Actual controller mining method and device based on stock right knowledge graph learning

Abstract

The invention discloses an actual controller mining method and device based on stock right knowledge graph learning. An actual control person is defined as a shareholder in a winning stable alliance, and members in the alliance make consistent decisions in the voting process and are controlled. According to the method, a plurality of complex relationships among shareholders in the equity network are organized and represented by constructing an equity knowledge graph. In order to mine a winning stable alliance, a multi-dimensional association aggregation neural network is designed, and the possibility of consistent action is predicted according to information of different types of relationships among shareholders. And setting a meta-path aggregation neural network based on the aggregation information, and encoding an equity structure by vertically capturing an equity holding path in the equity knowledge graph. In order to determine a winning stable alliance, the method proposes to control a neural network to simulate a voting process of shareholders. According to the method, the problems of low accuracy and high time cost in actual control person mining caused by a complex multi-dimensional relationship in the equity network can be effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of equity structure analysis, and specifically relates to a method and device for mining actual controllers based on equity knowledge graph learning. Background Art

[0002] There are millions of complex relationships between companies and their shareholders, forming a huge equity network. In this network, shareholders can not only control the company through the stocks they hold, but also use the relationships with other shareholders to expand their control. The actual controller refers to the shareholder who holds a sufficient number of voting shares in the company and can decide major decisions of the company. The actual controller is usually hidden in a multi-layered and complex equity structure. The key to identifying the actual controller is to discover the winning and stable alliance that can control the company's decision-making. Control calculation and identification of actual controlling shareholders are the core of financial risk governance and are of great significance to financial supervision.

[0003] Existing research calculates control rights through shareholding paths between shareholders. The weakest link model calculates control rights by summing up the minimum shareholding ratios of different shareholding paths between shareholders and the company; the matrix merger method assumes that control can be achieved as long as the company's voting weight exceeds the set threshold; the voting game method uses different indicators in game theory to simulate the voting process in the shareholder decision-making process. For example, Levy et al. measured shareholder control rights by analyzing all voting results and calculating the expected value of shareholders and companies voting in unison. Karos and Peters proposed the Shapley-Shubik power index, which takes indirect control into consideration. However, these methods ignore the diverse and complex relationships between shareholders and cannot accurately identify winning stable alliances, resulting in limited efficient mining of actual controllers. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art that only equity relations are considered to mine actual controllers, and the complex interactions between shareholders and different voting strategies during mining affect the accuracy of stable alliance identification, the present invention provides an actual controller mining method and device based on equity knowledge graph learning, so as to improve the accuracy and efficiency of actual controller mining.

[0005] According to one aspect of the present invention, a method for mining actual controllers based on equity knowledge graph learning is provided, comprising:

[0006] Step 1: Construct an equity knowledge graph based on the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the equity knowledge graph through entity alignment;

[0007] Step 2: In the equity knowledge graph described in step 1, a multi-dimensional association aggregation neural network RA is designed to automatically extract various relationship features between shareholders and discover stable alliances;

[0008] Step 3: Based on the multi-dimensional association aggregation neural network RA aggregating the equity structure information in step 2, a meta-path aggregation neural network PA is designed to capture the shareholder's shareholding path, and the voting rights information is retained according to the probability of the shareholder and the company co-occurring in the final captured path;

[0009] Step 4: Based on the embeddings generated by the multi-dimensional association aggregation neural network RA and the meta-path aggregation neural network PA in steps 2 and 3 above, a control neural network CoNN is designed to simulate the voting process by extracting the information of stock weights and stable alliances preserved in the entity embeddings to predict the control relationship between shareholders and companies.

[0010] As a further technical solution, step 1 includes:

[0011] Building a knowledge graph of equity Each triple is a de novo entity e i To the end entity e j There is a relationship r between them, ε is the entity set, e i ∈ε represents shareholders or companies, and each entity e i With a set of attributes {a i1 , ..., a im}Associated, It is a relationship set, including forward relationships and reverse relationships.

[0012] As a further technical solution, the multi-dimensional association aggregation neural network RA automatically extracts various relationship features between shareholders and discovers stable alliances, including: aggregating information from multi-hop neighbors to encode each relationship, capturing the mutual influence of different types of relationships through self-attention entity embedding, so as to capture the transitivity between relationships.

[0013] As a further technical solution, the multi-dimensional association aggregation neural network RA automatically extracts various relationship features between shareholders and finds stable alliances, further comprising:

[0014] Using the attributes of entities to aggregate the entities of multi-dimensional association neural network RA i Embedded i Initialize and calculate as follows:

[0015] u i =h o ( a i1, a i2 , ..., a im )

[0016] where h o is to set the attribute {a i1 , a i2 , ..., a im} is converted to an embedding function, o is the entity type; for each relationship Layer l Embedding The calculation is as follows:

[0017]

[0018] in N r (e i )={e j |(e i ,r,e j )∈G} represents e i Neighbors under relation r, l is a hyperparameter representing the number of hops and 0 <l<L;

[0019] Entity i The L-th layer embedding of different relations is expressed as:

[0020]

[0021] The self-attention mechanism is used to measure the mutual influence between different relations, and the entity e i The final embedding u on relation r ir The calculation method is as follows:

[0022]

[0023] Among them, M r is a trainable transformation matrix, α ir The coefficients representing the linear combination of the relations r;

[0024] A group of shareholders g The stability of the coalition composed of ∈εp a The calculation method of (g) is as follows:

[0025]

[0026] in, is the probability of consistent action, calculated by cosine similarity.

[0027] As a further technical solution, step 3 designs a meta-path aggregation neural network PA to traverse according to two predefined meta-path patterns, including:

[0028] Set meta path mode to capture the shareholder’s shareholding path, where l represents the pattern length, Set meta path mode to capture the shareholding paths of members in a stable coalition;

[0029] The equity path is captured by a random walk, where the i To e j The transition probability p(e i |e j ) is e i The normalized probability of neighbors, based on the above meta-path pattern, the transition probability p(e i |e j , t) is:

[0030]

[0031] According to the meta-path mode, r t Represents r a or s , is the corresponding weight of the walking edge in the tth step. Based on the transition probability, the meta-path aggregation neural network PA captures the shareholding path through graph traversal. According to the voting weight information of each entity pair, the frequency of co-appearance in the paths they generate is calculated and retained;

[0032] After generating the equity paths, the meta-path aggregation neural network PA learns entity embeddings by running the skip-gram model optimization.

[0033] As a further technical solution, for each pair of entities (e i , e j ), the embedding is calculated as follows:

[0034]

[0035] where v i Yes i The embedding of j Yes j The context embedding in the entity sequence, σ is the sigmoid function, calculated as σ(x) = 1 / (exp(x) + 1), N′(e i ) is e i The negative sampling set, N′(e i ) i is the noise distribution P(e i ) is a sample randomly selected from .

[0036] As a further technical solution, step 4 of designing the control neural network CoNN includes: on the one hand, by using the embedded v i and v j to assess the similarity of shareholders iThrough the shareholding path, the company j On the other hand, by changing v i The information of different types of stable alliances is extracted by transferring them to different embedding spaces, and then the final voting results of each stable alliance are analyzed to evaluate the shareholder e i Through stable winning alliances, the company j control.

[0037] According to one aspect of the present invention, a device for mining actual controllers based on equity knowledge graph learning is provided, comprising:

[0038] The first calculation module is used to construct a shareholding knowledge graph based on the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the shareholding knowledge graph through entity alignment;

[0039] The second computing module is used to design a multi-dimensional association aggregation neural network RA in the equity knowledge graph to automatically extract various relationship features between shareholders and discover stable alliances;

[0040] The third calculation module is used to design a meta-path aggregation neural network PA to capture the shareholder's shareholding path based on the multi-dimensional association aggregation neural network RA aggregating the equity structure information, and retain the voting rights information according to the probability of the shareholder and the company co-appearing in the finally captured path;

[0041] The fourth computing module is used to design a control neural network CoNN based on the embeddings generated by the multi-dimensional association aggregation neural network RA and the meta-path aggregation neural network PA. It simulates the voting process by extracting the information of stock weights and stable alliances saved in the entity embeddings to predict the control relationship between shareholders and companies.

[0042] According to one aspect of the present invention specification, there is provided a system for mining actual controllers based on equity knowledge graph learning, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the method for mining actual controllers based on equity knowledge graph learning.

[0043] According to one aspect of the present specification, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions enable the computer to execute the steps of a method for mining actual controllers based on equity knowledge graph learning.

[0044] The present invention captures the semantic and structural information in large-scale equity networks based on the characteristics of winning stable alliances by modeling the complex relationships between shareholders, and unearths hidden actual controllers. Compared with the prior art, the present invention has the following beneficial effects:

[0045] First, the actual controller mining problem is modeled as a control right prediction task for the first time, and it is proposed to uniformly model the complex relationships between shareholders as an equity knowledge graph to provide a unified representation for the rich semantic and structural information of the equity network.

[0046] Second, based on the characteristics of stable alliances, a learning method for equity knowledge graphs is proposed. A multi-relational aggregation neural network (RA) is designed to measure the impact of different relationships and mine potential alliances, and a meta-path aggregation neural network (PA) is designed to retain the voting rights information of members within the alliance.

[0047] Third, a Control Neural Network (CoNN) model is proposed to identify the winning stable coalition. The voting process is simulated through different embedding space transformations to identify the winning stable coalition and the actual controller is accurately identified through the fully connected layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction is given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 A flowchart of a method for mining actual controllers based on equity knowledge graph learning is provided for one embodiment of the present invention.

[0050] Figure 2 An example of mining actual controllers provided by an embodiment of the present invention.

[0051] Figure 3 An example of actual controller mining based on the equity knowledge graph is provided for one embodiment of the present invention.

[0052] Figure 4 An example of actual controller mining result analysis based on the equity knowledge graph provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention are arbitrarily combined with each other to form a new technical solution. This combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0054] In view of the problem that the existing technology only considers equity relations to mine actual controllers, and the complex interactions between shareholders and different voting strategies affect the accuracy of identifying stable alliances during mining, the present invention proposes an actual controller mining method based on equity knowledge graph learning, which aims to capture the semantic and structural information in large-scale equity networks according to the characteristics of winning stable alliances by modeling complex relationships between shareholders, and mine hidden actual controllers. The present invention mainly faces the following challenges: first, the relationship between shareholders is complex and multidimensional, which is difficult to represent and model; second, there is potential mutual influence between shareholders, and different types of association relationships have different degrees of influence on the formation of stable alliances; in addition, the interaction between shareholders and different voting strategies increase the complexity of the voting process, which in turn affects the accuracy of judging whether a stable alliance has won.

[0055] The various relationships among millions of shareholders form a huge equity network, and the relationships among shareholders are diverse, including relatives, company executives, parent-subsidiary companies, etc. Figure 2 As shown, for example, e6 gains control over e1 by controlling e2 and influencing e3, thus becoming the actual controller of e1. Due to their association, e6 and e3 are likely to act in concert, so e6 and e3 form a stable alliance. In contrast, the alliance formed by e5 and e3 or e5 and e6 is unstable. Through the shareholding paths e6→e2→e1 and e3→e1, the total shareholding ratio of the alliance (e3, e6) is 55%. According to the principle of one share, one vote, the voting weight of the alliance (e3, e6) is 55%>50%, which indicates that the alliance (e3, e6) is a stable winning alliance. However, if the stability of the alliance is ignored, e5 may be misidentified due to the involvement of two winning alliances (e5, e3) and (e5, e6).

[0056] like Figure 1As shown, the actual controller mining method based on equity knowledge graph learning provided by the embodiment of the present invention includes the following steps:

[0057] Step 1: Construct an equity knowledge graph based on the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the graph.

[0058] The equity knowledge graph in step 1 is defined as Each triple is a de novo entity e i To the end entity e j There is a relationship r between them. ε is an entity set, e i ∈ε represents shareholders or companies, and each entity e i With a set of attributes {a i1 , ..., a im}associated. is a set of relations, including forward relations and reverse relations. Specifically, Represents the equity relationship, each triple (e i , r s , e j ) are related to voting weight Associated.

[0059] Step 2: In the equity knowledge graph described in step 1, design RA to automatically extract various relationship features between shareholders and discover stable alliances. Figure 3 As shown in Figure 3, to capture the transitivity between relations, RA encodes each relation by aggregating information from multi-hop neighbors and generates entity embeddings through self-attention to capture the mutual influence of different types of relations.

[0060] Entity e in step 2 i Embedded i Initialized by the attributes of the entity, the specific calculation method is as follows:

[0061] u i =h o (a i1 , a i2 , ..., a im )

[0062] where h o is to set the attribute {a i1 , a i2 , ..., a im} is converted to an embedding function, where o is the entity type. For each relationship RA aggregates information from multiple neighbors multiple times. Specifically, in each iteration, for entity e i ∈ε, RA can aggregate additional neighbor information into e i The lth layer embedding The calculation method is as follows:

[0063]

[0064] Among them N r (e i ) = {e j |(e i , r, e j ) ∈ G} represents the neighbors of e i under the relationship r, where l is a hyperparameter representing the number of hops and 0 < l < L. The L-layer embedding of different relationships of the entity e i can be expressed as:

[0065]

[0066] The self-attention mechanism is used to measure the mutual influence between different relationships. The final embedding u i of the entity e ir on the relationship r is calculated as follows:

[0067]

[0068] Among them, M r is a trainable transformation matrix, and α ir represents the coefficient of the linear combination of the relationship r. The calculation method of α ir is as follows:

[0069]

[0070] Among them, m′ r and M′ r are trainable parameters of the relationship type r, and α ir represents the influence weight of the relationship r in forming a stable coalition of e i . Different influences of various types of relationships on the stable coalition are retained in the embedding.

[0071] Given a set of shareholders If there is at least one path recorded as e i and e j between e i e t ∈ ε g unanimous action path, then e i will make a unanimous decision with e j , and then make a unanimous decision with e j along this path. The stability p g of the coalition formed by a set of shareholders ε a(g) It is determined by the probability that all shareholders make consistent decisions in repeated games. The calculation method of alliance stability is as follows:

[0072]

[0073] If p a (g)>τ a , then the alliance is stable; otherwise, the alliance is unstable. The higher the probability of concerted action among shareholders, the stronger the stability of the alliance. Shareholders with concerted action relationships are likely to form stable alliances, so the discovery of stable alliances can be transformed into a prediction of concerted relationships and implemented based on the similarity of the generated embeddings. In order to incorporate the different effects of various relationships, the embeddings of all types of relationships are all need to be taken into account. i , e j ∈ε, probability of concerted action Calculated by cosine similarity:

[0074]

[0075] Among them U i and U j They are and Embedded splicing. Stability threshold τ a The model is adaptively adjusted through the training process of the model. The model learns the information of winning stable alliances by training on the equity knowledge graph and the real data set of actual controllers. Therefore, τ a It is dynamically determined based on the difference in the probability of shareholders acting in concert in stable alliances and unstable alliances.

[0076] Step 3: Based on the RA aggregated equity structure information described in step 2, PA is designed to capture the shareholder's shareholding path, and the shareholding ratio information is retained based on the probability of the shareholder and the company co-occurring in the final captured path.

[0077] like Figure 3 As shown in Figure 2, PA is traversed according to two predefined meta-path patterns. In order to capture the shareholder’s shareholding path, the meta-path pattern is set to Where l represents the pattern length, Through this multi-hop shareholding relationship, the voting weight obtained through the nested layer can be captured, and the long shareholding path can be captured by adjusting the value of l. In order to capture the shareholding path of members in a stable alliance, the meta-path mode is set to Since shareholders can expand their voting power by influencing other shareholders, a is set as the first step of the model, in which equity relationships are set up in subsequent steps to retain the voting rights of other shareholders within the alliance.

[0078] The equity path is captured by a random walk, where the i To e j The transition probability p(e i |e j ) is e i The normalized probability of neighbors. Based on the above meta-path pattern, the transition probability p(e i |e j , t) can be calculated as:

[0079]

[0080] According to the meta-path mode, r t Can represent r a or s , is the corresponding weight of the walking edge in the tth step. Based on this transition probability, PA captures the shareholding path through graph traversal, calculates the frequency of co-occurrence in the paths generated by each entity pair according to the voting weight information, and retains it.

[0081] After generating equity paths, PA optimizes the learned entity embeddings by running the skip-gram model. For each pair of entities in the same path (e i , e j ), the embedding is calculated as follows:

[0082]

[0083] where v i Yes i The embedding of j Yes j The context embedding in the entity sequence, σ is the sigmoid function, calculated as σ(x) = 1 / (exp(x) + 1), N′(e i ) is e i The negative sampling set, N′(e i ) i is the noise distribution P(e i ) is a sample randomly selected from .

[0084] Step 4: Based on the embeddings generated by RA and PA in steps 2 and 3 above, CoNN simulates the voting process by extracting the information of stock weights and stable alliances preserved in the entity embeddings to predict the control relationship between shareholders and companies.

[0085] The CoNN designed in step 4 uses the embedding v i and v j to assess the similarity of shareholders iThrough the shareholding path, the company j On the other hand, by changing v i The information of different types of stable alliances is extracted by transferring them to different embedding spaces, and then the final voting results of each stable alliance are analyzed to evaluate the shareholder e i Through stable winning alliances, the company j control.

[0086] like Figure 3 As shown, for company e j , if e i is one of the actual controllers, then i and e j There may be an equity path between them, leading to the embedding of v i and v j The similarity between them is high. CoNN first calculates v i and v j The cosine similarity between the two is calculated by concatenating the cosine score and the entity embedding into a single vector to merge the entity information and computing e through a fully connected hidden layer. i and e j The similarity between i，j The specific calculation method is as follows:

[0087] Sim i，j =σ(W(v join =v i ||cos(v i , v j )||v j ))

[0088] Where || represents concatenation and W is a trainable matrix.

[0089] For shareholders i , CoNN uses multiple layers in a multi-layer perceptron (MLP) to explore latent information for stable coalitions:

[0090] H im =tanh(M m V i )

[0091] Among them, H im , m∈{1, 2, ..., n} is the potential information in different embedding spaces, M m is a trainable transformation matrix.

[0092] For Company j , and shareholders i Other shareholders in the same stable alliance can be divided into direct holdings of e j Shares, indirect holdings j Shares and non-sharesj There are three types of shares, so CoNN uses three different MLPs to extract these three types of information. i Transformed into three embedding spaces with different winning conditions. CoNN determines these three embedding spaces by fitting the control relationship, and models the shareholder results under different conditions based on these three different embedding spaces to simulate the voting process. Whether the stable alliance wins depends on the similarity between the spatial information and the target company information. The specific calculation method is as follows:

[0093]

[0094]

[0095] Among them, H i Contains information about the probability of a stable coalition winning, Represents matrix multiplication. The similarity of stable alliances Sim′ i，j The calculation method is as follows:

[0096] Sim′ i，j =σ(W′(H i ||cos(H i , v j )||v j ))

[0097] Where W′ is a trainable matrix. Through the hidden layers, the linear integral of Sim and Sim′ is calculated to construct the CoNN. Probability The calculation method is as follows:

[0098]

[0099] Where β and β' are trainable parameters. To optimize the parameters of CoNN, set the coefficient L c To distinguish between actual controllers and non-actual controllers. The calculation method is as follows:

[0100]

[0101] where z i,j Yes i For j The calculation method of γ coefficient is:

[0102]

[0103] where τ z The threshold for determining whether a shareholder is the actual controller based on control rights.

[0104] The specific implementation of the present invention is as follows Figure 4 shown.

[0105] The equity structure of JS Bioengineering is relatively dispersed. According to official disclosure, the actual controller of the company is LK. Through the equity knowledge graph prediction, it can be known that LK is the actual controller and the control right is 0.8789. However, if the existing actual controller disclosure method is used, GS Investment will be incorrectly disclosed due to its largest shareholding ratio.

[0106] Further analysis of the actual controller prediction results on the equity knowledge graph shows that RA predicts the concerted actions between LK and a group of shareholders through various relationships between shareholders. These concerted shareholders form a stable alliance with a total voting weight of 30.35% and win the voting process. Expand control, where e k is a shareholder in the same stable alliance as LK. In addition, the present invention successfully predicts the top five shareholders of JS Bioengineering, namely LK, LY, YQ, ZH and WH, through equity knowledge graph representation learning, which indicates that the top five shareholders except LK all hold shares and can obtain control through a stable alliance.

[0107] This case study shows that shareholders with a low shareholding ratio can gain control by forming a winning and stable alliance with other shareholders.

[0108] The implementation basis of each embodiment of the present invention is to implement programmed processing through a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides an actual controller mining device based on equity knowledge graph learning, which is used to execute an actual controller mining method based on equity knowledge graph learning in the above method embodiment.

[0109] The device includes: a first calculation module, which is used to construct an equity knowledge graph according to the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the equity knowledge graph through entity alignment; a second calculation module, which is used to design a multidimensional association aggregation neural network RA in the equity knowledge graph to automatically extract various relationship features between shareholders and discover stable alliances; a third calculation module, which is used to design a meta-path aggregation neural network PA to capture the shareholder's shareholding path based on the aggregation of equity structure information by the multidimensional association aggregation neural network RA, and retain voting rights information based on the probability of co-occurrence of shareholders and companies in the final captured path; a fourth calculation module, which is used to design a control neural network CoNN based on the embedding generated by the multidimensional association aggregation neural network RA and the meta-path aggregation neural network PA, simulate the voting process by extracting the stock weights and stable alliance information stored in the entity embedding, and predict the control relationship between shareholders and companies.

[0110] An embodiment of the present invention provides an actual controller mining device based on equity knowledge graph learning, which adopts the above-mentioned modules to improve the accuracy and efficiency of actual controller mining, and solves the problem that the existing technology only considers equity relations to mine actual controllers, and the complex interactions between shareholders and different voting strategies during mining affect the accuracy of stable alliance identification.

[0111] It should be noted that the device embodiments provided by the present invention are not only used to implement the methods in the above-mentioned method embodiments, but also used to implement the methods in other method embodiments provided by the present invention. The only difference lies in the setting of corresponding functional modules, and the principles thereof are basically the same as the principles of the above-mentioned device embodiments provided by the present invention. As long as technical personnel in this field refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned device embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solutions, improve the modules in the above-mentioned device embodiments to obtain corresponding device class embodiments for implementing the methods in other method class embodiments.

[0112] Based on the same inventive concept as the above-mentioned embodiment, an embodiment of the present invention also provides an actual controller mining system based on equity knowledge graph learning, including a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the actual controller mining method based on equity knowledge graph learning.

[0113] Based on the same inventive concept as the above-mentioned embodiment, an embodiment of the present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable the computer to execute the steps of the actual controller mining method based on equity knowledge graph learning.

[0114] In summary, the actual controller mining method based on equity knowledge graph learning proposed in the present invention solves the problems of low accuracy and high time cost in discovering stable alliances caused by complex multivariate relationships in the equity network, and is applied to the field of financial technology. By constructing an equity knowledge graph to represent the semantic and structural information in the equity network, by aggregating different types of relationships, capturing the shareholding path in the equity knowledge graph to obtain stable alliances and voting rights information on the equity knowledge graph, and using a control neural network to simulate the shareholder voting process, determine the winning stable alliance, and ultimately achieve efficient actual controller mining.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A method for mining actual controllers based on equity knowledge graph learning, characterized in that: include: Step 1: Construct an equity knowledge graph based on the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the equity knowledge graph through entity alignment; Step 2: In the equity knowledge graph described in step 1, a multi-dimensional association aggregation neural network RA is designed to automatically extract various relationship features between shareholders and discover stable alliances; Step 3: Based on the multi-dimensional association aggregation neural network RA aggregating the equity structure information in step 2, a meta-path aggregation neural network PA is designed to capture the shareholder's shareholding path, and the voting rights information is retained according to the probability of the shareholder and the company co-occurring in the final captured path; Step 4: Based on the embeddings generated by the multi-dimensional association aggregation neural network RA and the meta-path aggregation neural network PA in steps 2 and 3 above, a control neural network CoNN is designed to simulate the voting process by extracting the information of stock weights and stable alliances preserved in the entity embeddings to predict the control relationship between shareholders and companies.

2. According to claim 1, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: Step 1 includes: Constructing the equity knowledge graph G = {(e i ,r,e j )|e i ,e j ∈ε,r∈R}, where each triple is a de novo entity e i To the end entity e j There is a relationship r between them, ε is the entity set, e i ∈ε represents shareholders or companies, and each entity e i With a set of attributes {a i1 ,…,a im }Associated, It is a relationship set, including forward relationships and reverse relationships.

3. According to claim 1, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: The multi-dimensional association aggregation neural network RA automatically extracts various relationship features between shareholders and discovers stable alliances, including: aggregating information from multi-hop neighbors to encode each relationship, capturing the mutual influence of different types of relationships by generating entity embeddings through self-attention, and capturing the transitivity between relationships.

4. According to claim 3, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: The multi-dimensional association aggregation neural network RA automatically extracts various relationship features between shareholders and finds stable alliances, further comprising: Using the attributes of entities to aggregate the entities of multi-dimensional association neural network RA i Embedded i Initialize and calculate as follows: u i =h o (a i1 ,a i2 ,…,a im ) where h o is to set the attribute {a i1 ,a i2 ,…,a im } is converted to an embedding function, o is the entity type; for each relationship Layer l Embedding The calculation is as follows: in N r (e i )={e j |(e i ,r,e j )∈G} represents e i Neighbors under relation r, l is a hyperparameter representing the number of hops and 0 <l<L; Entity i The L-th layer embedding of different relations is expressed as: The self-attention mechanism is used to measure the mutual influence between different relations, and the entity e i The final embedding u on relation r ir The calculation method is as follows: Among them, M r is a trainable transformation matrix, α ir The coefficients representing the linear combination of the relations r; A group of shareholders g The stability of the coalition composed of ∈εp a The calculation method of (g) is as follows: in, is the probability of consistent action, calculated by cosine similarity.

5. According to claim 1, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: Step 3: Design the meta-path aggregation neural network PA to traverse according to two predefined meta-path patterns, including: Set meta path mode to capture the shareholder’s shareholding path, where l represents the pattern length, Set meta path mode to capture the shareholding paths of members in a stable coalition; The equity path is captured by a random walk, where the i To e j The transition probability p(e i |e j ) is e i The normalized probability of neighbors, based on the above meta-path pattern, the transition probability p(e i |e j ,t) is: According to the meta-path mode, r t Represents r a or s , is the corresponding weight of the walking edge in the tth step. Based on the transition probability, the meta-path aggregation neural network PA captures the shareholding path through graph traversal. According to the voting weight information of each entity pair, the frequency of co-appearance in the paths they generate is calculated and retained; After generating the equity paths, the meta-path aggregation neural network PA learns entity embeddings by running the skip-gram model optimization.

6. According to claim 5, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: For each pair of entities in the same path (e i ,e j ), the embedding is calculated as follows: where v i Yes i The embedding of j Yes j The context embedding in the entity sequence, σ is the sigmoid function, calculated as σ(x) = 1 / (exp(x) + 1), N'(e i ) is e i Negative sampling set, N'(e i ) i is the noise distribution P(e i ) is a sample randomly selected from .

7. According to claim 1, a method for mining actual controllers based on equity knowledge graph learning is characterized in that: Step 4: Designing the control neural network CoNN includes: on the one hand, by using the embedded v i and v j to assess the similarity of shareholders i Through the shareholding path, the company j On the other hand, by changing v i The information of different types of stable alliances is extracted by transferring them to different embedding spaces, and then the final voting results of each stable alliance are analyzed to evaluate the shareholder e i Through stable winning alliances, the company j control.

8. A device for mining actual controllers based on equity knowledge graph learning, characterized in that: include: The first calculation module is used to construct a shareholding knowledge graph based on the shareholding relationship between shareholders and the company, and incorporate other relationships between shareholders into the shareholding knowledge graph through entity alignment; The second computing module is used to design a multi-dimensional association aggregation neural network RA in the equity knowledge graph to automatically extract various relationship features between shareholders and discover stable alliances; The third calculation module is used to design a meta-path aggregation neural network PA to capture the shareholder's shareholding path based on the multi-dimensional association aggregation neural network RA aggregating the equity structure information, and retain the voting rights information according to the probability of the shareholder and the company co-appearing in the finally captured path; The fourth computing module is used to design a control neural network CoNN based on the embeddings generated by the multi-dimensional association aggregation neural network RA and the meta-path aggregation neural network PA. It simulates the voting process by extracting the information of stock weights and stable alliances saved in the entity embeddings to predict the control relationship between shareholders and companies.

9. A system for mining actual controllers based on equity knowledge graph learning, characterized in that: It includes a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of a method for mining actual controllers based on equity knowledge graph learning as described in any one of claims 1 to 7.

10. A non-transitory computer readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the steps of a method for mining actual controllers based on equity knowledge graph learning as described in any one of claims 1 to 7.

Images