A method and system for identifying equity structure diagrams based on data processing

Abstract

This invention discloses a method and system for identifying equity structure diagrams based on data processing, relating to the field of data processing technology. The method includes: performing feature recognition on the equity structure diagram to obtain entity features and relationship features; constructing an identification model based on a graph neural network, using entity features as nodes and relationship features as edges; identifying nested relationships in the equity structure diagram through the identification model; wherein, nested relationships include relationships between equity entities; equity entities include shareholders and shareholders; and determining multi-level equity penetration paths, actual control paths, cross-shareholding relationships, and potential equity change relationships in the equity structure diagram based on the nested relationships.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method and system for identifying equity structure diagrams based on data processing. Background Technology

[0002] As companies grow larger and their operations become increasingly complex, equity structure diagrams, as tools for presenting key information such as equity distribution, shareholder relationships, and control, are becoming increasingly important. In many scenarios, such as mergers and acquisitions, investment analysis, and corporate governance research, the identification and analysis of equity structure diagrams have become fundamental and core elements.

[0003] Current methods for identifying equity structure diagrams have certain limitations. The main problem lies in their insufficient accuracy when dealing with complex nested relationships commonly found in equity structure diagrams. This can easily lead to misunderstandings of the actual relationships among equity holders, thereby affecting the reliability of decisions based on equity structure analysis.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method and system for identifying equity structure diagrams based on data processing to solve the above-mentioned technical problems.

[0006] This application provides a method for identifying equity structure diagrams based on data processing, including:

[0007] Feature recognition is performed on the equity structure diagram to obtain entity features and relationship features;

[0008] Using the entity features as nodes and the relationship features as edges, a recognition model based on a graph neural network is constructed.

[0009] The nesting relationships in the equity structure diagram are identified using the identification model; wherein, the nesting relationships include the relationships between equity entities; the equity entities include shareholders and shareholders.

[0010] Based on the nested relationships, the multi-level equity penetration paths, actual control paths, cross-shareholding relationships, and potential equity change relationships in the equity structure diagram are determined.

[0011] Furthermore, using the entity features as nodes and the relationship features as edges, a recognition model based on a graph neural network is constructed, including:

[0012] A two-layer adaptive modularity optimization algorithm is adopted to divide the holding community levels according to the preset shareholding ratio threshold. Among them, the node control feature vector is extracted by graph convolutional network, the control entropy difference between nodes is calculated, and a multi-level holding community tree with multiple edge weight constraints is generated.

[0013] A three-dimensional embedding system is constructed in the graph attention network, and node attributes, community level labels, and historical control trajectories are mapped to Euclidean space, hyperbolic space, and temporal space, respectively. Multidimensional association weights between nodes are calculated through a cross-space attention gating mechanism.

[0014] A three-level sampling strategy is implemented in the graph sampling and aggregation framework, including: for nodes within the holding community, deterministic sampling based on control entropy ranking is adopted to select nodes with control entropy higher than the community mean; for connecting edges across different holding communities, Monte Carlo sampling based on random walk is implemented, and the sampling weight of cross-community edges is dynamically adjusted through probability transition matrix; for node pairs in the circular shareholding structure, negative sampling of graph contrastive learning is implemented, and adversarial negative samples are generated by decoupling circular paths.

[0015] Furthermore, the dynamic optimization of the recognition model is achieved through the following mechanism:

[0016] A dynamic graph convolutional long short-term memory network is constructed. Time-varying control contribution tensors are generated through time slices. Bidirectional gradient propagation is performed along the holding community tree. Gated residual connections are used to capture the abrupt changes in control transfer.

[0017] We design a community stability discriminator based on spectral perturbation analysis, dynamically adjust the modularity of the objective function through a gated graph neural network, and correct the attention coefficient using the node influence propagation matrix to generate a hypergraph model containing the evolution trajectory of the control chain.

[0018] Furthermore, the relationships between equity entities include direct shareholding relationships and indirect shareholding relationships; the nested relationships in the equity structure diagram are identified through the identification model, including:

[0019] The nested relationship is decomposed into explicit direct shareholding links and implicit indirect shareholding links; wherein, the implicit indirect shareholding links are gradient-tracked through the hierarchical backtracking mechanism of the holding community tree to establish a shareholding ratio decay model.

[0020] In the identification model, a bidirectional control chain detection module is constructed. The forward chain aggregates shareholder behavior features along the time dimension, and the reverse chain locates potential control nodes through Monte Carlo tree search, generating a multi-level penetration path set with time-series labels.

[0021] Design a spatiotemporal dual attention evaluation module to apply composite weights to the nodes of the penetration path;

[0022] A structural similarity matrix is ​​constructed through graph comparison learning to handle cross-shareholding loops.

[0023] Furthermore, a structural similarity matrix is ​​constructed through graph comparison learning, and the following processing is applied to the cross-shareholding loop:

[0024] Graph isomorphic networks are used to detect the closure points of loop structures;

[0025] The equivalent control path of cyclical shareholding is separated by feature deentanglement;

[0026] Generate virtual control nodes and reconstruct them into a tree-like traversal path;

[0027] Output a topology-corrected graph containing nested ring relationships.

[0028] Furthermore, based on the nested relationship, the multi-level equity penetration path and actual control path in the equity structure diagram are determined, including:

[0029] Based on the virtual control nodes and physical nodes in the topology correction graph, at least two penetration paths at different levels are extracted through a hierarchical aggregation algorithm, and a weighted correlation channel is established between the paths.

[0030] In the three-dimensional embedding space, the hierarchical decay factor of hyperbolic space and the holding distance of Euclidean space are integrated to calculate the hierarchical control weight of nodes in each penetration path and classify the hierarchical affiliation of the penetration path.

[0031] The control entropy gradient is traced in reverse along the penetration path. When the difference in control effectiveness between adjacent levels exceeds the mutation threshold, the actual control node is determined and a minimum dominance path subgraph starting from that node is generated.

[0032] The control transfer trajectory in the hypergraph model is bound to the penetration path, the penetration depth and control strength confidence interval of each level are labeled, and the set of actual control paths containing level labels is output as a multi-level equity penetration path.

[0033] Furthermore, it further includes the dynamic verification and correction of the aforementioned multi-level equity penetration path:

[0034] By using a dynamic graph convolutional long short-term memory network, the temporal behavior differences of nodes at different levels in the penetration path are captured, and cross-level control transfer anomalies are identified.

[0035] A community stability discriminator is used to verify whether the penetration path runs through multiple holding community levels. If the jump between levels causes a break in control effectiveness, it is marked as a path to be repaired.

[0036] By using a control transfer probability model, we can detect whether there are hierarchical gaps in the penetration path in historical time series and fill in the broken path segments.

[0037] When an abnormal decay of control effectiveness is detected at a specific level, the shareholding ratio threshold is dynamically adjusted and the controlling community tree is reconstructed to generate an actual control path that conforms to the multi-level penetration logic.

[0038] Based on the spectral perturbation analysis results, the eigenvalue sensitivity and anti-interference capability of each level path are calculated, and a set of reference penetration paths with hierarchical stability labels is output; wherein, the set of reference penetration paths is used to correct the multi-level equity penetration paths.

[0039] Furthermore, based on the nested relationship, the cross-shareholding relationship and potential equity change relationship in the equity structure diagram are determined, including:

[0040] In the equivalent control path after feature deentanglement, the topology structure that has a bidirectional shareholding link and forms a closed loop is detected and marked as a candidate cross-shareholding loop.

[0041] By controlling the entropy difference model, the mutual control coefficients between nodes in the loop are quantified. If the closed-loop control effectiveness exceeds the shareholding ratio threshold, it is determined to be a cross-shareholding relationship.

[0042] Extract the third derivative mutation points of shareholding ratio in the historical control trajectory, and combine them with the control transfer probability output by the gated residual network to locate high volatility nodes as potential triggers for equity changes.

[0043] A model for the propagation of the influence of equity changes is constructed to simulate the impact of the transfer of control at the triggering source node on the global path. The output heat map with probability distribution serves as a visualization of cross-shareholding relationships and potential equity change relationships.

[0044] Furthermore, the curvature contraction rate of the cross-shareholding loop is calculated in hyperbolic space, and the stability fluctuation period of the loop is detected in temporal space to generate the spatiotemporal evolution feature vector of the loop.

[0045] The eigenvalue sensitivity parameters of the loop structure are obtained by spectral perturbation analysis, and the probability of loop disintegration and the range of chain reaction are calculated by combining the node influence propagation matrix.

[0046] We use adversarial generative networks to construct various equity restructuring schemes and evaluate the impact of different schemes on the cross-shareholding loop while maintaining the core structure of the actual control path.

[0047] By integrating the loop stability index, change heatmap, and simulation results, a decision support graph is generated that marks the optimal restructuring path and risk warning areas; wherein, the decision support graph is used to verify and optimize the determination of cross-shareholding and change relationship.

[0048] This application provides a data processing-based equity structure diagram recognition system, including:

[0049] The feature recognition module is used to identify features in the equity structure diagram to obtain entity features and relationship features;

[0050] The recognition model construction module is used to construct a recognition model based on a graph neural network by using the entity features as nodes and the relationship features as edges.

[0051] The nesting relationship identification module is used to identify the nesting relationships in the equity structure diagram through the identification model; wherein, the nesting relationship includes the relationship between equity entities; the equity entity includes the shareholder and the shareholder;

[0052] The equity structure diagram recognition module is used to determine the multi-level equity penetration path, actual control path, cross-shareholding relationship and potential equity change relationship in the equity structure diagram based on the nested relationship.

[0053] Based on the embodiments provided in this application, through feature recognition and recognition network models, complex nested relationships in equity structure diagrams can be identified more accurately, reducing misunderstandings of actual equity relationships and thus improving the reliability of decisions based on equity structure analysis. By accurately identifying multi-level equity penetration paths, actual control paths, cross-shareholding relationships, and potential equity change relationships in equity structure diagrams, more reliable data support can be provided for corporate mergers and acquisitions, investment analysis, corporate governance research, and other decisions, reducing decision-making risks. It can handle common complex nested relationships in equity structure diagrams, providing a powerful tool for in-depth analysis of corporate equity structure and helping to better understand the distribution of control and shareholder relationships within the company. Attached Figure Description

[0054] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0055] Figure 1 A flowchart illustrating an optional data processing-based equity structure diagram identification method according to an embodiment of this application;

[0056] Figure 2 A flowchart illustrating an optional data processing-based equity structure diagram identification method according to an embodiment of this application;

[0057] Figure 3 This is a structural diagram of an optional data processing-based equity structure diagram recognition system according to an embodiment of this application.

[0058] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0060] Optionally, such as Figure 1 As shown, this application provides a method for identifying equity structure diagrams based on data processing, including:

[0061] S101, Perform feature recognition on the equity structure diagram to obtain entity features and relationship features;

[0062] In this embodiment, entity characteristics refer to the inherent attributes and information of each entity (such as a company, shareholder, etc.) in the equity structure diagram. These characteristics are used to uniquely identify and describe each entity in the equity structure.

[0063] Feature recognition methods include: Optical Character Recognition (OCR): extracting text information from equity structure diagrams to identify entity information such as company names and shareholder names; Intelligent Text Detection Algorithm: combining a predefined equity structure keyword library to accurately locate and extract key entity information from the diagram; Natural Language Processing (NLP): for text descriptions containing entity information, NLP technology is used for entity recognition to extract entities such as companies and shareholders.

[0064] S102, using entity features as nodes and relationship features as edges, constructs a recognition model based on graph neural networks;

[0065] In this embodiment, the relationship features include equity relationship features and equity value features. The equity value features represent the shareholding ratio or number of shares held by the shareholder in the shareholding entity. By analyzing the equity value features on the multi-level equity penetration path, the shareholding ratio or control strength of the shareholder in the shareholding entity on the actual control path can be determined.

[0066] Components of Relationship Features: Equity Relationship Features: These indicate whether an equity relationship exists between a shareholder and a shareholder, i.e., whether there is direct or indirect shareholding behavior. This can be represented by lines or edges in the equity structure diagram; lines indicate the existence of an equity relationship. Equity Value Features: These further refine the equity relationship features, specifically specifying the shareholder's shareholding ratio or quantity in the shareholder's company. For example, a shareholding ratio of 30% or a shareholding quantity of 1 million shares quantifies the strength and specificity of the equity relationship. Analysis of Equity Value Features on Paths. Multi-Level Equity Penetration Paths: In complex equity structures, a shareholder may indirectly hold equity in a target company through multiple intermediate levels of companies. The path represents the complete equity link from the top-level shareholder to the bottom-level shareholder, including the shareholding relationship and equity value at each level.

[0067] Calculating the equity ratio or control strength along the actual control path: By analyzing the equity value characteristics along the path, the actual control ratio or control strength of the shareholder over the investee can be calculated. This helps identify the actual controlling party and key control points in the equity structure.

[0068] Assume the following shareholding structure: Shareholder A directly holds 40% of the shares in Company B. Company B holds 50% of the shares in Company C. Company C holds 60% of the shares in Company D.

[0069] Therefore, the multi-level equity penetration path from shareholder A to company D is A→B→C→D. By analyzing the equity value characteristics along this path, we can calculate that shareholder A's actual control over D is 40%×50%×60%=12%. This means that shareholder A indirectly holds 12% of the equity in company D through B and C, thus enabling us to assess its actual influence and degree of control over company D.

[0070] S103, Identify the nested relationships in the equity structure diagram through the identification model; wherein, the nested relationship includes the relationship between equity entities; equity entities include shareholders and shareholders.

[0071] S104. Based on the nested relationships, determine the multi-level equity penetration path, actual control path, cross-shareholding relationship and potential equity change relationship in the equity structure diagram.

[0072] In this embodiment of the application, when identifying nested relationships, the aggregation and update mechanism of graph neural networks is used to pass the information of nodes and edges layer by layer, thereby identifying complex equity structure relationships.

[0073] Furthermore, such as Figure 2 As shown, a recognition model based on a graph neural network is constructed by using entity features as nodes and relationship features as edges, including:

[0074] S201 employs a two-layer adaptive modularity optimization algorithm to divide the holding community levels according to a preset shareholding ratio threshold. Specifically, it extracts node control feature vectors through a graph convolutional network, calculates the control entropy difference between nodes, and generates a multi-level holding community tree with multiple edge weight constraints.

[0075] S202 constructs a three-dimensional embedding system in graph attention network, mapping node attributes, community level labels, and historical control trajectories to Euclidean space, hyperbolic space, and temporal space respectively, and calculates multidimensional association weights between nodes through a cross-space attention gating mechanism.

[0076] S203 implements a three-level sampling strategy within the graph sampling and aggregation framework.

[0077] Among them, GraphAttentionNetworks (GAT) is the graph attention network, and GraphSAGE is the graph sampling and aggregation framework.

[0078] In this embodiment of the application, a three-level sampling strategy is implemented in the graph sampling and aggregation framework, including:

[0079] For nodes within the holding community, deterministic sampling based on control entropy ranking is adopted to select nodes whose control entropy is higher than the community average.

[0080] For the connecting edges that cross different holding communities, Monte Carlo sampling based on random walk is implemented, and the sampling weights of the cross-community edges are dynamically adjusted through the probability transition matrix;

[0081] For node pairs in the circular shareholding structure, negative sampling is performed using graph contrastive learning to generate adversarial negative samples by decoupling the circular paths.

[0082] In this embodiment of the application, the modularity optimization function of the holding community is:

[0083]

[0084] in, The modularity score for the holding community measures the tightness of the control relationships between nodes within the community;

[0085] The set of holding communities partitioned using a two-layer adaptive modularity optimization algorithm; An index for a single holding community; For connecting nodes and A directed edge represents the edge from the shareholder. To the shareholding party Equity relationship; , For node indexes in the equity structure diagram; The first in the equity structure diagram Each node represents a shareholder entity; The first in the equity structure diagram Each node represents a shareholder entity; For the community The set of all internal edges; Community The set of all internal nodes; For the edge The control weight is calculated as follows: ,in, and This is the node control feature vector extracted by the convolutional network in the Chinese Library Classroom;

[0086] The sum of global edge weights. ; For nodes The control entropy value is calculated using the following formula: ;

[0087] For nodes For the The control probability of each community; This is an adaptive adjustment coefficient for the critical value of shareholding ratio. ; To control the smoothing parameter of entropy differences, the default value is the standard deviation of the feature vectors of nodes within the community.

[0088] In this embodiment of the application, the cross-spatial attention weight is calculated based on the following formula:

[0089]

[0090]

[0091] in, For nodes right Cross-spatial attention weights; Represents a node The attribute vector in Euclidean space includes static features such as shareholding ratio and registered capital; Represents a node The original hierarchical vectors in hyperbolic space are generated by mapping through the Poincaré sphere model; Represents a node The control trajectory vector in the time-series space records the changes in control intensity in the historical time slices; , , All are trainable parameter matrices, with dimensions respectively. , , ; The hierarchical attenuation coefficient is... , This represents the maximum layer depth of the penetration path; For nodes In the depth of the holding community tree, the root node has a depth of 0, and the depth increases layer by layer; This indicates element-wise multiplication, used to fuse vectors from different spaces to calculate the association weights between nodes; Represents a node The representation of the original hierarchical vectors in hyperbolic space after normalization transformation. This normalization transformation is to ensure that vectors in different spaces can be effectively associated with each other within a unified framework.

[0092] Based on the embodiments provided in this application, the two-layer adaptive modularity optimization algorithm, according to a preset shareholding ratio threshold, can accurately divide different levels of holding communities in the equity structure and generate a multi-level holding community tree with multiple edge weight constraints. This helps to more clearly understand the hierarchical relationships and control distribution in the corporate equity structure, providing a powerful tool for in-depth analysis of the internal control architecture of the enterprise. The three-dimensional embedding system constructed in the graph attention network maps node attributes, community level labels, and historical control trajectories to Euclidean space, hyperbolic space, and temporal space, respectively. This cross-space embedding method can more comprehensively capture the multi-dimensional features of nodes, calculate the multi-dimensional association weights between nodes through a cross-space attention gating mechanism, thereby more accurately identifying complex relationships between nodes and improving the model's ability to identify complex nested relationships in the equity structure graph. The three-level sampling strategy implemented in the graph sampling and aggregation framework adopts different sampling methods for different scenarios. For nodes within a holding community, deterministic sampling based on control entropy ranking selects nodes with control entropy higher than the community average, focusing on key nodes. For connections spanning different holding communities, Monte Carlo sampling based on random walks dynamically adjusts sampling weights through a probability transition matrix, flexibly acquiring key information about cross-community edges. For node pairs in a circular shareholding structure, negative sampling using graph contrastive learning generates adversarial negative samples by decoupling circular paths, effectively handling the complex relationship of circular shareholding. This three-level sampling strategy efficiently samples key information, improving the model's computational efficiency and accuracy.

[0093] Furthermore, the recognition model is dynamically optimized through the following mechanism:

[0094] A dynamic graph convolutional long short-term memory network is constructed. Time-varying control contribution tensors are generated through time slices, and bidirectional gradient propagation is performed along the holding community tree. Gated residual connections are used to capture the abrupt features of control transfer. The full name of the dynamic graph convolutional long short-term memory network is DirectedGraphConvolutionalNetwork-LongShort-TermMemory, abbreviated as DGCN-LSTM.

[0095] We design a community stability discriminator based on spectral perturbation analysis, dynamically adjust the modularity of the objective function through a gated graph neural network, and correct the attention coefficient using the node influence propagation matrix to generate a hypergraph model containing the evolution trajectory of the control chain.

[0096] Furthermore, the relationships between equity entities include direct and indirect shareholding relationships; through the identification model, the nested relationships in the equity structure diagram are identified, including:

[0097] The nested relationship is decomposed into explicit direct shareholding links and implicit indirect shareholding links; among them, the implicit indirect shareholding links are gradient-tracked through the hierarchical backtracking mechanism of the holding community tree to establish a shareholding ratio decay model.

[0098] A bidirectional control chain detection module is constructed in the identification model. The forward chain aggregates shareholder behavior features along the time dimension, and the reverse chain locates potential control nodes through Monte Carlo tree search, generating a multi-level penetration path set with time-series labels.

[0099] Design a spatiotemporal dual attention evaluation module to apply composite weights to the nodes of the penetration path;

[0100] A structural similarity matrix is ​​constructed through graph comparison learning to handle cross-shareholding loops.

[0101] Furthermore, a structural similarity matrix is ​​constructed through graph comparison learning, and the following processing is applied to the cross-shareholding loop:

[0102] Graph isomorphic networks are used to detect the closure points of loop structures;

[0103] The equivalent control path of cyclical shareholding is separated by feature deentanglement;

[0104] Generate virtual control nodes and reconstruct them into a tree-like traversal path;

[0105] Output a topology-corrected graph containing nested ring relationships.

[0106] In this embodiment, the closed loop of the circular shareholding system is detected based on the following formula:

[0107]

[0108] in, For nodes As a score for the loop closure point, a higher score indicates a greater probability of closure. The parameter matrix of the graph isomorphic network is generated through unsupervised contrastive learning pre-training. For nodes The set of predecessor nodes in the loop path is located by Monte Carlo tree search. The cross-space attention normalization weights are taken from the formula. The output;

[0109] This represents the number of graph convolutional layers, consistent with the depth of the graph convolutional network, and is the default value. It is 3; It is an infinite norm, used to capture the maximum deviation; This indicates element-wise multiplication, meaning that corresponding elements of two vectors are multiplied to obtain a new vector.

[0110] Furthermore, based on the nested relationships, the multi-level equity penetration paths and actual control paths in the equity structure diagram are determined, including:

[0111] Based on the virtual control nodes and entity nodes in the topology correction graph, at least two penetration paths at different levels are extracted through a hierarchical aggregation algorithm, and a weighted correlation channel is established between the paths.

[0112] In the three-dimensional embedding space, the hierarchical decay factor of hyperbolic space and the holding distance of Euclidean space are integrated to calculate the hierarchical control weight of nodes in each penetration path and classify the hierarchical affiliation of the penetration path.

[0113] The control entropy gradient is traced in reverse along the penetration path. When the difference in control effectiveness between adjacent levels exceeds the mutation threshold, the actual control node is determined and a minimum dominance path subgraph starting from that node is generated.

[0114] The control transfer trajectory in the hypergraph model is bound to the penetration path, the penetration depth and control strength confidence interval of each level are labeled, and the set of actual control paths containing level labels is output as a multi-level equity penetration path.

[0115] Furthermore, this includes the dynamic verification and correction of multi-level equity penetration paths:

[0116] By using a dynamic graph convolutional long short-term memory network, the temporal behavior differences of nodes at different levels in the penetration path are captured, and cross-level control transfer anomalies are identified.

[0117] A community stability discriminator is used to verify whether the penetration path runs through multiple holding community levels. If the jump between levels causes a break in control effectiveness, it is marked as a path to be repaired.

[0118] By using a control transfer probability model, we can detect whether there are hierarchical gaps in the penetration path in historical time series and fill in the broken path segments.

[0119] When an abnormal decay of control effectiveness is detected at a specific level, the shareholding ratio threshold is dynamically adjusted and the controlling community tree is reconstructed to generate an actual control path that conforms to the multi-level penetration logic.

[0120] Based on the spectral perturbation analysis results, the eigenvalue sensitivity and anti-interference capability of each level path are calculated, and a set of reference penetration paths with hierarchical stability labels is output; among them, the set of reference penetration paths is used to correct multi-level equity penetration paths.

[0121] Furthermore, based on the nested relationships, the cross-shareholding relationships and potential equity change relationships in the equity structure diagram are determined, including:

[0122] In the equivalent control path after feature deentanglement, the topology structure that has a bidirectional shareholding link and forms a closed loop is detected and marked as a candidate cross-shareholding loop.

[0123] By controlling the entropy difference model, the mutual control coefficients between nodes in the loop are quantified. If the closed-loop control effectiveness exceeds the shareholding ratio threshold, it is determined to be a cross-shareholding relationship.

[0124] Extract the third derivative mutation points of shareholding ratio in the historical control trajectory, and combine them with the control transfer probability output by the gated residual network to locate high volatility nodes as potential triggers for equity changes.

[0125] A model for the propagation of the influence of equity changes is constructed to simulate the impact of the transfer of control at the triggering source node on the global path. The output heat map with probability distribution serves as a visualization of cross-shareholding relationships and potential equity change relationships.

[0126] Furthermore, the curvature contraction rate of the cross-shareholding loop is calculated in hyperbolic space, and the stability fluctuation period of the loop is detected in temporal space to generate the spatiotemporal evolution feature vector of the loop.

[0127] The eigenvalue sensitivity parameters of the loop structure are obtained by spectral perturbation analysis, and the probability of loop disintegration and the range of chain reaction are calculated by combining the node influence propagation matrix.

[0128] We use adversarial generative networks to construct various equity restructuring schemes and evaluate the impact of different schemes on the cross-shareholding loop while maintaining the core structure of the actual control path.

[0129] By integrating the loop stability index, change heatmap, and simulation results, a decision support map is generated that marks the optimal restructuring path and risk warning areas. The decision support map is used to verify and optimize the determination of cross-shareholding and change relationships.

[0130] Optionally, such as Figure 3 As shown, this application provides a data processing-based equity structure diagram recognition system, including:

[0131] The feature recognition module 301 is used to perform feature recognition on the equity structure diagram to obtain entity features and relationship features;

[0132] The recognition model building module 302 is used to construct a recognition model based on a graph neural network by using entity features as nodes and relationship features as edges.

[0133] Nested relationship identification module 303 is used to identify nested relationships in the equity structure diagram through an identification model; wherein, nested relationships include relationships between equity entities; equity entities include shareholders and shareholders.

[0134] The equity structure diagram recognition module 304 is used to determine the multi-level equity penetration path, actual control path, cross-shareholding relationship and potential equity change relationship in the equity structure diagram based on the nesting relationship.

[0135] It should be noted that the embodiments implemented by the equity structure diagram recognition system based on data processing in this application can be referenced to the embodiments implemented by the equity structure diagram recognition method based on data processing, and will not be described in detail here.

[0136] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for identifying equity structure diagrams based on data processing, characterized in that, include: Feature recognition is performed on the equity structure diagram to obtain entity features and relationship features; Using the entity features as nodes and the relationship features as edges, a recognition model based on a graph neural network is constructed. Constructing a recognition model based on graph neural networks includes: A two-layer adaptive modularity optimization algorithm is adopted to divide the holding community levels according to the preset shareholding ratio threshold. Among them, the node control feature vector is extracted by graph convolutional network, the control entropy difference between nodes is calculated, and a multi-level holding community tree with multiple edge weight constraints is generated. A three-dimensional embedding system is constructed in the graph attention network, and node attributes, community level labels, and historical control trajectories are mapped to Euclidean space, hyperbolic space, and temporal space, respectively. Multidimensional association weights between nodes are calculated through a cross-space attention gating mechanism. A three-level sampling strategy is implemented in the graph sampling and aggregation framework, including: for nodes within the holding community, deterministic sampling based on control entropy ranking is adopted to select nodes with control entropy higher than the community mean; for connecting edges across different holding communities, Monte Carlo sampling based on random walk is implemented, and the sampling weight of cross-community edges is dynamically adjusted through probability transition matrix; for node pairs in the circular shareholding structure, negative sampling of graph contrastive learning is implemented, and adversarial negative samples are generated by decoupling circular paths. To achieve dynamic optimization of the recognition model, including: A dynamic graph convolutional long short-term memory network is constructed. Time-varying control contribution tensors are generated through time slices. Bidirectional gradient propagation is performed along the holding community tree. Gated residual connections are used to capture the abrupt changes in control transfer. Design a community stability discriminator based on spectral perturbation analysis, dynamically adjust the modularity to optimize the objective function through a gated graph neural network, and use the node influence propagation matrix to correct the attention coefficient, thereby generating a hypergraph model containing the evolution trajectory of the control chain; The nesting relationships in the equity structure diagram are identified using the identification model; wherein, the nesting relationships include the relationships between equity entities; the equity entities include shareholders and shareholders. Identify nested relationships in the equity structure diagram, including: The nested relationship is decomposed into explicit direct shareholding links and implicit indirect shareholding links; wherein, the implicit indirect shareholding links are gradient-tracked through the hierarchical backtracking mechanism of the holding community tree to establish a shareholding ratio decay model. In the identification model, a bidirectional control chain detection module is constructed. The forward chain aggregates shareholder behavior features along the time dimension, and the reverse chain locates potential control nodes through Monte Carlo tree search, generating a multi-level penetration path set with time-series labels. Design a spatiotemporal dual attention evaluation module to apply composite weights to the nodes of the penetration path; A structural similarity matrix is ​​constructed through graph comparison learning to handle cross-shareholding loops; Based on the nested relationships, the multi-level equity penetration paths, actual control paths, cross-shareholding relationships, and potential equity change relationships in the equity structure diagram are determined.

2. The equity structure diagram recognition method based on data processing according to claim 1, characterized in that, By constructing a structural similarity matrix through graph comparison learning, the following processing is applied to cross-shareholding loops: Graph isomorphic networks are used to detect the closure points of loop structures; The equivalent control path of cyclical shareholding is separated by feature deentanglement; Generate virtual control nodes and reconstruct them into a tree-like traversal path; Output a topology-corrected graph containing nested ring relationships.

3. The equity structure diagram recognition method based on data processing according to claim 2, characterized in that, Based on the nested relationship, the multi-level equity penetration path and actual control path in the equity structure diagram are determined, including: Based on the virtual control nodes and physical nodes in the topology correction graph, at least two penetration paths at different levels are extracted through a hierarchical aggregation algorithm, and a weighted correlation channel is established between the paths. In the three-dimensional embedding space, the hierarchical decay factor of hyperbolic space and the holding distance of Euclidean space are integrated to calculate the hierarchical control weight of nodes in each penetration path and classify the hierarchical affiliation of the penetration path. The control entropy gradient is traced in reverse along the penetration path. When the difference in control effectiveness between adjacent levels exceeds the mutation threshold, the actual control node is determined and a minimum dominance path subgraph starting from that node is generated. The control transfer trajectory in the hypergraph model is bound to the penetration path, the penetration depth and control strength confidence interval of each level are labeled, and the set of actual control paths containing level labels is output as a multi-level equity penetration path.

4. The equity structure diagram recognition method based on data processing according to claim 3, characterized in that, This further includes the dynamic verification and correction of the aforementioned multi-level equity penetration path: By using a dynamic graph convolutional long short-term memory network, the temporal behavior differences of nodes at different levels in the penetration path are captured, and cross-level control transfer anomalies are identified. A community stability discriminator is used to verify whether the penetration path runs through multiple holding community levels. If the jump between levels causes a break in control effectiveness, it is marked as a path to be repaired. By using a control transfer probability model, we can detect whether there are hierarchical gaps in the penetration path in historical time series and fill in the broken path segments. When an abnormal decay of control effectiveness is detected at a specific level, the shareholding ratio threshold is dynamically adjusted and the controlling community tree is reconstructed to generate an actual control path that conforms to the multi-level penetration logic. Based on the spectral perturbation analysis results, the eigenvalue sensitivity and anti-interference capability of each level path are calculated, and a set of reference penetration paths with hierarchical stability labels is output; wherein, the set of reference penetration paths is used to correct the multi-level equity penetration paths.

5. The equity structure diagram recognition method based on data processing according to claim 2, characterized in that, Based on the nested relationships, the cross-shareholding relationships and potential equity change relationships in the equity structure diagram are determined, including: In the equivalent control path after feature deentanglement, the topology structure that has a bidirectional shareholding link and forms a closed loop is detected and marked as a candidate cross-shareholding loop. By controlling the entropy difference model, the mutual control coefficients between nodes in the loop are quantified. If the closed-loop control effectiveness exceeds the shareholding ratio threshold, it is determined to be a cross-shareholding relationship. Extract the third derivative mutation points of shareholding ratio in the historical control trajectory, and combine them with the control transfer probability output by the gated residual network to locate high volatility nodes as potential triggers for equity changes. A model for the propagation of the influence of equity changes is constructed to simulate the impact of the transfer of control at the triggering source node on the global path. The output heat map with probability distribution serves as a visualization of cross-shareholding relationships and potential equity change relationships.

6. The equity structure diagram recognition method based on data processing according to claim 5, characterized in that, The curvature contraction rate of the cross-shareholding loop is calculated in hyperbolic space, and the stability fluctuation period of the loop is detected in temporal space to generate the spatiotemporal evolution feature vector of the loop. The eigenvalue sensitivity parameters of the loop structure are obtained by spectral perturbation analysis, and the probability of loop disintegration and the range of chain reaction are calculated by combining the node influence propagation matrix. We use adversarial generative networks to construct various equity restructuring schemes and evaluate the impact of different schemes on the cross-shareholding loop while maintaining the core structure of the actual control path. By integrating the loop stability index, change heatmap, and simulation results, a decision support graph is generated that marks the optimal restructuring path and risk warning areas; wherein, the decision support graph is used to verify and optimize the determination of cross-shareholding and change relationship.

7. A shareholding structure diagram recognition system based on data processing, the system implementing the method as described in claim 1, characterized in that, include: The feature recognition module is used to identify features in the equity structure diagram to obtain entity features and relationship features; The recognition model construction module is used to construct a recognition model based on a graph neural network by using the entity features as nodes and the relationship features as edges. The nesting relationship identification module is used to identify the nesting relationships in the equity structure diagram through the identification model; wherein, the nesting relationship includes the relationship between equity entities; the equity entity includes the shareholder and the shareholder; The equity structure diagram recognition module is used to determine the multi-level equity penetration path, actual control path, cross-shareholding relationship and potential equity change relationship in the equity structure diagram based on the nested relationship.

Images