A method, system, and related apparatus for detecting tax anomalies based on heterogeneous graphs.

Abstract

The present application provides a tax anomaly detection method based on a heterogeneous graph, including: obtaining tax data of each company; inputting the tax data into a preset graph database, and creating different types of nodes corresponding to the tax data; aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; inputting the meta-path instance level, meta-path intra-level, and meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model; and performing tax anomaly detection using the detection model. The present application utilizes the rich tax risk information of heterogeneous graphs to further improve the accuracy of tax evasion detection. The present application also provides a tax anomaly detection system, a computer-readable storage medium, and an electronic device based on a heterogeneous graph, which have the above-mentioned beneficial effects.

Description

Technical Field

[0001] The present application relates to the field of data detection, and in particular to a tax anomaly detection method, system and related devices based on heterogeneous graphs. Background Art

[0002] Current tax evasion detection methods can be divided into three categories: traditional tax audit methods, machine learning-based methods, and network-based methods. Traditional tax audit methods are divided into whistleblower case selection, manual case selection, and computer case selection. Traditional computer-based case selection mainly uses expert experience to define rules, and then establishes a rule-based system to screen abnormal financial indicators. However, these rules need to be defined manually and have a lag. In addition, as tax evasion behavior changes, manually defined rules often become invalid. Summary of the invention

[0003] The purpose of this application is to provide a tax anomaly detection method, system and related devices based on heterogeneous graphs, which can improve the accuracy and efficiency of tax anomaly detection.

[0004] In order to solve the above technical problems, this application provides a tax anomaly detection method based on heterogeneous graphs. The specific technical solution is as follows:

[0005] Obtain tax data for each company;

[0006] Inputting the tax data into a preset graph database and creating different types of nodes corresponding to the tax data;

[0007] Aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; wherein each node learns a low-dimensional embedding representation, and the embedding representation includes structural information of the heterogeneous graph and semantic information of the tax data;

[0008] Inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model;

[0009] The detection model is used to perform tax anomaly detection.

[0010] Optionally, after obtaining tax data, also include:

[0011] The tax data is cleaned up; the data cleansing includes at least one of data desensitization, data filling and abnormal data deletion.

[0012] Optionally, the meta-path instance-level aggregation of multiple meta-path information of each node includes:

[0013] All nodes on the meta-path instance are concatenated and multiplied with a specific type of matrix to learn the embedding representation of the meta-path instance.

[0014] Optionally, also include:

[0015] The Adam optimizer is used to back-propagate and update the hierarchical model parameters of the detection model.

[0016] Optionally, using the detection model to perform tax detection includes:

[0017] The detection model is used to detect tax evasion and concealed transactions.

[0018] Optionally, using the detection model to perform tax evasion detection and concealed transaction detection includes:

[0019] Outputting the learned node embeddings of each company using the detection model, and inputting the node embeddings into a support vector machine to perform binary classification on all companies and perform tax evasion detection;

[0020] The embeddings between companies are subjected to Hadamard product to obtain embedding identifiers of company pairs, and the embedding pairs are then input into the support vector machine for concealed transaction detection.

[0021] The present application also provides a tax anomaly detection system based on heterogeneous graphs, comprising:

[0022] Data acquisition module, used to obtain tax data of each company;

[0023] A node creation module, used for inputting the tax data into a preset graph database and creating different types of nodes corresponding to the tax data;

[0024] A meta-path aggregation module, for aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; wherein each node learns a low-dimensional embedding representation, and the embedding representation includes structural information of the heterogeneous graph and semantic information of the tax data;

[0025] A model building module, used for inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model;

[0026] The tax detection module is used to perform tax anomaly detection using the detection model.

[0027] Optional tax detection modules include:

[0028] A tax evasion detection unit, configured to output the node embedding learned by each company using the detection model, and input the node embedding into a support vector machine so as to perform binary classification on all companies and perform tax evasion detection;

[0029] The concealed transaction detection unit is used to perform Hadamard product on the embeddings between companies to obtain the embedding identifiers of company pairs, and then input the embedding pairs into the support vector machine for concealed transaction detection.

[0030] The present application also provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the above-mentioned method when executed by a processor.

[0031] The present application also provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of the above-mentioned method when calling the computer program in the memory.

[0032] The present application provides a tax anomaly detection method based on a heterogeneous graph, comprising: obtaining tax data of each company; inputting the tax data into a preset graph database, and creating different types of nodes corresponding to the tax data; aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; wherein each node learns a low-dimensional embedding representation, and the embedding representation includes structural information of the heterogeneous graph and semantic information of the tax data; inputting the meta-path instance level, meta-path intra-level, and meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model; and performing tax anomaly detection using the detection model.

[0033] This application models the tax scenario as a heterogeneous information network, taking into full account the three entities of companies, people and goods, as well as the rich interactive information between entities. By utilizing the rich tax risk information of heterogeneous graphs, the accuracy of tax evasion detection can be further improved. At the same time, semi-supervised learning can be performed on heterogeneous graphs, and the complex structural semantic information in the heterogeneous graphs can be learned through a small amount of labeled data, which is adaptive to tax scenarios. At the same time, the detection model generated by this application can not only detect tax evasion, but also perform a variety of downstream tasks such as hidden transaction detection and abnormal gang detection.

[0034] The present application also provides a tax anomaly detection system based on a heterogeneous graph, a computer-readable storage medium and an electronic device, which have the above-mentioned beneficial effects and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0036] Figure 1 A flowchart of a tax anomaly detection method based on a heterogeneous graph provided in an embodiment of the present application;

[0037] Figure 2 A schematic diagram of a network model of a tax heterogeneous graph provided in an embodiment of the present application;

[0038] Figure 3 A schematic diagram of the structure of a tax anomaly detection system based on a heterogeneous graph provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] Please refer to Figure 1 , Figure 1 A flowchart of a tax anomaly detection method based on a heterogeneous graph provided in an embodiment of the present application, the method comprising:

[0041] S101: Obtain tax data of each company;

[0042] This step aims to obtain the company's tax data. It should be noted that the tax data refers to the relevant data required for the tax application process, such as taxpayer identification number nsrsbh, taxpayer electronic file number nsrdzdah, legal person certificate number zjhm, invoice code fpdm, invoice number fphm and other information.

[0043] Of course, this step does not specifically limit how to obtain tax data. In addition, after obtaining the tax data, the tax data is cleaned, and the data cleaning may include at least one of data desensitization, data filling, and abnormal data deletion. Data desensitization can first recode the taxpayer identification number nsrsbh, taxpayer electronic file number nsrdzdah, legal person certificate number zjhm, invoice code fpdm, invoice number fphm and other information in the data, and reorder 0; randomly delete and add company names, and randomly fill in the newly generated names of investors and legal persons; parse the ID card number, and extract the three useful information of the place of issuance, year of birth and gender through the 15-digit or 18-digit ID card. Data filling requires deletion of data items with serious missing items. For other data items, mean interpolation is used to take the average or mode. Under the condition that the missing type is random missing, maximum likelihood estimation can be used. For data with obvious abnormalities, they can be deleted directly.

[0044] S102: inputting the tax data into a preset graph database, and creating different types of nodes corresponding to the tax data;

[0045] This step requires creating nodes corresponding to the tax data. The preset graph database is not limited here. Preferably, this step can store the data in the neo4j graph database, create various types of nodes and edges, and the nodes and edges contain various attribute information. You can also further use BERT, word2vec, and one hot encoding technology to learn the initial low-dimensional embedding representation of each node.

[0046] S103: aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level;

[0047] S104: inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model;

[0048] When executing this step, each node learns a low-dimensional embedding representation, which contains the structural information of the heterogeneous graph and the semantic information of the tax data. The detection model required by this embodiment mainly performs tax anomaly detection by three parts: meta-path instance-level aggregation, meta-path intra-level aggregation, and meta-path inter-level aggregation. Meta-path is the basic semantic exploration method on heterogeneous graphs. Complex tax risk information is explored through meta-paths. Each node contains multiple meta-path instances. First, meta-path instance-level aggregation is used to aggregate multiple meta-path information of each node. Secondly, the information of different meta-path instances is aggregated using the meta-path intra-level aggregation. Finally, the information of different meta-paths is aggregated using the meta-path inter-level aggregation. The specific construction is as follows:

[0049] Step 1. Meta-path instance-level aggregation

[0050] Meta-path is the basic semantic exploration method in heterogeneous graphs, and each node contains multiple meta-path instances. Each instance contains different information. For example, the company-person-company meta-path expresses the semantics that one person has invested in two companies; the company-goods-company meta-path expresses that the two companies have bought and sold the same goods and have similar business scopes. Different meta-path instances contain very different semantic information, so the key issue of meta-path instance-level aggregation is how to learn the rich information brought by different meta-path instances. This patent learns the embedded representation of the meta-path instance by concatenating all the nodes on the meta-path instance and multiplying them with a specific type of matrix.

[0051] Step 2. Meta-path intra-level aggregation

[0052] Each node contains multiple meta-path instances based on each meta-path, and each meta-path instance contains different semantic information. By introducing intra-level aggregation of meta-paths, the importance of each meta-path instance to the node is different, so different attention coefficients are assigned to each meta-path instance, and finally the embedded representation of each meta-path is learned.

[0053] Step 3. Meta-path level aggregation

[0054] The embedded representation of each meta-path is learned through step 2. Each node contains multiple meta-paths, and each meta-path expresses different meanings. For example, company-person-company expresses the common investor relationship between the two companies, and company-goods-company expresses the common business scope relationship between the two companies, which contains different deep semantic information. In order to better learn the embedded representation of nodes from multiple meta-paths, meta-path level aggregation is used to learn the implicit information of nodes from multiple meta-paths. Finally, each node learns a low-dimensional embedded representation, which fully captures the complex structural information and rich tax risk semantic information in the heterogeneous graph.

[0055] For details, see Figure 2 , Figure 2 The network model of the tax heterogeneous graph is shown. First, three types of nodes (company nodes, person nodes, and goods nodes) and six types of edges (transaction relationships between companies, investment or holding relationships between people and companies, kinship relationships between people, buying relationships between companies and goods, selling relationships between companies and goods, and category relationships between goods) are created in the graph database. The nodes and edges contain a variety of attribute information, and then the desensitized data is stored in the database.

[0056] First, meta-paths are used to explore semantics. This embodiment uses multiple meta-paths containing different semantics, such as company-company, company-person-company, company-person-person-company, company-goods-company, company-goods-goods-company, to explore the complex structure and rich tax risk information in the heterogeneous graph. Then, a hierarchical attention model is used as a detection model to detect tax evasion.

[0057] For meta path instance level aggregation:

[0058] Metapath is the basic semantic exploration method in heterogeneous graphs. Each node contains multiple metapath instances. Each instance contains different risk information. For example, the company-person-company metapath expresses the semantics that one person has invested in two companies; the company-goods-company metapath expresses that two companies have bought and sold the same goods and have similar business scopes. Different metapath instances contain very different risk semantic information, so the key issue of metapath instance-level aggregation is how to learn the rich tax risk information brought by different metapath instances.

[0059] The task of meta-path instance-level aggregation is to learn the feature representation of node meta-path instances. Each node contains one or more instances based on a meta-path, and each instance of a meta-path contains a lot of information. The purpose of meta-path instance-level aggregation is to better mine the deep features of heterogeneous data sources, complex structural information, and rich semantic information in meta-path instances, and finally output meta-path instance-level embedding.

[0060] However, heterogeneous graphs contain many types of node types (such as companies, people, and goods), which contain completely different attribute information, which have different feature spaces and may even have different feature dimensions. Therefore, if used directly, it will seriously affect the effect of tax evasion detection or hidden transaction detection. Therefore, the nodes are first projected into the same feature space using the feature transformation matrix. The node projection process is as follows:

[0061] h i =P·x i ;

[0062] Where P is the tax node projection matrix, x iis the initial tax feature vector of company i, h i is the tax feature vector of company i after projection.

[0063] After the initial feature vector of the node is projected through the transformation matrix, all nodes are in a unified feature space. In the semantic exploration method based on meta-path, researchers often discard all intermediate nodes in the meta-path and transform the complex heterogeneous graph aggregation problem into a simple homogeneous graph problem, which will undoubtedly lead to a large amount of information loss. Therefore, this patent retains the intermediate node information on the meta-path, which is formally expressed as:

[0064]

[0065] in, is the kth instance of company i under meta-path p, which concatenates the nodes on the meta-path. P is the meta-path-specific tax risk semantic learning matrix, and σ(·) is the nonlinear activation function.

[0066] For meta-path level aggregation:

[0067] Each node contains multiple meta-path instances based on each meta-path, and each meta-path instance contains different risk semantic information. By introducing intra-meta-path aggregation, the importance of each meta-path instance to the node is different, so different attention coefficients are assigned to each meta-path instance, and finally the embedded representation of each meta-path is learned.

[0068] After meta-path instance-level aggregation, each node learns a set of meta-path-based instance-level embeddings. This patent constructs meta-path intra-level aggregation to learn a set of instance-level embeddings after meta-path instance-level aggregation. Since the embedding of each meta-path instance contains different risk semantic information. The most commonly used method is to directly take the average of the instance-level embeddings, which is equivalent to treating each meta-path instance with equal weight. In fact, the contribution of each meta-path instance embedding to the node is different. Inspired by the attention mechanism, the meta-path intra-level attention mechanism is introduced to learn the impact of different meta-path instances on the node, which is formally expressed as:

[0069]

[0070] Among them, k P is the parameterized attention vector related to the tax semantics within the meta-path, is the tax risk importance coefficient of the jth instance. Then the normalization operation is performed:

[0071]

[0072] in, is the normalized tax risk attention value.

[0073] The embedding of node i based on the meta-path level can be represented as a set of instance embedding representations learned through meta-path instance-level aggregation and related importance coefficients as follows:

[0074]

[0075] in is the risk embedding representation learned by company i based on multiple meta-path instances of meta-path P.

[0076] For meta-path level aggregation:

[0077] After intra-meta-path level aggregation, each node learns a set of meta-path level embeddings. In heterogeneous graphs, each meta-path contains different deep semantic information. In order to better learn the embedding representation of nodes from multiple meta-paths, a novel inter-meta-path attention mechanism is proposed to learn the implicit information of nodes from multiple meta-paths.

[0078]

[0079] Among them, v P is the meta-path-level tax risk parameterized attention vector, is the tax importance coefficient between meta-paths. The same softmax processing is performed:

[0080]

[0081] in, is the normalized tax risk attention value.

[0082] Finally, the meta-path specific embeddings are aggregated using the learned attention coefficients to obtain the final company risk features as follows:

[0083]

[0084] The label information is a record of tax evasion by a small number of companies. Therefore, this embodiment performs semi-supervised learning in a heterogeneous graph, uses a cross entropy loss function to calculate the loss, and then uses the Adam optimizer to back-propagate to update the three-level model parameters of meta-path instance-level aggregation, meta-path intra-level aggregation, and meta-path inter-level aggregation to obtain a detection model.

[0085] S105: Using the detection model to perform tax anomaly detection.

[0086] This step directly applies the detection model to detect tax anomalies. Specifically, the detection model can be used to detect tax evasion and concealed transactions. Specifically, the detection model can be used to output the node embedding learned by each company, and the node embedding can be input into the support vector machine to perform binary classification on all companies and perform tax evasion detection. For concealed transaction detection, the embeddings between companies can be Hadamard products to obtain the embedding identifiers of company pairs, and then the embedding pairs can be input into the support vector machine for concealed transaction detection.

[0087] The embodiment of the present application models the tax scenario as a heterogeneous information network, taking into full account the three entities of companies, people and goods, as well as the rich interactive information between entities. The rich tax risk information of heterogeneous graphs can further improve the accuracy of tax evasion detection. At the same time, semi-supervised learning can be performed on heterogeneous graphs, and the complex structural semantic information in the heterogeneous graphs can be learned through a small amount of labeled data, which is adaptive to tax scenarios. At the same time, the detection model generated by the present application can not only detect tax evasion, but also perform a variety of downstream tasks such as hidden transaction detection and abnormal gang detection.

[0088] The following is an introduction to a tax anomaly detection system based on a heterogeneous graph provided in an embodiment of the present application. The tax anomaly detection system based on a heterogeneous graph described below and the tax anomaly detection method based on a heterogeneous graph described above can refer to each other.

[0089] See also Figure 3 , Figure 3 The present application also provides a tax anomaly detection system based on a heterogeneous graph, including:

[0090] Data acquisition module, used to obtain tax data of each company;

[0091] A node creation module, used for inputting the tax data into a preset graph database and creating different types of nodes corresponding to the tax data;

[0092] A meta-path aggregation module, for aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; wherein each node learns a low-dimensional embedding representation, and the embedding representation includes structural information of the heterogeneous graph and semantic information of the tax data;

[0093] A model building module, used for inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model;

[0094] The tax detection module is used to perform tax anomaly detection using the detection model.

[0095] Based on the above embodiment, as a preferred embodiment, the tax detection module includes:

[0096] A tax evasion detection unit, configured to output the node embedding learned by each company using the detection model, and input the node embedding into a support vector machine so as to perform binary classification on all companies and perform tax evasion detection;

[0097] The concealed transaction detection unit is used to perform Hadamard product on the embeddings between companies to obtain the embedding identifiers of company pairs, and then input the embedding pairs into the support vector machine for concealed transaction detection.

[0098] The present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed, the steps provided in the above embodiment can be implemented. The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0099] The present application also provides an electronic device, which may include a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps provided in the above embodiment may be implemented. Of course, the electronic device may also include various network interfaces, power supplies and other components.

[0100] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the system provided in the embodiment, since it corresponds to the method provided in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0101] Specific examples are used herein to illustrate the principles and implementation methods of the present application, and the description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

[0102] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A tax anomaly detection method based on heterogeneous graphs, characterized in that: include: Obtain tax data for each company; Inputting the tax data into a preset graph database and creating different types of nodes corresponding to the tax data; Aggregating multiple meta-path information of each node at the meta-path instance level, aggregating information of different meta-path instances at the meta-path intra-level, and learning implicit information of the node from different meta-path instances at the meta-path inter-level; wherein each node learns a low-dimensional embedding representation, and the embedding representation includes structural information of the heterogeneous graph and semantic information of the tax data; Inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model; Using the detection model to perform tax anomaly detection; Aggregating multiple meta-path information of each node at the meta-path instance level includes: Concatenate all nodes on the meta-path instance and multiply them with a specific type of matrix to learn the embedding representation of the meta-path instance; Wherein, using the detection model to perform tax detection includes: Outputting the learned node embeddings of each company using the detection model, and inputting the node embeddings into a support vector machine to perform binary classification on all companies and perform tax evasion detection; The embeddings between companies are subjected to Hadamard product to obtain embedding identifiers of company pairs, and the embedding pairs are then input into the support vector machine for concealed transaction detection.

2. The tax anomaly detection method according to claim 1, characterized in that: After obtaining tax data, it also includes: The tax data is cleaned up; the data cleansing includes at least one of data desensitization, data filling and abnormal data deletion.

3. The tax anomaly detection method according to claim 1, characterized in that: Also includes: The Adam optimizer is used to back-propagate and update the hierarchical model parameters of the detection model.

4. A tax anomaly detection system based on heterogeneous graphs, characterized in that: include: Data acquisition module, used to obtain tax data of each company; A node creation module, used for inputting the tax data into a preset graph database and creating different types of nodes corresponding to the tax data; A meta-path aggregation module is used to concatenate all nodes on the meta-path instance, multiply them with a matrix of a specific type to learn the embedded representation of the meta-path instance, aggregate the information of different meta-path instances using the intra-meta-path level, and learn the implicit information of the nodes from different meta-path instances using the inter-meta-path level; wherein each node learns a low-dimensional embedded representation, and the embedded representation contains the structural information of the heterogeneous graph and the semantic information of the tax data; A model building module, used for inputting the meta-path instance level, the meta-path intra-level and the meta-path inter-level into the heterogeneous graph, calculating the loss using a cross entropy loss function, and obtaining a detection model; A tax detection module, used for performing tax anomaly detection using the detection model; The tax detection module includes: A tax evasion detection unit, configured to output the node embedding learned by each company using the detection model, and input the node embedding into a support vector machine so as to perform binary classification on all companies and perform tax evasion detection; The concealed transaction detection unit is used to perform Hadamard product on the embeddings between companies to obtain the embedding identifiers of company pairs, and then input the embedding pairs into the support vector machine for concealed transaction detection.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of tax anomaly detection based on heterogeneous graphs as described in any one of claims 1 to 3 are implemented.

6. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, the steps of tax anomaly detection based on heterogeneous graphs as described in any one of claims 1 to 3 are implemented.

Images