Corrugated set-based enterprise taxpayer associated tax risk assessment method

Abstract

The invention discloses an enterprise taxpayer associated tax risk assessment method based on a ripple set. An associated tax risk assessment process is divided into the following steps: 1) constructing a tax risk knowledge graph database; 2) selecting seed nodes, and generating candidate sub-graphs; and 3) identifying the high tax risk subgraph. According to the invention, by introducing the multi-relation fusion knowledge graph, the capability of identifying tax risks hidden in enterprise taxpayer groups is improved; efficient sub-graph analysis is realized on a large-scale graph through a ripple set and a threshold mechanism; the influence of static rule dependence and subjective experience on associated tax risk assessment is reduced through a classic subgraph mode, and the method is superior to an existing method in accuracy, expandability and calculation efficiency, and has very high application value and popularization value.

Description

Technical Field

[0001] This invention relates to the field of corporate taxpayer tax risk management, specifically to a method for assessing related tax risks of corporate taxpayers based on ripple sets, which belongs to the intersection of artificial intelligence and tax risk assessment and management. Background Technology

[0002] With the development of the digital economy and the increasing complexity of business models, there is a growing trend of covert tax planning and even tax evasion using methods such as related-party transactions, cross-regional operations, and shell companies. Traditional tax risk assessment methods typically rely on static financial data, making it difficult to effectively identify and quantify these cross-entity, dynamically evolving tax risks. Therefore, the need for more advanced methods to improve the accuracy and efficiency of tax risk assessment has become a challenge for researchers in this field.

[0003] Researchers in industry and academia are applying graph computing and artificial intelligence technologies to associated risk analysis. Ji Peng et al. proposed a "method, system, equipment, and storage medium for public opinion early warning and risk propagation analysis," constructing a knowledge graph containing basic enterprise information. This method focuses on analyzing unstructured public opinion texts from outside the enterprise, automatically mining nodes associated with high-risk nodes, but neglects internal operational data. Zhang Ze et al. proposed "a method for assessing associated risks of entities," focusing on equity relationships. This method combines basic entity information, business transactions, financial status, and graph theory indicators such as degree and betweenness centrality to analyze enterprises and identify key entities with risks, ignoring the impact of transactions, collaborations, and other related relationships on associated risk assessment. Xu Ming et al. proposed "an artificial intelligence-based method and system for individual stock risk early warning," employing a multi-level attenuation mechanism and multiple adjustment factors to calculate individual stock risk based on the risk transmission characteristics of the stock market. However, it is insufficient in identifying cross-entity transmission tax avoidance patterns. Summary of the Invention

[0004] This invention discloses a method for assessing the related tax risks of corporate taxpayers based on ripple set theory. It constructs a corporate tax risk knowledge graph, defines eight entity relationship types and a relationship strength matrix, and uses the ripple set algorithm to generate candidate subgraphs. The method then identifies high-tax-risk subgraphs using their tax and structural features. Specifically, it includes: 1) constructing a tax risk knowledge graph database; 2) selecting seed nodes and generating candidate subgraphs; and 3) identifying high-tax-risk subgraphs. In detail, the method of this invention includes the following steps:

[0005] A. Construct a tax risk knowledge graph database, with the following specific steps:

[0006] A1. Define a tax risk knowledge graph G={E, R}, where E and R represent entity and relation sets, respectively. E contains two entity types: natural persons and corporate taxpayers, and R contains eight relation types from Table 1.

[0007] A1.1 Define corporate taxpayers and natural persons as entity types in the tax risk knowledge graph G;

[0008] A1.2 defines eight types of relationships between corporate taxpayers, between corporate taxpayers and natural persons, and between natural persons, as shown in Table 1:

[0009] Table 1 Relationship Types

[0010]

[0011] Where CC represents the relationship between corporate taxpayer entities, CO represents the relationship between natural person entities and corporate taxpayer entities, OO represents the relationship between natural person entities, A and B represent corporate taxpayers, and C and D represent natural persons.

[0012] A1.3 Define the attributes of the relationship type between entities, as shown in Table 2:

[0013] Table 2 Relationship Type Attributes

[0014]

[0015] A2. Extract relevant data on investment information, shareholding information, tax payment information, and transaction information from the tax management system to construct relationships between corporate taxpayers, between corporate taxpayers and natural persons, and between natural persons, denoted as relationships (subject, object);

[0016] A3. Extract the attribute data corresponding to the relationship from the tax management system, and use it as the attribute value of the relationship, denoted as relationship (subject, object).attribute;

[0017] B. Select seed nodes and generate candidate subgraphs G. s The specific steps are as follows:

[0018] B1. The main factors for selecting the seed node u include the following characteristics:

[0019] - Transactions and related corporate taxpayers with significant tax rate differences;

[0020] - Taxpayers whose related-party transactions constitute a high percentage of their total transactions;

[0021] - Taxpayers whose businesses frequently engage in large-scale financial transactions;

[0022] B2. Calculate the weight coefficient of each attribute for the four types of relationships. Taking the transaction relationship as an example, calculate the weight coefficient ω for the number of transactions, the total transaction amount, and the total transaction value. j (j=1,2), the specific steps are as follows:

[0023] The four types of relationships mentioned include: transaction relationship, payment on behalf relationship, equity relationship, and investment relationship;

[0024] B2.1 Let C be the set of all instances of transaction relationships CC_transaction(A, B) in the knowledge graph database. i = {c i | c i =(e i h , e i t )}(i=1,2,...,N ), where N is the number of instances, e i h Representing relation c i The head entity, e i t Represents the tail entity; the two attribute values ​​of N transaction relationships constitute an N×2 original relation matrix Z=[z ij (i=1,2,...,N; j=1,2), where z ij This represents the value of the j-th attribute in the i-th relation;

[0025] B2.2 Standardize the original relation matrix Z. All attributes of each relation type are positive indices. Standardize using formula (1) to obtain the standardized matrix Y = [y ij ];

[0026] (1)

[0027] Wherein, min(z) j ) and max(z j ) are the maximum and minimum values ​​of the j-th attribute, respectively;

[0028] B2.3 uses formula (2) to calculate c for each relation instance. i The weight p of the j-th attribute in the attribute set j ij The proportion matrix P = [p] is obtained. ij ];

[0029] (2)

[0030] B2.4 The information entropy e of the j-th attribute of the relation is calculated using formula (3). j ;

[0031] (3)

[0032] If p ij =0, then define p ij ln(p ij )=0;

[0033] B2.5 Information Entropy e j The larger the value, the greater the difference between different relation instances and the stronger the distinguishing ability. Therefore, this attribute should be given a higher weight. The weight coefficient ω of the j-th attribute in the relation is calculated using formula (4). j ;

[0034] (4)

[0035] B3. Use a graph search method to propagate the seed node u layer by layer to generate the ripple set S of u. The specific steps are as follows:

[0036] B3.1 Define the set of associated nodes at level 0 of the seed node u as follows: ;

[0037] B3.2 Construct the first-level candidate associated node set of seed node u The specific steps are as follows:

[0038] B3.2.1 Define the first-level associated node set as , ;

[0039] B3.2.2 Calculate the original relationship strength between a single node in the set and node u, taking a single node u as an example. 1 For example, the specific steps are as follows:

[0040] B3.2.2.1 Remember u and u 1 The number of relations between them is a, where a ≤ 4;

[0041] B3.2.2.2 Calculate u and u 1 The strength of the relationship between each pair; taking a transactional relationship as an example, u and u 1 The standardized relational attribute values ​​between them are y 11 y 12 The weight coefficients ω of the two attributes calculated in step B2 j (j=1,2), use formula (5) to calculate u and u 1 The strength of the transactional relationship between them;

[0042] (5)

[0043] B3.2.2.3 Based on statistical analysis of historical audit cases and expert evaluation, and combined with Saaty's 1-9 scale method, the importance weights of different types of relationships in tax risk transmission are set as β = [β1, β2, β3, β4]. These weights reflect the relative importance of each type of relationship in risk transmission.

[0044] B3.2.2.4 Using the calculation method in B3.2.2.2, we obtain u and u 1 The strength of each relationship between u and u is calculated based on the weights β of the four relationships. 1 Strength of the original relationship between ;

[0045] B3.2.3 Following the method in step B3.2.2, calculate the set of associated nodes at level 1. The strength of the original relationship W between each node and u u ui , i=1,2,...,L;

[0046] B3.2.4 Standardize the original relationship strength (W) between L associated nodes and u. u u1 W u u2 ,…W u uL The original relation strength is standardized using formula (6);

[0047] (6)

[0048] B3.2.5 will The standard relation strength between each node and the seed node u is represented by a triple. or , where r i The values ​​represent node u and node u i The strength of the standardized relationship w between them u ui ;

[0049] B3.2.6 Screening for relationships that meet standardized strength w u ui The associated nodes corresponding to ≥d, where d is a configurable parameter, form the first-level candidate associated node set by selecting nodes that meet the criteria. , (L1≤L);

[0050] B3.3 Construct the first-level ripple subset of seed node u Let L1 be the number of triples in the set, and the set of triples is represented as follows:

[0051]

[0052] Among them, h, h i Represents the head node, r i The value represents the relationship between node u and node u. i The strength of the standardized relationship w between them u ui ,t、t i represents the tail node, i represents the node number, and G represents the tax risk assessment knowledge graph;

[0053] B3.4 For each node u in the existing outermost ripple subset of u i (i=1,2,...,L1) are used as temporary seed nodes. Following steps B3.1, B3.2, and B3.3, the first-level ripple subset of each temporary seed node is constructed. ;

[0054] B3.5 Subsets of the first-level ripples from these temporary seed nodes Combined, they form a new ripple subset S of u. 2 , means as follows:

[0055]

[0056] The number of triples in the set is denoted as L2;

[0057] B3.6 Following the methods in steps B3.4 and B3.5, iteratively expand to generate the ripple subset S of the subsequent levels of the seed node u. k (k=3,4,..,H), the specific steps are as follows;

[0058] B3.6.1 Following the method in step B3.4, extract the existing outermost ripple subset S. k-1 The tail node of each triple in the array serves as a temporary seed node u. i Construct the first-level ripple subset for each temporary node. ;

[0059] B3.6.2 Following the method in step B3.5, extract the first-level ripple subset of all temporary nodes. Merge, as the outermost ripple subset S to be generated for u k ;

[0060] B3.7 Merge the k-layer ripple subsets into a ripple set for seed node u. , H represents the ripple subset of the k-th layer, where H is the largest ripple layer.

[0061] B4. Merge and recombine the triples (h, r, t) of each ripple subset in the ripple set S to generate candidate subgraphs G for the seed nodes. s The specific steps are as follows:

[0062] B4.1 Extract all head nodes h and tail nodes t to form candidate subgraph G. s The set of nodes V s ;

[0063] B4.2 Extract the edge r of the triple (h, r, t) as a candidate subgraph G s The set of edges E s The standardized relation strength value w of r u ui , as the weight of the edge;

[0064] C. Identify high-risk tax avoidance sub-maps, the specific steps are as follows:

[0065] C1. Construct a database of high-risk tax avoidance models. The specific steps are as follows:

[0066] C1.1 Based on historical audit cases and tax expert knowledge, define a set of typical high-risk tax avoidance patterns P={p1,p2,...,p r}, where r is the total number of tax avoidance schemes, and the specific steps are as follows:

[0067] C1.1.1 Each of the aforementioned tax avoidance schemes p j (1≤j≤r) is represented by a subgraph with specific node roles, topological structure, and relationships;

[0068] C1.1.2 Example P1: A tax avoidance pattern exists where "five nodes, P01, P02, C01, C02, C03, C04, and C05, constitute a tax avoidance pattern for related-party transactions." The tax avoidance pattern diagram is shown below. Figure 3 As shown, where:

[0069] • P01 and P02 are natural persons with social relationships, and C01, C02, C03, C04, and C05 are enterprise nodes, and there exists CO_owner(P01, C01), CC_shareholding(C03, C05), and CO_owner(P02, C02).

[0070] • There exists a CC_transaction(C05, C04) with significant differences in the tax rates involved;

[0071] • The risk path meets the requirements of multi-layer penetration and hierarchical depth, and E meets the requirements of tail nodes of multiple risk paths;

[0072] C1.2 represents p for each tax avoidance modelj (1≤j≤r) Extract node attributes and topological features to generate the corresponding pattern feature vector Fp. j Taking example p1, the feature vector Fp1 contains the following core dimensions:

[0073] C1.2.1 Specific types of corporate taxpayer relationships that affect related-party tax risks;

[0074] C1.2.2 Threshold for Tax Rate Difference Between Nodes at Both Ends of a Specific Relationship Type , among which, T C05 T C04 This represents the tax rate for the head node C05 and the tail node C04;

[0075] C1.2.3 Threshold for the percentage of related transaction amount between the two nodes of a specific relationship type Re C05 Represents the total revenue of node C05, Re CC α represents the total revenue generated from transactions between nodes C04 and C05. If there are no transactions between the nodes, α is 0.

[0076] Finally obtained Where type represents the relation type;

[0077] C2. Starting from the enterprise taxpayer node, identify the candidate risk subgraph G. ce The specific steps are as follows:

[0078] C2.1 In candidate subgraph G s Find two corporate taxpayer nodes, a and b, that have a transaction relationship.

[0079] C2.2 Starting from nodes a and b, perform ripple propagation and reverse tracing in the candidate subgraph to find the actual controller nodes c and d corresponding to nodes a and b, respectively. Extract the relationship type, tax rate difference, and transaction amount ratio of nodes a and b to obtain the candidate subgraph G. s eigenvector F Gce ;

[0080] C2.3 Detect whether natural person nodes c and d have social relationships; if they do, and they conform to pattern p j If the threshold condition is met, it is marked as a candidate risk subgraph G. ce ;

[0081] C3. Starting from the natural person node, identify the candidate risk subgraph G. cr The specific steps are as follows:

[0082] C3.1 In candidate subgraph G s Search for natural person nodes a and b that have social relationships;

[0083] C3.2 Starting from nodes a and b, perform ripple propagation direction tracing in the candidate subgraph to find the enterprise taxpayer nodes c and d controlled by nodes a and b respectively. Extract the node relationship type, tax rate difference, and transaction amount ratio to obtain the candidate subgraph G. s eigenvector F Gcr ;

[0084] C3.3 Detect whether there are transaction relationships between enterprise taxpayer nodes c and d that meet the threshold conditions; if so, and they conform to pattern p j If the threshold condition is met, it is marked as a candidate risk subgraph G. cr ;

[0085] C4. Calculate tax avoidance patterns and candidate risk subgraphs G ce G cr Based on feature similarity, high-tax-risk subgraphs are identified through the following steps:

[0086] C4.1 Calculate the risk subgraph G for each candidate. ce G cr And each tax avoidance model p j The feature similarity (1≤j≤r) is calculated using the following steps:

[0087] C4.1.1 Extract candidate risk subgraph G using the method in step C1.2. ce G cr eigenvector F Gce F Gcr ;

[0088] C4.1.2 Calculate the tax avoidance model p using formulas (7) and (8) respectively. j and candidate subgraph G ce G cr Feature similarity of (j=1,2,…,r);

[0089] (7)

[0090] (8)

[0091] C4.2 If sim(p) j G ce )≥θs or sim(p j G cr )≥θ s Then determine G ce or G cr For the high-tax-risk subgraph; where θ s These are configurable parameters.

[0092] Compared with existing technologies, this invention has the following advantages: by introducing a knowledge graph with multi-relation fusion, this invention improves the ability to identify tax risks hidden in corporate taxpayer groups; by using ripple sets and threshold mechanisms, it achieves efficient subgraph analysis on large-scale graphs; and by using classic subgraph patterns, it reduces the impact of static rule dependence and subjective experience on the assessment of related tax risks. It is superior to existing methods in terms of accuracy, scalability, and computational efficiency, and has high application and promotion value. Attached Figure Description

[0093] Figure 1 Flowchart of a method for assessing related-party tax risks of corporate taxpayers based on ripple sets

[0094] Figure 2 : Schematic diagram of ripple set

[0095] Figure 3 Example of a tax avoidance model

[0096] Figure 4 Candidate subgraph Detailed Implementation

[0097] This invention discloses a method for assessing the related tax risks of corporate taxpayers based on ripple sets. The method includes: 1) constructing a tax risk knowledge graph database; 2) selecting seed nodes and generating candidate subgraphs; 3) identifying high-tax-risk subgraphs. The following description, in conjunction with the accompanying drawings, uses 300 corporate taxpayers and 350 individuals, with 1400 relationship data points, to clearly and completely describe the technical solution of this invention. The parameters in the method are examples and can be adjusted based on expert experience and industry background in actual use cases.

[0098] 1. Construct a tax risk knowledge graph database. The specific steps are as follows:

[0099] 1.1 Extract relevant data on investment, shareholding, tax payment on behalf, and transaction information from the tax management system to construct relationships between corporate taxpayers, between corporate taxpayers and natural persons, and between natural persons, denoted as relationships (subject, object); some relationship data are shown in Table 3:

[0100] Table 3. Partial Relationship Data Between Entities

[0101]

[0102] 1.2 Extract the attribute data corresponding to the relationship from the tax management system and use it as the attribute value of the relationship, denoted as relationship (subject, object).attribute; the attribute data of some transaction relationships are shown in Table 4:

[0103] Table 4. Attribute data of some transaction relationships

[0104]

[0105] 2. Select a seed node and generate a candidate subgraph G. The specific steps are as follows:

[0106] 2.1 Select a seed node based on the features in step B1, and select C2 as the seed node;

[0107] 2.2 Calculate the weight coefficient of each attribute for transaction relationship and equity relationship. Taking transaction relationship as an example, the weight coefficients of the transaction relationship attribute for the number of transactions and the transaction amount in the past year are ω. j (j=1,2), the specific steps are as follows:

[0108] 2.2.1 The two attribute values ​​of 50 transaction relationships constitute a 500×2 original relationship matrix Z=[z ij (i=1,2,...,50; j=1,2), the matrix Z is shown below:

[0109]

[0110] 2.2.2 The original relation matrix Z is standardized. All attributes of each relation type are positive indices. Standardization is performed using formula (1) to obtain the standardized matrix Y = [y ij The matrix Y is shown below;

[0111]

[0112] 2.2.3 Calculate c for each relation instance using formula (2). i The weight p of the j-th attribute in the attribute set j ij The proportion matrix P = [pij] is obtained;

[0113]

[0114] 2.2.4 The information entropy e of the j-th attribute of the relation is calculated using formula (3). j The information entropy e1 of the number of transactions in the past year is 0.929664, and the information entropy e2 of the transaction amount in the past year is 0.956814.

[0115] 2.2.5 The weight coefficient ω of the j-th attribute in the relation is calculated using formula (4). j The weighting coefficient ω1 for the number of transactions in the past year is 0.49, and the weighting coefficient ω2 for the transaction amount in the past year is 0.51.

[0116] 2.2.6 Since there is only one attribute to the equity relationship, the weighting coefficient for the percentage of shares held in the other party's taxpayer enterprise is 1.0;

[0117] 2.3 The seed node C1 is propagated layer by layer using a graph search method to generate the ripple set S of C1. The specific steps are as follows:

[0118] 2.3.1 The set of associated nodes at level 0 of seed node C1 is: Level 1 associated node set ;

[0119] 2.3.2 Constructing the first-level candidate associated node set of seed node C1 The specific steps are as follows:

[0120] 2.3.2.1 Calculate the original relationship strength between a single node in the set and node C1. Taking a single node C2 as an example, the specific steps are as follows:

[0121] 2.3.2.1.1 There is a CC_transaction(C1, C2) between node C1 and node C2.

[0122] 2.3.2.1.2 Calculate the strength of CC_transaction(C1, C2) based on the weight coefficients ω of the two attributes calculated in step 2.1.5. j (j=1,2), the relationship strength of CC_transaction(C1,C2) is calculated using formula (5), and the relationship strength of CC_transaction(C1,C2) is 0.0685;

[0123] 2.3.2.1.3 Based on statistical analysis of historical audit cases and expert evaluation, and combined with Saaty's 1-9 scale method, the importance weights β=[0.4, 0.2, 0.3, 0.1] of different types of relationships in tax risk transmission are set. These weights reflect the relative importance of each type of relationship in risk transmission.

[0124] 2.3.2.1.4 Using the calculation method in step B3.2.2.2, obtain the relationship strength of each relationship between C1 and C2, and obtain the original relationship strength between C1 and C2. ;

[0125] 2.3.2.1.5 Following the method in step B3.2.2.2, calculate the set of associated nodes at level 1. The original relation strength between each node and node C1, and the original relation strength of CC_transaction(C1, C3). The original relation strength of CC_transaction(C1, C6) ,;

[0126] 2.3.2.1.6 Standardize the original relation strength (W) between two related nodes C2, C3, and C6 and node C1. C1 C2 W C1 C3 W C1 C6 The original relation strengths were standardized using formula (6), resulting in standardized relation strengths of 0.7685, 0.1955, and 0.0359, respectively.

[0127] 2.3.2.1.7 will The standard relation strength between each node and the seed node C1 is represented by the triples (C1, 0.7685, C2), (C1, 0.1955, C3), and (C1, 0.0359, C6).

[0128] 2.3.2.1.8 Screening for relationships that meet the standardized strength w u ui For nodes corresponding to ≥0.11458, nodes that meet the criteria will form a first-level candidate set of related nodes. ;

[0129] 2.3.3 Constructing the first-level ripple subset of seed node C2 , means as follows:

[0130] S 1 ={(C1, 0.7685, C2), (C1, 0.1955, C3), (P01, 1.0, C1)}

[0131] 2.3.4 Using nodes C2 and C3 as temporary seed nodes, construct the first-layer ripple set for each temporary node according to the methods in 2.3.1, 2.3.2, and 2.3.3. , ;

[0132] 2.3.5 Subset the ripple and Merge, as a new layer of ripple subset S of node C1 2 , means as follows:

[0133] S 2 ={(C5,1.0,C2),(C3,0.3992,C4),(C3,0.6007,C5), (P01, 1.0, P02), (P01,1.0, P03)}

[0134] 2.3.6 Following the methods in steps 2.3.4 and 2.3.5, iteratively expand to generate the subsequent hierarchical ripple subset S of the seed node C1. k (k=3,4,..,5), which are S... 3 ={(P02, 1.0, C5), (P03,1.0,C4),(P01,1.0,C2)}, , ;

[0135] 2.3.7 Combine the 5-layer ripple subsets into a ripple set S={(C1, 0.7685, C2), (C1,0.1955, C3), (P01, 1.0, C1), (C5,1.0,C2), (C3,0.3992,C4), (C3,0.6007,C5),(P01, 1.0, P02), (P01, 1.0, P03), (P02, 1.0, C5), (P03,1.0,C4), (P01,1.0,C2)};

[0136] 2.4 Merge and recombine the triples (h, r, t) of each ripple subset in the ripple set S to generate candidate subgraphs G for the seed nodes. s The specific steps are as follows:

[0137] 2.4.1 Extract all head nodes h and tail nodes t to form a candidate subgraph G. s The set of nodes V s ={C1,C2,C3,C4,C5,P01,P02,P03};

[0138] 2.4.2 Extract all edges r between nodes in the knowledge graph as candidate subgraphs G s The set of edges E s ={(C1, C2),(C1, C3), (P01, C1), (C5, C2), (C3, C4), (C3, C5), (P01, P02), (P01, P03),(P02, C5), (P03, C4), (P01, C2)}, where w is the standardized relation strength value of r. u ui As the weight of the edge; candidate subgraph G s like Figure 4 As shown;

[0139] 3. Identify high-risk tax avoidance sub-maps. The specific steps are as follows:

[0140] 3.1 Based on historical audit cases and tax expert knowledge, a set of typical high-risk tax avoidance patterns, P={p1,p2,p3}, is defined. The specific steps are as follows:

[0141] 3.1.1 Extract p for each tax avoidance pattern j (1≤j≤r) is represented by a subgraph with specific node roles, topological structure, and relationships. Taking p1 as an example, the feature vector Fp1 contains the following core dimensions:

[0142] 3.1.1.1 The specific type of corporate taxpayer relationship affecting related-party tax risks is a transactional relationship;

[0143] 3.1.1.2 The threshold for the difference in tax rates between the two ends of a transaction is 10%;

[0144] 3.1.1.3 The threshold for the proportion of transaction amount between the two ends of the transaction relationship is 25%;

[0145] Finally, the eigenvector F of p1 is obtained. p1 =[1, 10, 25];

[0146] 3.1.2 Repeat the method in 3.1.1 to obtain the eigenvector F of p2. p2 =[1,12, 50] and the eigenvector F of p3 p2 =[1, 16, 50];

[0147] 3.2 Identifying candidate risk subgraphs G starting from the enterprise taxpayer node ce The specific steps are as follows:

[0148] 3.2.1 In candidate subgraph G s Find the two corporate taxpayer nodes C1 and C2 that have a transaction relationship;

[0149] 3.2.2 Starting from nodes C1 and C2, perform ripple propagation and reverse tracing in the candidate subgraph to find the natural person nodes P01 and P03 corresponding to nodes C1 and C2, respectively. Extract the relationship type, tax rate difference, and transaction amount ratio of nodes C1 and C2 to obtain the candidate subgraph G. s eigenvector F Gce =[1, 10, 35.814];

[0150] 3.2.3 Detecting the existence of social relationships between natural person nodes P03 and P02, and the feature vector F Gce If the threshold condition of pattern p1 is met, then label G. s For candidate risk subgraph G ce ;

[0151] 3.3 Calculating Tax Avoidance Patterns and Candidate Risk Subgraph G ce G cr Based on the feature similarity of [the graph] and [the graph], we can identify subgraphs with high tax risk. The specific steps are as follows:

[0152] 3.3.1 Candidate Subgraph G ce eigenvector F Gce =[1, 10, 35.814];

[0153] 3.3.2 Calculate the tax avoidance pattern p1 and candidate subgraph G using formula (7). ce Feature similarity is used to obtain the similarity sim(p1, G). ce The value is 0.9941;

[0154] 3.4.3 sim(p1, G ce If )≥0.5, then G is determined. ce This is a sub-graph representing high tax risk.

[0155] This invention provides a method for assessing related tax risks among corporate taxpayers. It selects nodes with anomalous characteristics as seed nodes, discovers candidate subgraphs with strong relationships through ripple set analysis, and identifies high-risk subgraphs using pattern matching. This invention integrates the advantages of knowledge graphs and ripple sets, and through a construction-filtering-pattern matching process, it discovers hidden high-risk related tax relationships within corporate taxpayer groups, demonstrating broad application and promotion value.

[0156] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A method for assessing related-party tax risks of corporate taxpayers based on ripple sets, comprising the following steps: A. Construct a tax risk knowledge graph database, with the following specific steps: A1. Define a tax risk knowledge graph G={E, R}, where E and R represent entity and relation sets, respectively. E contains two entity types: natural persons and corporate taxpayers, and R contains eight relation types from Table 1. A2. Extract relevant data on investment information, shareholding information, tax payment information, and transaction information from the tax management system to construct relationships between corporate taxpayers, between corporate taxpayers and natural persons, and between natural persons, denoted as relationships (subject, object); A3. Extract the attribute data corresponding to the relationship from the tax management system, and use it as the attribute value of the relationship, denoted as relationship (subject, object).attribute; B. Select seed nodes to generate candidate subgraphs G. The specific steps are as follows: B1. The main factors for selecting the seed node u include the following characteristics: Transactions involving significant tax rate differences and related corporate taxpayers; Taxpayers whose related-party transactions constitute a high percentage of their total transactions; Taxpayers whose businesses frequently engage in large-scale financial transactions; B2. Calculate the weight coefficient of each attribute for the four types of relationships. Taking the transaction relationship as an example, calculate the weight coefficient ω of the two attributes: the number of transactions and the transaction amount. j (j=1,2); B3. Use a graph search method to propagate the seed node u layer by layer to generate the ripple set S of u. The specific steps are as follows: B3.1 Define the set of associated nodes at level 0 of the seed node u as follows: ; B3.2 Construct the first-level candidate associated node set of seed node u The specific steps are as follows: B3.2.1 Define the first-level associated node set as , ; B3.2.2 Calculate the original relationship strength between a single node in the set and node u, taking a single node u as an example. 1 For example, the specific steps are as follows: B3.2.2.1 Remember u and u 1 The number of relations between them is a, where a ≤ 4; B3.2.2.2 Calculate u and u 1 The strength of the relationship between each pair; taking a transactional relationship as an example, u and u 1 The standardized relational attribute values ​​between them are y 11 y 12 The weight coefficients ω of the two attributes calculated in step B2 j (j=1,2), use formula (5) to calculate u and u 1 The strength of the transactional relationship between them; B3.2.2.3 Based on statistical analysis of historical audit cases and expert evaluation, and combined with Saaty's 1-9 scale method, the importance weights of different types of relationships in tax risk transmission are set as β = [β1, β2, β3, β4]. These weights reflect the relative importance of each type of relationship in risk transmission. B3.2.2.4 Using the calculation method in B3.2.2.2, we obtain u and u 1 The strength of each relationship between u and u is calculated based on the weights β of the four relationships. 1 Strength of the original relationship between ; B3.2.3 Following the method in step B3.2.2, calculate the set of associated nodes at level 1. The strength of the original relationship W between each node and u u ui , i=1,2,...,L; B3.2.4 Standardize the original relationship strength (W) between L associated nodes and u. u u1 W u u2 ,…W u uL The original relation strength is standardized using formula (6); B3.2.5 will The standard relation strength between each node and the seed node u is represented by a triple. or , where r i The values ​​represent node u and node u i The strength of the standardized relationship w between them u ui ; B3.2.6 Screening for relationships that meet standardized strength w u ui The associated nodes corresponding to ≥d, where d is a configurable parameter, form the first-level candidate associated node set by selecting nodes that meet the criteria. , (L1≤L); B3.3 Constructing the seed node u to obtain the first-level ripple subset The number of triples in a set is denoted by , and the set of L1 triples is represented as follows: in, h、h i Indicates the head node, r i The value represents the relationship between node u and node u. i The strength of the standardized relationship w between them u ui ,t、 t i Indicates the tail node. i Indicates the node's sequence number. G This represents a knowledge graph for tax risk assessment. B3.4 For each node u in the existing outermost ripple subset of u i (i=1,2,...,L1) are used as temporary seed nodes. Following steps B3.1, B3.2, and B3.3, the first-level ripple subset of each temporary seed node is constructed. ; B3.5 Subsets of the first-level ripples from these temporary seed nodes Combined, they form a new ripple subset S of u. 2 , means as follows: The number of triples in the set is denoted as L2; B3.6 Following the methods in steps B3.4 and B3.5, iteratively expand to generate the ripple subset S of the subsequent levels of the seed node u. k (k=3,4,..,H), the specific steps are as follows; B3.6.1 Following the method in step B3.4, extract the existing outermost ripple subset S. k-1 The tail node of each triple in the array serves as a temporary seed node u. i Construct the first-level ripple subset for each temporary node. ; B3.6.2 Following the method in step B3.5, extract the first-level ripple subset of all temporary nodes. Merge, as the outermost ripple subset S to be generated for u k ; B3.7 Merge the k-layer ripple subsets into a ripple set for seed node u. , H represents the ripple subset of the k-th layer, where H is the largest ripple layer. B4. Merge and recombine the triples (h, r, t) of each ripple subset in the ripple set S to generate candidate subgraphs G for the seed nodes. s ; C. Identify high-risk tax avoidance sub-maps, the specific steps are as follows: C1. Construct a database of high-risk tax avoidance models. The specific steps are as follows: C1.1 Based on historical audit cases and tax expert knowledge, define a set of typical high-risk tax avoidance patterns P={p1,p2,...,p r }, where r is the total number of tax avoidance schemes, and the specific steps are as follows: C1.1.1 Each of the aforementioned tax avoidance schemes p j (1≤j≤r) is represented by a subgraph with specific node roles, topological structure, and relationships; C1.1.2 Example P1: There exists a tax avoidance pattern where "five nodes, P01, P02, C01, C02, C03, C04, and C05, constitute a tax avoidance pattern for related-party transactions." The tax avoidance pattern diagram is shown in Figure 3, where: • P01 and P02 are natural persons with social relationships, and C01, C02, C03, C04, and C05 are enterprise nodes, and there exists CO_owner(P01, C01), CC_shareholding(C03, C05), and CO_owner(P02, C02). • There exists a CC_transaction(C05, C04) with significant differences in the tax rates involved; • The risk path meets the requirements of multi-layer penetration and hierarchical depth, and E meets the requirements of tail nodes of multiple risk paths; C1.2 represents p for each tax avoidance model j (1≤j≤r) Extract node attributes and topological features to generate the corresponding pattern feature vector Fp. j Taking example p1, the feature vector Fp1 contains the following core dimensions: C1.2.1 Specific types of corporate taxpayer relationships that affect related-party tax risks; C1.2.2 Threshold for Tax Rate Difference Between Nodes at Both Ends of a Specific Relationship Type , among which, T C05 T C04 This represents the tax rate for the head node C05 and the tail node C04; C1.2.3 Threshold for the percentage of related transaction amount between the two nodes of a specific relationship type Re C05 Represents the total revenue of node C05, Re CC α represents the total revenue generated from transactions between nodes C04 and C05. If there are no transactions between the nodes, α is 0. Finally obtained Where type represents the relation type; C2. Starting from the enterprise taxpayer node, identify the candidate risk subgraph G. ce ; C3. Starting from the natural person node, identify the candidate risk subgraph G. cr ; C4. Calculate tax avoidance patterns and candidate risk subgraphs G ce G cr Based on feature similarity, high-tax-risk subgraphs are identified through the following steps: C4.1 Calculate the risk subgraph G for each candidate. ce G cr And each tax avoidance model p j The feature similarity (1≤j≤r) is calculated using the following steps: C4.1.1 Extract candidate risk subgraph G using the method in step C1.

2. ce G cr eigenvector F Gce F Gcr ; C4.1.2 Calculate the tax avoidance model p using formulas (7) and (8) respectively. j and candidate subgraph G ce G cr Feature similarity of (j=1,2,…,r); (7) (8) C4.2 If sim(p) j G ce )≥θs or sim(p j G cr )≥θ s Then determine G ce or G cr For the high-tax-risk subgraph; where θ s These are configurable parameters.

2. The method for assessing related-party tax risks of corporate taxpayers based on ripple sets as described in claim 1, constructing a tax risk knowledge graph database; the specific steps are as follows: A1.1 Define corporate taxpayers and natural persons as entity types in the tax risk knowledge graph G; A1.2 Defines eight relationship types between corporate taxpayers, between corporate taxpayers and natural persons, and between natural persons, and extracts data from the system database to construct relationship data between instances; A1.3 Defines the attributes of the relationship type between entities.

3. The knowledge graph-based tax risk assessment method for corporate taxpayers as described in claim 1, starting from the corporate taxpayer node, identifies candidate risk subgraphs G. ce The specific steps are as follows: C2.1 In candidate subgraph G s Find two corporate taxpayer nodes, a and b, that have a transaction relationship. C2.2 Starting from nodes a and b, perform ripple propagation and reverse tracing in the candidate subgraph to find the actual controller nodes c and d corresponding to nodes a and b, respectively. Extract the relationship type, tax rate difference, and transaction amount ratio of nodes a and b to obtain the candidate subgraph G. s eigenvector F Gce ; C2.3 Detect whether natural person nodes c and d have social relationships; if they do, and they conform to pattern p j If the threshold condition is met, it is marked as a candidate risk subgraph G. ce .

4. The knowledge graph-based tax risk assessment method for corporate taxpayers as described in claim 1, starting from the natural person node, identifies candidate risk subgraph G. cr The specific steps are as follows: C3.1 In candidate subgraph G s Search for natural person nodes a and b that have social relationships; C3.2 Starting from nodes a and b, perform ripple propagation direction tracing in the candidate subgraph to find the enterprise taxpayer nodes c and d controlled by nodes a and b respectively. Extract the node relationship type, tax rate difference, and transaction amount ratio to obtain the candidate subgraph G. s eigenvector F Gcr ; C3.3 Detect whether there are transaction relationships between enterprise taxpayer nodes c and d that meet the threshold conditions; if so, and they conform to pattern p j If the threshold condition is met, it is marked as a candidate risk subgraph G. cr .

Images