogeneous graph neural network based on node-level attention, which converges the neighborhood node features obtained from the above heterogeneous graph construction, learns the neighbor node weight based on the meta path by defining different meta paths, and aggregates them to obtain the embedding representation of the node with specific semantics; (2) Construct a heterogeneous graph neural network layer based on semantic-level attention, which jointly learns the weight of each meta path and fuses the semantic-specific node embedding representation obtained in the previous step through semantic-level attention; (3) Combine the above different neural network layers to form a hierarchical attention mechanism-based heterogeneous graph neural network; A semi-supervised optimization module for semi-supervised optimization of the heterogeneous graph neural network model in the absence of labeled data A meta path mask module for extracting important factors affecting data node judgment; The specific implementation method is as follows: (1) Define a meta path mask matrix, which is used to multiply the original heterogeneous graph with the mask matrix to obtain a new masked heterogeneous subgraph; (2) Randomly initialize the meta path mask matrix, input the new heterogeneous subgraph into the above optimized heterogeneous graph neural network to obtain a new output result, and compare it with the output of the original heterogeneous subgraph; (3) Define a loss function between the subgraph output and the original graph output, and optimize the mask matrix to minimize the loss; (4) Multiply the final optimized mask matrix with the original heterogeneous graph to obtain the masked heterogeneous subgraph, wherein the remaining neighbor nodes are the important factors affecting the node classification result.

Images