A method for equity penetration data association analysis and hidden relationship mining

Abstract

The present application provides a method for equity penetration data association analysis and hidden relationship mining, including: performing semantic analysis on the text of the shareholders' meeting resolution through a natural language processing method, extracting sentences containing keywords, and judging whether there is shareholder voting rights delegation behavior in the resolution, where the keywords include delegation, agency, and authorization; calculating the voting weight of each shareholder according to the shareholding ratio and voting rights delegation ratio of the shareholders in the delegation chain, tracing the final shareholding ratio of the hidden controller, and determining the actual control of each shareholder over the target company; extracting an association network centered on the hidden controller in the equity relationship network, using a graph neural network to process multiple types of nodes and edge relationships in the equity network, and quantifying the influence and connectivity of entities in the hidden controller's associated network.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for equity penetration data association analysis and hidden relationship mining. Background Art

[0002] In the equity penetration correlation analysis, the study of the characteristics of shareholders' meeting resolutions faces a complex technical problem. When the voting rights of shareholders' meetings are exercised in a principal-agent model, especially when the voting rights delegation chain is nested in multiple layers, how to accurately identify the actual control relationship becomes a difficult problem. This multi-layered nested principal-agent model may lead to the opacity of the equity structure, making it difficult to identify hidden shareholders. Although the principal-agent model itself is a common way to exercise voting rights, when the delegation chain is too long and the nesting level is too deep, the difficulty of tracking the actual controller is greatly increased. This complex delegation relationship may be intentionally used to hide the real equity control structure. Therefore, how to accurately identify the actual controller in a complex principal-agent relationship, deeply analyze the characteristics and patterns of the delegation chain, and reveal the actual control relationship hidden behind the complex structure, so as to provide strong technical support for identifying hidden shareholders, has become a key technical problem that needs to be solved in the equity penetration correlation analysis. Summary of the invention

[0003] The present invention provides a method for equity penetration data association analysis and hidden relationship mining, which mainly includes:

[0004] Through natural language processing methods, semantic analysis is performed on the text of shareholders' meeting resolutions, sentences containing keywords are extracted, and it is determined whether there is shareholder voting rights delegation in the resolution. Keywords include delegation, agency and authorization; if there is a shareholders' meeting resolution with voting rights delegation, the principal and the principal are identified, and regular expressions are used to match the principal's shareholding ratio. A directed graph of voting rights delegation between shareholders is constructed based on shareholder nodes and delegation edges to form a topological structure of the delegation chain; the shareholder voting rights delegation chain is traversed, the out-degree and in-degree of each shareholder node are counted, and the centrality index of each shareholder node is calculated according to the network topology analysis method to determine whether the centrality index of each shareholder node is higher than the preset centrality threshold. If the shareholder node is higher than the preset centrality threshold, it is considered that the shareholder corresponding to the shareholder node higher than the preset centrality threshold is the actual controller hidden behind the delegation chain; the historical voting rights delegation data of shareholder nodes higher than the preset centrality threshold are obtained, and the nodes higher than the preset centrality threshold are analyzed. The position and role of the shareholder corresponding to the shareholder node in the entrustment chain. If the frequency of the shareholder appearing at the key node of the entrustment chain is higher than the preset frequency threshold, and the connected entrustment chain is deeply nested, the shareholder is determined to be a hidden controller; the attribute characteristics of the shareholder node are obtained, and the graph neural network algorithm is used to predict the probability value of the actual control of each shareholder node based on the attribute characteristics of the shareholder node and the topological structure of the entrustment chain. If the probability value of the actual control of the shareholder node is higher than the preset probability value, the shareholder corresponding to the shareholder node is determined to be a hidden controller; according to the shareholding ratio and voting power delegation ratio of the shareholders in the entrustment chain, the voting weight of each shareholder is calculated, the final shareholding ratio of the hidden controller is traced, and the actual control of each shareholder over the target company is determined; the association network centered on the hidden controller in the equity relationship network is extracted, and the multi-type node and edge relationships in the equity network are processed using the graph neural network, and the influence and connectivity of the entities in the hidden controller's associated network are quantified.

[0005] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0006] The present invention discloses a method for equity penetration data association analysis and hidden relationship mining. The method extracts voting rights delegation information and constructs a directed graph of voting rights delegation among shareholders by semantically analyzing the text of shareholders' meeting resolutions. By using network topology analysis and graph neural network algorithm, combined with shareholder attribute characteristics, the centrality index and actual control probability of each shareholder node are calculated to identify the actual controller hidden behind the delegation chain. Further combined with equity penetration analysis, the final shareholding ratio of the hidden controller is traced to determine its actual control over the target company. Finally, the associated party network of the hidden controller is extracted from the equity relationship knowledge graph to quantify the influence of each entity. The present invention can effectively identify hidden controllers in complex equity structures, reveal their actual control relationships, and provide an important basis for corporate governance and supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 The present invention is a flowchart of a method for equity penetration data association analysis and hidden relationship mining. DETAILED DESCRIPTION

[0008] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0009] like Figure 1 In this embodiment, a method for equity penetration data association analysis and hidden relationship mining may specifically include:

[0010] S101. Perform semantic analysis on the text of the shareholders' meeting resolution through natural language processing methods, extract sentences containing keywords, and determine whether there is any shareholder voting rights delegation behavior in the resolution. Keywords include delegation, agency and authorization.

[0011] The text of the shareholders' meeting resolution is received and punctuation marks are removed through text cleaning to obtain a semantic index database; a neural network is used to perform semantic matching on text keywords based on the semantic index database, and a semantic keyword combination is calculated through an edit distance algorithm; a sequence labeling method is used to label the grammatical components of sentences for the semantic keyword combination, and semantic dependency relationship data is obtained through conditional random field labeling; a text vector space model is used to measure the relevance of the context based on the semantic dependency relationship data, and if the context similarity is greater than a similarity threshold, it is determined to be a voting right delegation object; a text topic model is established for the voting right delegation object, and semantic clustering is performed through a latent Dirichlet distribution algorithm, and if the clustering result belongs to the voting right delegation class, delegation relationship labeling data is obtained; semantic labeling is performed on the text paragraphs based on the delegation relationship labeling data, and the relationship between the principal and the trustee is matched through a semantic rule base, and if the matching degree is greater than the matching degree threshold, the voting right delegation result is obtained.

[0012] Specifically, the text of the shareholders' meeting resolution is preprocessed according to the text segmentation rules, punctuation marks and special characters are removed by text cleaning, and the text is segmented by part-of-speech tagging method to establish a text semantic index database. A neural network is used to semantically match the text keyword combination for the text semantic index database, and the similarity of the keywords delegation, agency, and authorization is calculated by the edit distance algorithm. If the semantic similarity is greater than the preset threshold, the semantic keyword combination is obtained. For the semantic keyword combination, the sequence tagging method is used to annotate the grammatical components of the sentences containing the keywords, and the conditional random field is used to annotate the semantic dependency relationship between the main and subordinate clauses. If the semantic dependency is greater than the preset threshold, it is determined to be a valid sentence. The text vector space model is used to measure the relevance of the context of the keywords in the valid sentence, and the text context is calculated by semantic rules. If the context similarity is greater than the preset threshold, it is determined to be a sentence related to voting rights. A text topic model is established for the sentences related to voting rights, and the sentences are semantically clustered by the latent Dirichlet distribution algorithm. If the clustering result belongs to the voting rights delegation class, it is determined that there is a voting rights delegation relationship. According to the voting rights delegation relationship, the text paragraphs are semantically annotated, and the relationship between the principal and the trustee is matched through the semantic rule library. If the matching degree is greater than the preset threshold, the final voting rights delegation object is obtained. In the text processing of the shareholders' meeting resolution, the text segmentation rule uses punctuation and paragraph marks as the separation basis, segments the original text, and obtains the standardized text paragraphs by removing special characters such as spaces and line breaks. Then, based on the part-of-speech tagging method, the nouns, verbs and other parts of speech are identified, and the semantic index corresponding to the words and parts of speech is constructed. For keyword semantic matching, the neural network model determines the degree of semantic association by calculating the cosine similarity between word vectors. For example, the similarity between "full power of attorney" and "delegation" is 0.85, which is higher than the preset threshold of 0.8, and is determined to be a similar semantic combination. The edit distance algorithm calculates the word form change distance. The edit distance between "delegation authorization" and "authorization delegation" is 2, which is lower than the preset distance 4, and is determined to be an equivalent expression. In the grammatical component annotation, the sequence annotation method identifies the sentence components. For example, "Shareholder Party A entrusts Party B to attend the meeting on his behalf" is annotated as the grammatical structure of "subject + attributive + predicate + object + adverbial + complement". The conditional random field calculates the semantic dependency probability to be 0.92, which is higher than the 0.9 threshold, and is determined to be a valid sentence structure for voting rights delegation. The text vector space model measures the context relevance. Taking "entrusted to attend" as the central word, the words in the previous and next text windows are extracted to construct the context vector. The calculated vector angle cosine value is 0.88, which is higher than the 0.85 threshold, indicating that the context has a strong voting right relevance. The topic model clusters the sentences based on the latent Dirichlet distribution algorithm, setting the number of topics to 5, including voting rights delegation, shareholder identity, meeting agenda, voting procedures, and other matters. By calculating the distribution probability of the sentence under each topic, the highest probability of 0.75 corresponds to the voting rights delegation topic.In the matching of entrustment relationships, the semantic rule library contains rule templates such as entrustor qualifications, entrusted authority, and entrustment period, and performs pattern matching on the text. For example, the matching degree of the rule "the entrustor has shareholder qualifications" is 0.95, which is higher than the 0.9 threshold, confirming that the entrustment relationship is compliant and valid. Through multiple rule matching and threshold judgment, the voting rights entrustment objects and their specific authority scope in the text are finally identified. In the semantic annotation stage, the system will structurally annotate the identified elements such as entrustor, trustee, and entrustment authority to form a complete description of the voting rights entrustment relationship including the identity information of both entrusting parties, the scope of entrusted matters, and the entrustment period. The accuracy of the annotation results reached 0.93, and the recall rate was 0.89. The comprehensive performance meets the actual application needs.

[0013] S102. If there is a shareholders' meeting resolution for voting rights delegation, identify the principal and the delegate, use regular expressions to match the principal's shareholding ratio, and construct a directed graph of voting rights delegation between shareholders based on shareholder nodes and delegation edges to form a topological structure of the delegation chain.

[0014] The voting rights delegation text is parsed according to the preset entity recognition rules to obtain entity attribute data including the principal identifier, the principal identifier and the shareholding value; an attribute graph data model is established through the entity attribute data, and the attribute graph data model includes shareholder entity nodes and delegation relationship edges, and the delegation relationship edges store the shareholding value and delegation time information; the delegation relationship edges between the shareholder entity nodes are verified, and if the principal's shareholding ratio is higher than the shareholding ratio threshold and the delegation period meets the rule requirements, a directed edge with a weighted value is established in the attribute graph data model; a depth-first search algorithm is used to perform path search on the directed edge with a weighted value, and a delegation chain topology sequence is obtained by accumulating the path weight values, and the delegation chain topology sequence marks the actual ownership relationship of the voting rights.

[0015] Specifically, the voting rights delegation text is subjected to entity extraction according to the regular matching rules, the entrusting party and the entrustee are annotated by the named entity recognition method, the text pattern matching is used to extract the entrusting party's shareholding value information, and an entity attribute table containing the shareholder name, the number of shares held, and the entrustment time is established for the extraction results. A graph database model is designed for the entity attribute table, and the data structure of the shareholder entity node and the entrustment relationship edge is established by defining a graph structure pattern containing node identification, node type, and node attributes, and the attribute graph model is used to store the relationship data between nodes and edges. According to the preset entrustment rule set, the entrusting parties are authenticated, the number of entrustment voting rights is determined by the equity ratio threshold, the entrustment period is verified by the timeliness rule, and a directed edge with a weight value is established for the verified entrustment relationship. Path calculation is performed for the directed edge with a weight value, and the path between nodes is recursively searched by the depth-first search algorithm. The weights of multiple entrustment paths are accumulated by the shortest path algorithm to obtain a complete entrustment chain topology sequence. According to the topological sequence of the delegation chain, a directed graph structure is constructed. The path is closed-loop verified by the connectivity test method. The nodes are hierarchically divided by the topological sorting algorithm to determine the in-degree and out-degree values ​​of each node. Cycle detection is performed on the directed graph structure. The loop structure is identified by the depth-first traversal algorithm. The priority rule is used to handle the cyclic delegation situation. The final directed acyclic graph is constructed and the actual ownership of the voting rights is marked. In the text of the shareholders' meeting resolution, the shareholder identity information and shareholding data are identified by regular expressions. For example, in the text "Shareholder Zhang holds 3 million shares of the company, accounting for 15% of the total share capital, and entrusts Li to attend the shareholders' meeting and exercise voting rights on his behalf", the regular pattern of "holding.? shares.? million shares" is used to extract the number of shares held, and the shareholding ratio is matched based on "accounting for.?%". The delegation relationship is identified through "entrust.? on behalf of", thereby establishing an entity attribute table containing shareholder code, number of shares held, and shareholding ratio. In the graph database design, the node attributes include basic information such as shareholder code, shareholder type, and number of shares held. The edge attributes include relational data such as entrusted shares, entrusted period, and entrusted authority, such as node attributes {shareholder code: SH001, number of shares held: 3 million shares}, edge attributes {entrusted shares: 15%, entrusted start date: 20240301, entrusted end date: 20240630}, and a complete data structure is constructed through the attribute graph model. In the verification of entrustment rules, multiple identities of both parties are verified. The shareholding ratio verification requires that the actual number of shares held by the entrusting party is not less than the entrusted voting share. The timeliness rule requires that the entrustment period is within the valid range. For example, a shareholder holds 5 million shares, entrusts 3 million shares for voting, and the entrustment period is 3 months. It is verified that it meets the requirements of the entrustment rules. In the path calculation process, all possible paths from the starting node to the target node are found through depth-first search, such as A entrusts B 15% equity, B entrusts C 10% equity, and C entrusts D 5% equity. The complete entrustment transfer chain is obtained through weight accumulation calculation, and the actual voting rights are finally determined.In the directed graph structure analysis, the in-degree of a node indicates the number of times it has been delegated, and the out-degree indicates the number of times it has been delegated externally. For example, if the in-degree of node B is 1, it indicates that it has been delegated once, and the out-degree is 1, it indicates that it has been delegated externally once. The nodes are divided into different levels through topological sorting, and the hierarchical relationship of delegation transfer is identified. In the processing of cyclic delegation, when a loop structure of A→B→C→A is found, the actual ownership of voting rights is determined according to the priority rules. For example, according to the principle of "time priority", the delegation relationship with the earliest establishment time takes precedence, and other conflicting delegation relationships automatically become invalid, thereby eliminating the ambiguity of voting rights ownership caused by cyclic delegation, and finally constructing a clear directed acyclic graph structure.

[0016] S103. Traverse the shareholder voting rights delegation chain, count the out-degree and in-degree of each shareholder node, calculate the centrality index of each shareholder node according to the network topology analysis method, and determine whether the centrality index of each shareholder node is higher than the preset centrality threshold. If the shareholder node is higher than the preset centrality threshold, it is considered that the shareholder corresponding to the shareholder node higher than the preset centrality threshold is the actual controller hidden behind the delegation chain.

[0017] A breadth-first search algorithm is used to obtain the delegated out-degree value and the entrusted in-degree value of the shareholder node, and the relative degree weight of the node is calculated based on the delegated out-degree value and the entrusted in-degree value; the connection weight value between the node and the adjacent nodes is calculated based on the relative degree weight, and the initial centrality value of the node is obtained through eigenvector iteration calculation; the maximum and minimum value normalization processing is performed on the initial centrality value, and the control range value of the node is calculated through the path analysis method; the comprehensive score of the node is calculated based on the control range value, and if the comprehensive score is higher than the preset centrality threshold, the node is determined to be the actual controller node.

[0018] Specifically, the structure of the shareholder voting rights delegation graph is traversed, and the breadth-first search is used to count the delegated out-degree value and the entrusted in-degree value of each shareholder node. The out-degree value and the in-degree value are weighted by the node degree calculation formula to obtain the relative degree weight, and the basic degree distribution sequence is constructed for all shareholder nodes.

[0019]

[0020] , W i represents the relative degree weight of node i, α and β are the weight coefficients of out-degree and in-degree respectively, n represents the total number of nodes in the network, and D out (i) represents the out-degree value of the i-th node, D in(i) represents the in-degree value of the i-th node, and the denominator represents the sum of the out-degree and in-degree of all nodes. The eigenvector parameters are set according to the basic degree distribution sequence. By calculating the connection weight value of each node and the adjacent node, the indirect influence of the node is iteratively attenuated, and the initial centrality value of the node is obtained by the eigenvector algorithm. The initial centrality value is normalized, and the value is mapped to the interval of zero to one by the maximum and minimum value standardization method. The control power of the node in the entrustment chain is quantitatively calculated, and the control range value of the node is obtained by the path analysis method. The preset centrality threshold is determined according to the historical data statistical method, the threshold is dynamically calibrated by the node clustering method, and the threshold is segmented by the centrality interval division rule. A node scoring system is established for the control range value, the comprehensive score of the node is calculated by the weighted summation method, and the importance of the node is ranked by the score ranking method. The threshold is judged according to the comprehensive score of the node, and the nodes above the threshold are identified by the preset centrality threshold, and the actual controller node is obtained by the network convergence judgment method. In the equity entrustment network analysis, the node degree statistics reflect the strength of the shareholder's entrustment relationship. Taking the equity structure of a company as an example, shareholder A delegated voting rights 3 times and accepted 2 voting rights entrustments. The out-degree value is 3 and the in-degree value is 2. The relative degree weight is 0.25 after weighted calculation, indicating that the shareholder has strong connectivity in the entrustment network. In the process of calculating the characteristic vector of the node, the transmission attenuation of the adjacent relationship is considered. For example, shareholder B directly accepts the voting rights entrustment of A, the connection weight is 1.0, and the weight attenuation of C indirectly affected by B is 0.8. After iterative calculation, the initial centrality value of node B is 0.85, reflecting the important position of B in the network. The maximum and minimum value standardization method is used in the normalization processing stage to map the centrality values ​​of all nodes to the range of 0 to 1. For example, the original value of 0.85 is converted to 0.92, which is convenient for comparative analysis between different nodes. Through path analysis, it is found that the control range of this node involves 5 downstream nodes, and the control range value is 0.88. In the process of determining the threshold, the distribution law of centrality is obtained by analyzing historical data, and the clustering method is used to divide the nodes into three categories: high, medium and low. The high influence threshold is set to 0.85, the medium influence threshold is set to 0.65, and the low influence threshold is set to 0.45 to achieve a reasonable division of the importance of nodes. The node scoring system integrates multiple dimensional indicators, with a control range weight of 0.4, a centrality weight of 0.3, and a connection strength weight of 0.3. The various indicators of a certain node are weighted to obtain a comprehensive score of 0.78, ranking third among all nodes, showing strong control capabilities. Based on the preset threshold of 0.75, the node is judged and it is found that the comprehensive score of shareholder D is 0.88, which is significantly higher than the threshold, and the proportion of voting rights controlled by him reaches 35%. The stability of the control relationship is verified by network convergence analysis, and D is finally confirmed as the actual controller of the company, forming a complete control right judgment chain.In actual applications, a listed company has undergone multiple voting rights delegations, forming a complex delegation network. Node E controls more than 40% of the voting rights through multi-layer delegation relationships. Its centrality index is 0.95, far exceeding the preset threshold of 0.75. Combined with its control range and comprehensive score, it clearly reveals the actual control relationship hidden behind the delegation chain.

[0021] S104. Obtain historical voting rights delegation data of shareholder nodes that are higher than a preset centrality threshold, analyze the position and role of shareholders corresponding to shareholder nodes that are higher than the preset centrality threshold in the delegation chain, and if the frequency with which a shareholder appears at a key node in the delegation chain is higher than a preset frequency threshold, and the connected delegation chain has a deep nesting layer, then the shareholder is determined to be a hidden controller.

[0022] The historical entrustment records of high-centrality shareholder nodes are obtained by screening the high-centrality shareholder nodes from the shareholder nodes through a preset centrality threshold; the balance of entrustment shares and temporal continuity are verified by using data validity verification rules according to the historical entrustment records, so as to obtain a standardized data set including both entrusting parties and the number of entrustments; a depth-first search algorithm is used to traverse the entrustment chain for the standardized data set, and a node weighted frequency value is obtained by calculating a hierarchical weight function, wherein the hierarchical weight function adopts a weighting method in which the first-level weight value is a preset value that increases layer by layer; if the node weighted frequency value is higher than the preset frequency threshold and the number of nested levels of the entrustment chain exceeds the preset hierarchical threshold, the node is marked as a hidden controller.

[0023] Specifically, high-centrality shareholder nodes are screened according to the centrality threshold, all historical entrustment records are obtained through time span statistics, and the entrustment records are cleaned by text format standardization and numerical range verification. A historical data table containing entrustment time, entrustment object, and entrustment share is established for each node. The entrustment record is verified by data validity verification rules for the historical data table, and abnormal data is identified through entrustment share balance verification and time series continuity check. A standardized data set containing both entrusting parties and the number of entrustments is established for the verified records. A depth-first search algorithm is used to traverse the entrustment chain for the standardized data set, and the hierarchical position of the node in different chains is identified through the path tracing method. The basic frequency data is obtained by counting the number of node occurrences. A hierarchical weight function is established based on the basic frequency data. The importance of deep nodes is quantified by setting the weight of the first level to 1 and increasing by 0.5 layer by layer. The frequency value after considering the hierarchical weighting is calculated for each node.

[0024] .W lRepresents the weight value of the lth layer, l represents the number of levels, the initial level weight is 1, and the weight increases by 0.5 for each additional level. Set a frequency benchmark threshold for the weighted frequency value, determine the normal frequency distribution range through historical data statistics, and use the mean plus standard deviation method to calculate the frequency threshold value. Perform a preliminary screening of nodes based on the frequency threshold, count the number of nested levels of the entrustment chain, and use the recursive counting method to obtain the actual nesting depth value. Set a level benchmark threshold for the nested depth value, determine the regular nested level range through the entrustment chain statistics method, and use the quantile method to calculate the level threshold value. Perform a double judgment based on the frequency threshold and the level threshold, and mark the nodes that meet the weighted frequency value higher than the frequency threshold and the number of nested levels exceeding the level threshold as hidden controllers. In the processing of equity entrustment data, the historical records of high-centrality shareholder nodes contain multiple formats. For example, the entrustment text "Shareholder Zhang entrusts Li to exercise the voting rights of his 25% shares on 20240301" needs to be standardized to extract key information, including the entrusting parties, entrustment date, and entrustment ratio. At the same time, the numerical range is verified to ensure that the entrustment ratio is between 0% and 100%. In the data validity verification, the balance relationship of the entrustment share is emphasized. In a case, shareholder Wang holds 40% of the shares, and entrusts 35% to Chen and 10% to Liu, respectively. The total entrustment of 45% exceeds the actual shareholding. The abnormal data is identified through the share balance verification, and the entrustment records are eliminated to ensure data accuracy. During the traversal of the entrustment chain, it was found that the node Zhao was located in the 2nd, 3rd, 3rd, 4th, and 5th layers in 5 different chains, with a basic frequency of 5 times. The weighted frequency was calculated based on the hierarchical position, with a weight of 1.5 for the 2nd layer, 2.0 for the 3rd layer, 2.5 for the 4th layer, and 3.0 for the 5th layer, and the final weighted frequency value was 13.5. The determination of the frequency benchmark threshold was based on historical data analysis. By statistically analyzing the entrustment records in the past three years, it was found that the weighted frequency distribution range of normal nodes was 2.0 to 8.0, with a mean of 5.0 and a standard deviation of 1.5. The frequency threshold was set to the mean plus two times the standard deviation, i.e. 8.0. A value above this threshold indicates that the node appears frequently in the entrustment chain. In the hierarchical complexity analysis, a company has a 6-layer delegation chain of A→B→C→D→E→F. The actual nesting depth calculated by the recursive method is 6, while the nesting level of the conventional delegation chain ranges from 2 to 4 layers. The hierarchical threshold calculated by the 75% quantile method is 4. Exceeding this threshold indicates that the delegation structure is abnormally complex. In a specific case, the weighted frequency value of shareholder Zheng is 12.5, which is higher than the frequency threshold of 8.0, and the delegation chain in which he participates reaches the deepest 7 layers, exceeding the hierarchical threshold of 4. At the same time, it is found that the shareholder actually controls more than 30% of the voting rights through multi-layer nesting, which meets the characteristics of a hidden controller. In another case, although Zhou's weighted frequency value reached 9.5, the deepest delegation chain was only 3 layers, which did not exceed the hierarchical threshold and did not meet the judgment conditions of a hidden controller.The comprehensive judgment results show that a total of 15 suspected hidden controller nodes were found in the equity structures of 10 listed companies, of which 8 nodes had weighted frequency values ​​and nested levels exceeding the threshold at the same time. Their hidden controller identities were confirmed through subsequent verification, reflecting the accuracy of the dual threshold judgment method.

[0025] S105. Obtain the attribute characteristics of the shareholder node, and use the graph neural network algorithm to predict the probability value of the actual control of each shareholder node based on the attribute characteristics of the shareholder node and the topological structure of the delegation chain. If the probability value of the actual control of the shareholder node is higher than the preset probability value, the shareholder corresponding to the shareholder node is determined to be a hidden controller.

[0026] A type coding method is used for the shareholder information to obtain the classification labeling values ​​of natural person shareholders and institutional shareholders, and a numerical feature vector of the shareholding ratio is obtained through normalization processing; a graph neural network structure is constructed according to the numerical feature vector, and the graph neural network structure includes a feature aggregation layer, a spatial attention layer and a node update layer; a graph convolution operator is used to perform aggregation operations on the numerical feature vector and the adjacency relationship, and a node embedding vector of the fused structural information is obtained through attention coefficient calculation; a cross entropy loss function value is calculated for the node embedding vector, and if the loss function value is greater than the loss function threshold, a back propagation method is used to optimize and adjust the network parameters.

[0027] Specifically, node attribute features are extracted from the shareholder information database, and natural person shareholders and institutional shareholders are classified and labeled by type coding methods. Normalization is used to convert the shareholding ratio into numerical values. The correlation score is calculated for the length of shareholding and the position of directors, supervisors and senior managers, and a node feature vector of fixed dimension is constructed. A graph neural network structure is constructed for the node feature vector of fixed dimension. By designing a three-layer network architecture including feature aggregation layer, spatial attention layer and node update layer, the input dimension and output dimension of each layer of the network are fixedly configured. A historical control case library is constructed based on the training data set, and the known control cases are divided into positive and negative samples by manual annotation methods. The training set and validation set are divided by cross-validation method, and the network parameters are scaled. The graph convolution operator is used to aggregate node features and adjacency relationships for the network architecture, and the information of different neighbor nodes is weighted by calculating the attention coefficient to obtain the node embedding vector that integrates the structural information. Feature mapping is performed through a multi-layer perceptron according to the node embedding vector, and the prediction error is calculated by the cross entropy loss function. The network parameters are optimized and adjusted by the back propagation method to obtain a trained recognizer. The validation set is used to evaluate the performance of the output results of the identifier. The hidden controller identification model is comprehensively evaluated by calculating the accuracy, recall rate, precision and other indicators to obtain the optimal probability threshold. The preset probability threshold is calibrated according to the evaluation results of the validation set, and the threshold is fine-tuned by the example verification method. If the control probability is higher than the calibrated threshold, the node is determined to be a hidden controller. The confidence evaluation is performed on the judgment result, and the reliability score is scored by calculating the difference between the predicted probability and the threshold. The stability of the judgment result is verified by a multi-round verification method. In the processing of shareholder attribute features, the node type encoding adopts the unique hot encoding method, the natural person shareholder is encoded as [1,0], the institutional shareholder is encoded as [0,1], the shareholding ratio is converted to the [0,1] interval through maximum and minimum normalization, the length of shareholding is calculated in months and then normalized, and the directors, supervisors and senior managers are assigned weight scores of 0.2 to 1.0 according to their job levels, and finally a 64-dimensional feature vector is constructed. In the design of the graph neural network architecture, the dimension of the feature aggregation layer is 64×32, and the local representation of the node is obtained by weighted averaging the features of neighboring nodes. The dimension of the spatial attention layer is 32×16, and the structural similarity and feature similarity between nodes are considered when calculating the attention coefficient. The dimension of the node update layer is 16×8, which integrates the original features of the nodes with the aggregated features. In the historical control case, shareholder A of a listed company actually controls 35% of the shares through multi-layer entrustment and is marked as a positive sample, and shareholder B only holds 5% of the shares and has no obvious control behavior and is marked as a negative sample. The 5-fold cross-validation method is used to divide and obtain 800 training samples and 200 validation samples.During the node embedding process, the neighbor nodes with degree centrality of 0.8 and feature similarity of 0.9 are given an attention weight of 0.95, while the neighbor nodes with degree centrality of 0.3 and feature similarity of 0.4 are only given a weight of 0.25, highlighting the information contribution of important nodes. In the model training stage, the calculation result of the loss function is reduced from the initial 0.85 to 0.15, and the accuracy rate, recall rate and precision rate of 0.92, 0.89 and 0.90 are achieved on the validation set. The optimal probability threshold is determined to be 0.75 through ROC curve analysis. In the instance verification, it is found that the control probability of shareholder C is 0.82, which exceeds the threshold of 0.75. Its characteristics are that it holds a 20% stake, serves as the vice chairman, and controls an additional 15% of the equity through a 3-layer nested delegation chain. The comprehensive indicators show that it has strong hidden control characteristics. The confidence assessment shows that the predicted probability of shareholder C differs from the threshold by 0.07, and the predicted result remains stable after 10 rounds of cross-validation, while the control probability of shareholder D is 0.76, which is only 0.01 different from the threshold. The predicted results fluctuate in multiple rounds of validation, and more evidence is needed to make a judgment. In actual application, 50 shareholder nodes of a group company were analyzed, and 3 high-probability hidden controllers were identified, with control probabilities of 0.88, 0.85 and 0.83, respectively, which are significantly higher than the threshold, and they show obvious control characteristics in multiple dimensions such as feature performance, delegation chain, and control scope, verifying the effectiveness of the model.

[0028] S106. Calculate the voting weight of each shareholder based on the shareholding ratio and voting power delegation ratio of the shareholders in the delegation chain, trace the ultimate shareholding ratio of the hidden controller, and determine the actual control of each shareholder over the target company.

[0029] According to the shareholder register, the direct shareholding records and voting rights delegation records are obtained, the shareholding ratios are standardized through normalization, and the basic voting weights are obtained by linear combination method; a shareholding relationship diagram is constructed for the basic voting weights, the shareholding relationship diagram is layered through topological sorting method, and the penetration shareholding ratio is calculated by recursive traversal algorithm; a weight attenuation coefficient is set according to the penetration shareholding ratio, and an independent control coefficient is obtained by calculating the weight attenuation coefficient, and the independent control coefficients are grouped and merged by related party determination method to obtain group control power; a hierarchical clustering algorithm is used to group the group control power, and the group results are cross-validated through historical voting records, and the actual controller is marked for the high control power group that passes the verification.

[0030] Specifically, direct shareholding records and voting rights delegation records are extracted from the shareholder register database, and the shareholding ratio is standardized through normalization. The weighted sum of the shareholding ratio and the delegated voting rights ratio is calculated, and the basic voting weight is calculated using the linear combination method. A shareholding relationship diagram is constructed for the basic voting weight, and the multi-layer shareholding structure is stratified through the topological sorting method. The shareholding ratio of each shareholder is calculated layer by layer using the recursive traversal algorithm, and the penetration shareholding ratio is obtained by multiplying the shareholding values ​​of each layer. The weight attenuation coefficient is set according to the penetration shareholding ratio, and the weight is calculated by setting the direct shareholding weight to 1.0 and attenuating it by 0.1 for each additional layer of indirect shareholding. The independence of shareholders at all levels is evaluated to obtain the independent control coefficient. Voting rights aggregation rules are established for the independent control coefficient, and the relationship of concerted action is identified through the related party determination method. The voting rights of related shareholders are summed up by the group merging method to obtain the group control power. The control power benchmark value is set according to the group control power, and the control power mean and standard deviation are calculated through the equity structure statistical method. The significant control power threshold is determined by the mean plus two standard deviation methods. Hierarchical clustering is performed for the significant control threshold. By setting the number of clusters to 3, the hierarchical clustering algorithm is used to group the control values, and high control groups, medium control groups and low control groups are obtained. Control is verified based on the grouping results, and the control grouping is cross-validated through historical voting records. The accuracy of the control grouping is verified by consistency test method. The high control group that passes the verification is finally marked, and the main controller is determined by the control ranking method. Shareholders with high ranking and significant control are marked as actual controllers. In the equity structure analysis, the calculation of basic voting weights needs to comprehensively consider direct shareholding and voting rights delegation. For example, shareholder A directly holds 25% of the company's shares and obtains 15% of the voting rights delegation from other shareholders. After normalization, the basic voting weight is 0.4, which reflects the initial voting ability of the shareholder. The construction of the equity relationship diagram highlights the complexity of the multi-layer shareholding structure. In a real case, shareholder A holds 60% of the shares of Company B, Company B holds 40% of the shares of Company C, and Company C holds 30% of the shares of the target company. The three-layer shareholding relationship is identified through topological sorting, and the cumulative multiplication calculation shows that A's penetration shareholding ratio of the target company is 7.2%. The weight decay calculation reflects the decreasing control of indirect shareholding. Taking the above case as an example, A, as the third-layer indirect shareholder, has a weight decay coefficient of 0.7 and an actual control of 5.04%, while the weight coefficient of shareholder D, who directly holds 30%, is 1.0, and the actual control remains at 30%, reflecting the control advantage of direct shareholding. In the identification of related parties, shareholders E and F hold 18% and 12% of the shares respectively. The two have a kinship relationship and their historical voting records are highly consistent. They are identified as persons acting in concert through the group merging method. The group control after the merger reaches 30%, which significantly improves its overall control position.The determination of the control power threshold is based on empirical statistics. The analysis of the equity structure of 100 listed companies found that the mean control power was 15% and the standard deviation was 5%. Based on this, the significant control power threshold was set at 25%. Shareholders above this threshold have strong actual control power. The hierarchical clustering results show that in the equity structure of a certain company, the control power range of the high control power group is 25%-35%, including 2 shareholder groups, the medium control power group range is 15%-25%, including 3 shareholder groups, and the low control power group range is below 15%, including multiple dispersed shareholders. In the cross-validation process, the voting records of the shareholders' meeting in the past three years were analyzed, and it was found that the support rate of the high control power group in the voting on major matters reached 95%, and its proposal approval rate reached 90%, which verified the accuracy of the control power grouping. The final marking shows that the company has two main control groups, which hold 33% and 28% of the voting rights respectively, and realize the joint control of the company through a multi-layer shareholding structure and voting rights delegation. This multi-layer control structure is relatively common in practice. Another case shows that an investment group has achieved actual control over a number of downstream listed companies through a five-layer shareholding arrangement. Although the final shareholding ratio does not exceed 10%, control over major decisions is ensured through voting rights delegation and concerted action arrangements.

[0031] S107. Extract the association network centered on the hidden controller in the equity relationship network, use graph neural network to process the multi-type node and edge relationships in the equity network, and quantify the influence and connectivity of entities in the hidden controller's associated network.

[0032] Obtain the neighboring nodes of hidden controllers in the equity relationship knowledge graph, traverse the neighboring nodes through the breadth-first search method, and obtain the related party network including shareholder entities, company entities and executive entities; construct a heterogeneous graph neural network based on the related party network, extract features of equity relations, investment relations and employment relations through the entity embedding layer, relationship conversion layer and feature aggregation layer, and obtain a multi-dimensional feature matrix; use the eigenvector method to calculate the degree centrality, betweenness centrality and closeness centrality of the nodes for the multi-dimensional feature matrix, and use the clustering method to divide the nodes into communities; use the random walk method to serialize and sample the nodes according to the community division results, and use the link prediction method to calculate the association probability between entity node pairs to obtain the association strength matrix between entities.

[0033] Specifically, the neighboring nodes of hidden controllers are extracted according to the knowledge graph of equity relations, and the related entities are obtained within the range of 1 to 3 hops through the breadth-first search method. The types of shareholder entities, company entities, and executive entities are annotated, and the related party network is generated using the subgraph construction rule. A heterogeneous graph neural network architecture is designed for the related party network. By constructing a three-layer structure including an entity embedding layer, a relationship conversion layer, and a feature aggregation layer, weight values ​​are assigned to equity relations, investment relations, and employment relations, respectively, to obtain a multidimensional feature matrix. The centrality of the entity node is calculated according to the multidimensional feature matrix, and the degree centrality, betweenness centrality, and proximity centrality of the node are weightedly calculated by the eigenvector method to obtain the comprehensive centrality index of the node. A threshold interval is set for the comprehensive centrality index, and the nodes are divided into three levels: core layer, middle layer, and edge layer by the clustering method. The label propagation algorithm is used to divide the nodes into communities. The modularity index is calculated based on the community division results, and the validity of the community structure is verified by setting a modularity threshold. Communities with modularity higher than the threshold are retained and a community relationship graph is constructed. Random walk sampling is performed on the community relationship graph. The nodes are serialized and sampled by setting the walk depth and the number of walks. The jump probability method is used to balance the local structure and the global structure. The entity embedding vector is constructed according to the sampling sequence. The association probability between entity node pairs is calculated by the link prediction method to obtain the association strength matrix between entities. The influence evaluation function is designed for the association strength matrix. The indicators of the three dimensions of node centrality, community status, and association strength are integrated by the weighted combination method. The cross-validation method is used to verify the accuracy of the evaluation results. In the analysis of the related party network, the breadth-first search method starts from the hidden controller node and expands layer by layer. For example, a hidden controller is associated with 5 holding companies within 1 hop, 15 shareholding companies within 2 hops, and 25 executives within 3 hops, forming a related party network containing 45 entity nodes. The design of the heterogeneous graph neural network fully considers the importance of different types of relationships, assigning a weight of 0.5 to the equity relationship, a weight of 0.3 to the investment relationship, and a weight of 0.2 to the position relationship. The entity embedding dimension is set to 128 dimensions, and different types of node information are integrated into a unified feature space through a three-layer network structure. During the calculation of centrality, the degree centrality of a core node is 0.85, indicating that it has extensive direct connections, the betweenness centrality is 0.78, indicating that it plays an important role in information transmission, and the proximity centrality is 0.92, reflecting that the path to other nodes is short. Combining these three indicators, the comprehensive centrality of the node is 0.85. The community division results show that the network of related parties forms three main communities. The first community contains 12 nodes with a modularity of 0.75, the second community contains 8 nodes with a modularity of 0.68, and the third community contains 6 nodes with a modularity of 0.62, all of which are higher than the modularity threshold of 0.6, indicating that the community structure is significant.The random walk sampling is set to 10 depth, 100 walks, and 0.15 jump probability. In a certain sampling sequence, starting from the hidden controller node, it passes through its holding company, chairman, affiliated company, financial director and other nodes in turn, capturing a complex association path. The entity embedding results show that the node pairs closely related to the hidden controller have a higher cosine similarity in the vector space, such as the similarity of its holding company is 0.92, the similarity of companies where relatives serve as executives is 0.85, and the similarity of unrelated companies is only 0.25, reflecting the difference in association strength. In the impact assessment, the centrality index of a key node is 0.85, the community status index is 0.78, and the association strength index is 0.82. The weighted combination results in a comprehensive influence score of 0.83. The accuracy and recall rates obtained through 5-fold cross validation are 0.88 and 0.85, which confirm the reliability of the evaluation results. In an actual case, a certain group formed a closely related network through complex equity structure and personnel arrangements. The related parties were distributed in three main communities. The core community included the group headquarters and major holding companies, the secondary community included shareholding companies and supply chain companies, and the peripheral community included related companies controlled by relatives. The hidden control relationship was accurately revealed through multi-dimensional indicator analysis.

[0034] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for equity penetration data association analysis and hidden relationship mining, characterized in that: The method comprises: Through natural language processing methods, semantic analysis is performed on the text of shareholders' meeting resolutions, sentences containing keywords are extracted, and it is determined whether there is shareholder voting rights delegation in the resolution. Keywords include delegation, agency and authorization; if there is a shareholders' meeting resolution with voting rights delegation, the principal and the delegate are identified, and regular expressions are used to match the principal's shareholding ratio. A directed graph of voting rights delegation between shareholders is constructed based on shareholder nodes and delegation edges to form a topological structure of the delegation chain; the shareholder voting rights delegation chain is traversed, the out-degree and in-degree of each shareholder node are counted, and the centrality index of each shareholder node is calculated according to the network topology analysis method to determine whether the centrality index of each shareholder node is higher than the preset centrality threshold. If the shareholder node is higher than the preset centrality threshold, it is considered that the shareholder corresponding to the shareholder node higher than the preset centrality threshold is the actual controller hidden behind the delegation chain; the historical voting rights delegation data of shareholder nodes higher than the preset centrality threshold are obtained, and the shareholders higher than the preset centrality threshold are analyzed. The position and role of the shareholder corresponding to the node in the delegation chain. If the frequency of the shareholder appearing at the key node of the delegation chain is higher than the preset frequency threshold, and the number of nested layers of the connected delegation chain exceeds the level threshold, the shareholder is determined to be a hidden controller; the attribute characteristics of the shareholder node are obtained, and the graph neural network algorithm is used to predict the probability value of the actual control of each shareholder node based on the attribute characteristics of the shareholder node and the topological structure of the delegation chain. If the probability value of the actual control of the shareholder node is higher than the preset probability value, the shareholder corresponding to the shareholder node is determined to be a hidden controller; according to the shareholding ratio and voting power delegation ratio of the shareholders in the delegation chain, the voting weight of each shareholder is calculated, the final shareholding ratio of the hidden controller is traced, and the actual control of each shareholder over the target company is determined; the association network centered on the hidden controller in the equity relationship network is extracted, and the graph neural network is used to process the multi-type nodes and edge relationships in the equity network, and the influence and connectivity of the entities in the hidden controller's associated network are quantified.

2. The method according to claim 1, characterized in that: The text of the shareholders' meeting resolution is semantically analyzed by natural language processing methods to extract sentences containing keywords, and determine whether there is any shareholder voting right delegation behavior in the resolution. Keywords include delegation, agency and authorization, including: Receive the shareholder meeting resolution text and remove punctuation marks through text cleaning to obtain a semantic index database; Using a neural network to perform semantic matching on text keywords according to the semantic index database, and obtaining a semantic keyword combination by using an edit distance algorithm; Using a sequence labeling method to label the grammatical components of the sentence for the semantic keyword combination, and obtaining semantic dependency relationship data through conditional random field labeling; Using a text vector space model to measure the relevance of the context according to the semantic dependency data, if the context similarity is greater than a similarity threshold, the context is determined to be a voting right delegation object; A text topic model is established for the voting right delegation object, and semantic clustering is performed by latent Dirichlet distribution algorithm. If the clustering result belongs to the voting right delegation class, delegation relationship annotation data is obtained; The text paragraphs are semantically annotated according to the delegation relationship annotation data, and the relationship between the principal and the trustee is matched through the semantic rule library. If the matching degree is greater than the matching degree threshold, the voting right delegation result is obtained.

3. The method according to claim 1, characterized in that If there is a shareholder meeting resolution for voting rights delegation, the entrusting party and the entrusting party are identified, and the entrusting party's shareholding ratio is matched using a regular expression. A directed graph of voting rights delegation between shareholders is constructed based on shareholder nodes and delegation edges to form a topological structure of the delegation chain, including: Parse the voting rights delegation text according to the preset entity identification rules to obtain entity attribute data including the delegation party identifier and the delegation party identifier and the shareholding value; Establishing an attribute graph data model through the entity attribute data, wherein the attribute graph data model includes shareholder entity nodes and delegation relationship edges, and the delegation relationship edges store shareholding values ​​and delegation time information; Verify the delegation relationship edge between the shareholder entity nodes. If the entrusting party's shareholding ratio is higher than the shareholding ratio threshold and the entrustment period meets the rule requirements, establish a directed edge with a weight value in the attribute graph data model. A depth-first search algorithm is used to search for paths of the directed edges with weighted values, and a delegation chain topology sequence is obtained by accumulating the path weight values. The delegation chain topology sequence marks the actual ownership relationship of the voting rights.

4. The method according to claim 1, characterized in that: The traversal of the shareholder voting rights delegation chain, counting the out-degree and in-degree of each shareholder node, calculating the centrality index of each shareholder node according to the network topology analysis method, and judging whether the centrality index of each shareholder node is higher than the preset centrality threshold. If the shareholder node is higher than the preset centrality threshold, it is considered that the shareholder corresponding to the shareholder node higher than the preset centrality threshold is the actual controller hidden behind the delegation chain, including: A breadth-first search algorithm is used to obtain the delegated out-degree value and the entrusted in-degree value of the shareholder node, and the relative degree weight of the node is calculated according to the delegated out-degree value and the entrusted in-degree value; Calculating the connection weight value between the node and the adjacent nodes according to the relative degree weight, and obtaining the initial centrality value of the node by iterative calculation of the feature vector; The initial centrality value is normalized to the maximum and minimum values, and the control range value of the node is calculated by the path analysis method; The comprehensive score of the node is calculated according to the control range value. If the comprehensive score is higher than the preset centrality threshold, the node is determined to be the actual controller node.

5. The method according to claim 1, characterized in that The obtaining of historical voting rights delegation data of shareholder nodes above the preset centrality threshold, analyzing the position and role of the shareholder corresponding to the shareholder node above the preset centrality threshold in the delegation chain, and if the frequency of the shareholder appearing at the key node of the delegation chain is higher than the preset frequency threshold, and the connected delegation chain has a deep nesting layer, then determining that the shareholder is a hidden controller includes: Obtaining historical entrustment records of high-centrality shareholder nodes, wherein the high-centrality shareholder nodes are screened from shareholder nodes through a preset centrality threshold; Based on the historical entrustment records, the data validity verification rules are used to verify the entrustment share balance and time series continuity, and a standardized data set containing both entrusting parties and the entrustment quantity is obtained; A depth-first search algorithm is used to traverse the delegation chain for the standardized data set, and a node weighted frequency value is calculated by a hierarchical weight function, wherein the hierarchical weight function adopts a weighting method in which the first-level weight value is a preset value that increases layer by layer; If the weighted frequency value of a node is higher than the preset frequency threshold and the number of nested levels of the delegated chain exceeds the preset level threshold, the node will be marked as a hidden controller.

6. The method according to claim 1, characterized in that The method of obtaining the attribute characteristics of the shareholder node and using a graph neural network algorithm to predict the probability value of the actual control of each shareholder node based on the attribute characteristics of the shareholder node and the topological structure of the delegation chain, and if the probability value of the actual control of the shareholder node is higher than a preset probability value, then determining that the shareholder corresponding to the shareholder node is a hidden controller includes: For shareholder information, the type coding method is used to obtain the classification label values ​​of natural person shareholders and institutional shareholders, and the numerical feature vector of shareholding ratio is obtained through normalization processing; Constructing a graph neural network structure according to the numerical feature vector, wherein the graph neural network structure includes a feature aggregation layer, a spatial attention layer, and a node update layer; A graph convolution operator is used to perform aggregation operation on the numerical feature vector and the adjacency relationship, and a node embedding vector of fused structural information is obtained by calculating the attention coefficient; The cross entropy loss function value is calculated for the node embedding vector. If the loss function value is greater than the loss function threshold, the back propagation method is used to optimize and adjust the network parameters.

7. The method according to claim 1, characterized in that The above method calculates the voting weight of each shareholder based on the shareholding ratio and voting power delegation ratio of the shareholders in the delegation chain, traces the final shareholding ratio of the hidden controller, and determines the actual control of each shareholder over the target company, including: Obtain direct shareholding records and voting rights delegation records based on the shareholder register, standardize the shareholding ratio through normalization, and use the linear combination method to obtain the basic voting weight; Constructing a shareholding relationship diagram for the basic voting weights, stratifying the shareholding relationship diagram by a topological sorting method, and calculating the penetration shareholding ratio by a recursive traversal algorithm; A weight attenuation coefficient is set according to the penetration shareholding ratio, an independent control coefficient is obtained by calculating the weight attenuation coefficient, and the independent control coefficients are grouped and merged using the related party determination method to obtain group control power; A hierarchical clustering algorithm is used to group the control power of the groups, and the grouping results are cross-validated through historical voting records. The actual controllers of the high-control groups that pass the verification are marked.

8. The method according to claim 1, characterized in that The extraction of the association network centered on the hidden controller in the equity relationship network uses a graph neural network to process multiple types of nodes and edge relationships in the equity network, and quantifies the influence and connectivity of entities in the hidden controller's associated party network, including: Obtain the neighboring nodes of the hidden controller in the equity relationship knowledge graph, traverse the neighboring nodes by using the breadth-first search method, and obtain the associated party network including the shareholder entity, the company entity, and the executive entity; A heterogeneous graph neural network is constructed based on the related party network, and features of equity relationships, investment relationships and employment relationships are extracted through the entity embedding layer, relationship conversion layer and feature aggregation layer to obtain a multi-dimensional feature matrix; The degree centrality, betweenness centrality and closeness centrality of the nodes are calculated using the eigenvector method for the multidimensional feature matrix, and the nodes are divided into communities using a clustering method; According to the community division results, the random walk method is used to serialize the sampling of nodes, and the association probability between entity node pairs is calculated by the link prediction method to obtain the association strength matrix between entities.

Images