A model construction method for associated party identification based on reinforcement learning

Abstract

The present application relates to the technical field of financial data analysis, and more particularly to a model construction method for associated party identification based on reinforcement learning, comprising: constructing a heterogeneous enterprise correlation graph; determining an exploration strategy for the associated party based on a directional center representation parameter; determining whether there is a potential associated party in the exploration path based on correlation complexity, so as to determine whether the proportion of the potential associated party is qualified according to the correlation strength; determining whether there is an abnormal associated party based on a path connectivity representation value, so as to determine the type of the abnormal associated party according to the absolute value of the difference; taking the associated party identification result in the associated party identification process as associated party identification data to establish an associated party identification model, determining whether the identification accuracy of the associated party identification model is qualified based on the identification accuracy, and determining the adjustment of the acquisition cycle length according to the ratio; identifying the associated party of the enterprise to be identified based on the trained reinforcement learning model, and the present application improves the accuracy of associated party identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data analysis, and in particular to a model construction method for associated party identification based on reinforcement learning. BACKGROUND

[0002] In the financial field, accurately identifying associated parties is crucial for risk prevention, compliance management, and conflict of interest detection. Traditional methods mainly rely on manual review and simple rule matching, but these methods are inefficient and prone to errors when faced with complex and variable relationships and massive data, making it difficult to meet actual needs. Reinforcement learning, as an important branch of machine learning, learns the optimal behavior strategy through the interaction between the agent and the environment based on the reward signals fed back by the environment.

[0003] Chinese Patent Application Publication No. CN114020933A discloses a method for identifying associated parties and relationships in commercial banks, which includes obtaining shareholder data from trading exchanges, external business data from banks, and internal employee data from banks, and preliminarily determining the list of associated parties according to the predetermined path of the shareholder data, external business data, and internal employee data of the bank, as well as the associated party information. The final associated party list is screened according to the calculation of the total holding ratio after the penetration of the associated parties in the associated party list, and the mapping of the associated relationship, and the associated relationship of each associated party is determined. The invention also proposes an identification device for associated parties and relationships in commercial banks, which not only considers the shareholder information of commercial banks, but also comprehensively considers the information of bank employees and close relatives, effectively improving the efficiency and accuracy of associated party and relationship identification, reducing the transaction risk of commercial banks, and promoting the safe and stable operation of commercial banks.

[0004] However, the prior art has the following problems: the potential associated party relationship in the associated party is not finely processed, resulting in low accuracy of potential associated party identification, and thus low accuracy of associated party identification. SUMMARY

[0005] Therefore, the present application provides a model construction method for associated party identification based on reinforcement learning to overcome the problem of low accuracy of potential associated party identification due to the lack of fine processing of potential associated parties in the associated party in the prior art, thereby improving the accuracy of associated party identification.

[0006] To achieve the above-mentioned purpose, the present application provides a model construction method for associated party identification based on reinforcement learning, comprising:

[0007] Obtaining the equity relationship data and the position relationship data of a plurality of enterprises, and constructing a heterogeneous enterprise association relationship graph;

[0008] The exploration strategy for related parties is determined based on the orientation center representation parameter of the enterprise relationship graph.

[0009] The existence of potential related parties in the exploration path is determined by the association complexity of the path nodes based on the corresponding exploration strategy, and the proportion of potential related parties is determined based on the association strength of the exploration path with potential related parties.

[0010] The existence of abnormal related parties is determined based on the path connectivity representation value of the heterogeneous enterprise relationship graph where no potential related parties exist, and the type of abnormal related party is determined based on the absolute value of the difference between the path connectivity representation value and the preset path connectivity representation value.

[0011] The related party identification results in the related party identification process are used as related party identification data to establish a related party identification model. Based on the identification accuracy of the related party identification model on the real-time acquired enterprise information data, it is determined whether the identification accuracy of the related party identification model is qualified. The acquisition cycle length of equity relationship data and employment relationship data is determined according to the ratio of the identification accuracy to the preset identification accuracy.

[0012] The adjusted related party identification results are used as the optimized related party identification dataset to train the reinforcement learning model, and the related parties of the enterprise to be identified are identified based on the trained reinforcement learning model.

[0013] Furthermore, under the condition of constructing the heterogeneous enterprise relationship graph, the exploration strategy is determined to be an upward exploration strategy based on the comparison result that the direction center representation parameter of the heterogeneous enterprise relationship graph is less than or equal to the first preset direction center representation parameter. The upward exploration strategy is to explore upwards from the nodes in the heterogeneous enterprise relationship graph.

[0014] Furthermore, based on the comparison result that the directional center representation parameter of the heterogeneous enterprise relationship graph is greater than the first preset directional center representation parameter and less than or equal to the second preset directional center representation parameter, the exploration strategy is determined to be a downward exploration strategy, which is to explore downwards from the nodes in the heterogeneous enterprise relationship graph.

[0015] Furthermore, based on the comparison result that the directional center representation parameter of the heterogeneous enterprise relationship graph is greater than the second preset directional center representation parameter, the exploration strategy is determined to be a hybrid exploration strategy, which is to explore both above and below the nodes in the enterprise relationship graph simultaneously.

[0016] Furthermore, under the condition of determining the exploration strategy, based on the comparison result that the association complexity of the path nodes of the corresponding exploration strategy is less than or equal to the preset association complexity, it is determined that there are no potential related parties, and based on the comparison result that the path connectivity representation value of the heterogeneous enterprise relationship graph is less than or equal to the preset path connectivity representation value, it is determined that there are abnormal related parties.

[0017] Furthermore, given a defined exploration strategy, the existence of potential related parties is determined based on the comparison results of the path nodes of the corresponding exploration strategy having a correlation complexity greater than a preset correlation complexity. The proportion of potential related parties in the exploration path is determined to be qualified based on the comparison results of the correlation strength of the exploration path being greater than a preset correlation strength.

[0018] Furthermore, under the condition that the proportion of potential related parties is qualified and there are abnormal related parties, the abnormal related parties are determined to be structural fracture abnormal related parties based on the comparison result that the absolute value of the difference between the path connectivity characterization value and the preset path connectivity characterization value is less than or equal to the preset absolute value of the difference.

[0019] Furthermore, based on the comparison result that the absolute value of the difference between the path connectivity representation value and the preset path connectivity representation value is greater than the preset absolute value of the difference, the abnormal associated party is determined to be a closed-loop abnormal associated party.

[0020] Furthermore, based on the comparison result that the identification accuracy of the related party identification model for real-time acquired enterprise information data is less than the preset identification accuracy, it is determined that the identification accuracy of the related party identification model is unqualified, and based on the comparison result that the ratio of the identification accuracy to the preset identification accuracy is less than or equal to the preset ratio, it is determined to increase the acquisition cycle duration by a first preset duration adjustment coefficient.

[0021] Furthermore, based on the comparison result that the ratio of the recognition accuracy to the preset recognition accuracy is greater than the preset ratio, the acquisition cycle duration is increased by a second preset duration adjustment coefficient.

[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a heterogeneous enterprise relationship graph, determines the exploration strategy of related parties based on the direction center representation parameter, judges whether there are potential related parties based on the relationship complexity of the path nodes of the exploration strategy, and determines whether the proportion of potential related parties is qualified based on the relationship strength. At the same time, it uses the path connectivity representation value of the path without potential related parties to judge whether there are abnormal related parties, thereby achieving accurate identification of related parties, improving the efficiency of related party identification, enhancing risk identification capabilities, adapting to the comprehensive mining of complex relationships, and thus improving the accuracy of related party identification.

[0023] Furthermore, this invention determines the exploration strategy by using the direction center characterization parameter, and judges whether there are potential related parties based on the path node association complexity under the exploration strategy. This improves the efficiency of identifying potential and abnormal related parties in enterprise relationships, reduces computational complexity, improves system efficiency, and enhances identification accuracy.

[0024] Furthermore, the present invention determines whether the proportion of potential related parties in the exploration path is qualified by comparing the correlation strength with the preset correlation strength, and compares the path connectivity characterization value with the preset path connectivity characterization value to determine whether there are abnormal related parties, thereby improving the accuracy of identifying abnormal related parties and their types, improving data utilization efficiency, and enhancing the system's adaptability.

[0025] Furthermore, this invention identifies real-time acquired enterprise information data through a related party identification model, and judges whether the model's identification accuracy is qualified by comparing the identification accuracy rate with the preset identification accuracy rate, thereby adjusting the acquisition cycle time, dynamically optimizing data acquisition, adapting to changes in actual needs, improving system performance and resource utilization efficiency, enhancing model identification accuracy, and thus improving the accuracy of related party identification. Attached Figure Description

[0026] Figure 1 This is a flowchart of a model construction method for association identification based on reinforcement learning, according to an embodiment of the present invention.

[0027] Figure 2 A flowchart for determining whether a potential related party exists in the exploration path according to an embodiment of the present invention;

[0028] Figure 3 This is a flowchart illustrating whether the proportion of potential related parties in the exploration path is qualified, as described in this embodiment of the invention.

[0029] Figure 4 A flowchart for determining whether there are abnormal related parties in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0031] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0032] It should be noted that the data in this embodiment are all derived from a comprehensive analysis and evaluation of historical test data and corresponding historical test results from the three months prior to this test. Those skilled in the art will understand that the determination of the above-mentioned parameters for any single item in this invention can be achieved by selecting the value with the highest percentage based on the data distribution as the preset standard parameter, using weighted summation to obtain the value as the preset standard parameter, substituting each historical data point into a specific formula and using the value obtained from that formula as the preset standard parameter, or other selection methods, as long as the invention can clearly define different specific situations in the single-item judgment process through the obtained values.

[0033] Please see Figure 1 As shown, it is a flowchart of the model construction method for association identification based on reinforcement learning in an embodiment of the present invention.

[0034] The present invention provides a method for constructing a model for identifying related parties based on reinforcement learning, comprising:

[0035] Step S1: Obtain equity relationship data and employment relationship data of several enterprises, and construct a heterogeneous enterprise relationship graph;

[0036] Step S2: Determine the exploration strategy for related parties based on the directional center representation parameter of the enterprise relationship graph;

[0037] Step S3: Determine whether there are potential related parties in the exploration path based on the association complexity of the path nodes of the corresponding exploration strategy, and determine whether the proportion of potential related parties is qualified based on the association strength of the exploration path with potential related parties.

[0038] Step S4: Determine whether there are any abnormal related parties based on the path connectivity representation value of the heterogeneous enterprise relationship graph where no potential related parties exist, and determine the type of abnormal related party based on the absolute value of the difference between the path connectivity representation value and the preset path connectivity representation value.

[0039] Step S5: Use the related party identification results in the related party identification process as related party identification data to establish a related party identification model. Based on the identification accuracy of the related party identification model on the real-time acquired enterprise information data, determine whether the identification accuracy of the related party identification model is qualified. Adjust the acquisition cycle length of equity relationship data and employment relationship data according to the ratio of the identification accuracy to the preset identification accuracy.

[0040] Step S6: Use the adjusted related party identification results as the optimized related party identification dataset to train the reinforcement learning model, and identify the related parties of the enterprise to be identified based on the trained reinforcement learning model.

[0041] Specifically, this invention constructs a heterogeneous enterprise relationship graph and determines the exploration strategy for related parties based on the direction center representation parameter. It determines whether there are potential related parties based on the relationship complexity of the path nodes of the exploration strategy, and determines whether the proportion of potential related parties is qualified based on the relationship strength. At the same time, it uses the path connectivity representation value of the path without potential related parties to determine whether there are abnormal related parties. This achieves accurate identification of related parties, improves the efficiency of related party identification, enhances risk identification capabilities, and adapts to the comprehensive mining of complex relationships, thereby improving the accuracy of related party identification.

[0042] In this embodiment of the invention, the acquisition period for the equity relationship data and the employment relationship data is one month. It is understood that the equity relationship and employment relationship data of an enterprise change frequently. A one-month period can ensure the timeliness and accuracy of the data, while avoiding the increased cost caused by excessively frequent data updates. In practice, the acquisition period can be determined according to the actual situation. No specific limitation is made here, and it will not be elaborated further.

[0043] In this embodiment of the invention, the equity relationship data refers to the ownership relationship formed between enterprises through investment and holding, including entity information data, shareholding relationship data, and change record data.

[0044] In this embodiment of the invention, the employment relationship data refers to the associations formed by personnel holding management positions in an enterprise, including personnel information data, employment record data, and part-time relationship data, etc.

[0045] In this embodiment of the invention, obtaining the equity relationship data and employment relationship data of an enterprise requires authorization from the enterprise.

[0046] In this embodiment of the invention, the enterprise information data includes equity relationship data and employment relationship data, and the real-time acquisition frequency is 5 times / min.

[0047] Specifically, in this embodiment of the invention, the acquired equity relationship data and employment relationship data are cleaned, and the cleaned equity relationship data and employment relationship data are parsed and hyper-edge relationships are constructed. The entity parsing process is to match enterprise nodes based on the unified social credit code and to identify natural person nodes based on hash fingerprint technology. The construction of hyper-edge relationships is the geometric average of equity weight and employment weight. The above nodes and edge relationships are integrated into a graph database to form a heterogeneous enterprise relationship graph.

[0048] In this embodiment of the invention, the data cleaning and entity parsing processes are conventional processes and will not be described in detail here. The graph database is the Neo4j graph database.

[0049] During implementation, the equity weight is calculated based on the shareholding ratio, penetration level, and data update time, while the job weight is calculated based on the job level, the number of concurrent companies, and the length of service.

[0050] Specifically, the equity weight and the appointment weight can be calculated according to the following formula, expressed as follows:

[0051]

[0052] In the formula, W e Indicates equity weight, where P is the percentage of shares held, L is the look-through level, and T is the equity weight. c -T u Update the data to reflect the present time; W p This indicates the job weight, where E is the weight assigned to the position (e.g., a legal person's weight is 1.0), N is the number of companies the same natural person holds concurrently, and D is the length of service in months (more than 36 months is counted as 36).

[0053] Specifically, in this embodiment of the invention, under the condition of determining to construct the heterogeneous enterprise relationship graph, the exploration strategy of the related parties is determined based on the direction center representation parameter of the heterogeneous enterprise relationship graph and the preset direction center representation parameter;

[0054] If the direction center representation parameter is less than or equal to the first preset direction center representation parameter, then the exploration strategy is determined to be an upward exploration strategy.

[0055] If the direction center representation parameter is greater than the first preset direction center representation parameter and less than or equal to the second preset direction center representation parameter, then the exploration strategy is determined to be a downward exploration strategy.

[0056] If the direction center representation parameter is greater than the second preset direction center representation parameter, then the exploration strategy is determined to be a hybrid exploration strategy.

[0057] In this embodiment of the invention, the first preset direction center representation parameter is 0.3, which is obtained when the direction center representation parameter of several historical upward exploration strategies is taken as the maximum value. The second preset direction center representation parameter is 0.7, which is obtained when the direction center representation parameter of several historical mixed exploration strategies is taken as the average value.

[0058] In this embodiment of the invention, the upward exploration strategy is to explore upwards from the nodes in the heterogeneous enterprise relationship graph, the downward exploration strategy is to explore downwards from the nodes in the heterogeneous enterprise relationship graph, and the hybrid exploration strategy is to explore both upwards and downwards from the nodes in the heterogeneous enterprise relationship graph simultaneously.

[0059] During implementation, the directional center representation parameter is the product of out-degree centrality and weight 0.5 plus the product of in-degree centrality and weight 0.5. The out-degree centrality is the ratio of the number of edges output by the i-th node to the total number of nodes, and the in-degree centrality is the ratio of the number of edges received by the i-th node to the total number of nodes.

[0060] Please see Figure 2 As shown, it is a flowchart of an embodiment of the present invention for determining whether there are potential related parties in the exploration path.

[0061] Specifically, in this embodiment of the invention, under the condition of determining the exploration strategy, the existence of potential related parties in the exploration path is determined based on the comparison result of the association complexity of the path nodes of the corresponding exploration strategy and the preset association complexity.

[0062] If the association complexity is less than or equal to the preset association complexity, then it is determined that there are no potential related parties.

[0063] If the association complexity is greater than the preset association complexity, then a potential related party is determined to exist.

[0064] In this embodiment of the invention, the preset association complexity is set to 0.85. The preset association complexity is obtained by taking the maximum value of the association complexity of a number of historical cases in which no potential related parties exist. However, the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0065] In this embodiment of the invention, the path node is an enterprise node or natural person node on a related path in a heterogeneous enterprise relationship graph, and the exploration path is a complete link formed by starting from the starting node and extending along the heterogeneous enterprise relationship graph.

[0066] In the implementation process, the complexity of the association is the sum of several hyper-edge relationships divided by the number of equity relationships, multiplied by the sum of several hyper-edge relationships divided by the number of appointment relationships.

[0067] Specifically, this invention determines the exploration strategy by using the direction center characterization parameter, and judges whether there are potential related parties based on the path node association complexity under the exploration strategy. This improves the efficiency of identifying potential and abnormal related parties in enterprise relationships, reduces computational complexity, improves system efficiency, and enhances identification accuracy.

[0068] Please see Figure 3 As shown, it is a flowchart for determining whether the proportion of potential related parties in the exploration path is qualified according to an embodiment of the present invention.

[0069] Specifically, in this embodiment of the invention, under the condition that potential related parties exist, the proportion of potential related parties in the exploration path is determined to be qualified based on the comparison result of the association strength of the exploration path with potential related parties and the preset association strength.

[0070] If the correlation strength is less than or equal to the preset correlation strength, then the proportion of potential related parties in the exploration path is determined to be unqualified.

[0071] If the correlation strength is greater than the preset correlation strength, then the proportion of potential related parties in the exploration path is determined to be qualified.

[0072] In this embodiment of the invention, the preset association strength is 0.82. The preset association strength is obtained by taking the maximum value of the association strength when the proportion of several potential related parties in the past is unqualified. However, the above value is not limited to this. Those skilled in the art can also adjust the value according to actual needs.

[0073] During implementation, the association strength is the average of several hyperedge relationships of the exploration path with potential related parties.

[0074] In one embodiment, if the proportion of potential related parties is deemed unqualified, the exploration depth of the corresponding exploration strategy is increased. For key nodes on the exploration path, the keywords of the key nodes are matched with keywords in publicly available big data. Based on the comparison result of the matching overlap degree and the preset matching overlap degree, the association relationship between the keywords of the key nodes and the keywords in the big data is determined. If the matching overlap degree is less than or equal to the preset matching overlap degree, it is determined that there is a clear connection; if the matching overlap degree is greater than the preset matching overlap degree, it is determined that there is a clear equity relationship. The proportion of matching data with clear connections and equity relationships in all matching data is compared with the preset proportion to determine whether the association of a single key node is valid. If the proportion is less than or equal to the preset proportion, it is determined that the association of a single key node is invalid; if the proportion is greater than the preset proportion, it is determined that the association of a single key node is valid.

[0075] In this embodiment of the invention, the key node is an enterprise node or a natural person node, the keyword is an enterprise name, business scope, position, etc., the explicit reference is direct business dealings or cooperation agreements, the explicit equity relationship is shareholding ratio or equity control chain, etc., and the matching data is key node data.

[0076] In this embodiment of the invention, the preset matching overlap rate is 95%, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0077] In this embodiment of the invention, the preset ratio is 83%, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0078] During implementation, the overlap rate is the ratio of the frequency of keywords appearing at key nodes to the frequency of keywords appearing in the big data.

[0079] During implementation, the ratio is the ratio of the total number of matching data with clear orientation and equity relationship to the total number of matching data in big data.

[0080] Understandably, analyzing the connections between key nodes and reassessing the proportion of potential stakeholders in the exploration path is crucial. If more key nodes with established connections are discovered after in-depth exploration, raising the proportion of potential stakeholders to a satisfactory level, then it can be considered that there are more potential stakeholders in the exploration path, and further in-depth investigation is necessary. If the proportion of potential stakeholders is still unsatisfactory, it can be basically determined that the relationship between potential stakeholders in the exploration path is weak, and the subsequent attention to it can be reduced.

[0081] Please see Figure 4 As shown, it is a flowchart for determining whether there are abnormal related parties in an embodiment of the present invention.

[0082] Specifically, in this embodiment of the invention, the existence of abnormal related parties is determined by comparing the path connectivity representation value of the heterogeneous enterprise relationship graph in which no potential related parties exist with the preset path connectivity representation value.

[0083] If the path connectivity representation value is less than or equal to the preset path connectivity representation value, then it is determined that there is an abnormal associated party;

[0084] If the path connectivity representation value is greater than the preset path connectivity representation value, then it is determined that there are no abnormal related parties.

[0085] In this embodiment of the invention, the preset path connectivity representation value is 0.79, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0086] During implementation, the path connectivity representation value is the ratio of the average value of the hyperedge relationships of the exploration path without potential associated parties to the number of nodes.

[0087] Specifically, in this embodiment of the invention, under the condition that the proportion of potential related parties is qualified and there are abnormal related parties, the type of abnormal related party is determined according to the comparison result of the absolute value of the difference between the path connectivity characterization value and the preset path connectivity characterization value and the preset absolute value of the difference;

[0088] If the absolute value of the difference is less than or equal to the preset absolute value of the difference, then the abnormal associated party is determined to be the abnormal associated party of structural fracture.

[0089] If the absolute value of the difference is greater than the preset absolute value of the difference, then the abnormal related party is determined to be a closed-loop abnormal related party;

[0090] The absolute value of the difference is the absolute value of the difference between the path connectivity representation value and the preset path connectivity representation value.

[0091] In this embodiment of the invention, the preset absolute value of the difference is 0.35, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0092] In this embodiment of the invention, the structural fracture abnormal related party is the missing node or hierarchical break in the enterprise's related path, and the closed-loop abnormal related party is the presence of circular shareholding, cross control or implicit closed loop in the related path.

[0093] Specifically, in this embodiment of the invention, a random forest regression model is used as the model for identifying related parties.

[0094] Specifically, this invention determines whether the proportion of potential related parties in the exploration path is qualified by comparing the correlation strength with the preset correlation strength, and compares the path connectivity characterization value with the preset path connectivity characterization value to determine whether there are abnormal related parties. This improves the accuracy of identifying abnormal related parties and their types, improves data utilization efficiency, and enhances the adaptability of the system.

[0095] Specifically, in this embodiment of the invention, given the aforementioned related party identification model, the accuracy of the related party identification model is determined to be qualified based on a comparison between the accuracy of the related party identification model in identifying real-time acquired enterprise information data and a preset accuracy rate.

[0096] If the recognition accuracy is less than the preset recognition accuracy, then the recognition accuracy of the related party recognition model is determined to be unqualified.

[0097] If the recognition accuracy rate is greater than or equal to the preset recognition accuracy rate, then the recognition accuracy of the related party recognition model is determined to be qualified.

[0098] In this embodiment of the invention, the preset recognition accuracy rate is 95%. The preset recognition accuracy rate is obtained by averaging the recognition accuracy rates of several historical recognition accuracy rates that meet the requirements. However, the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0099] During implementation, the identification accuracy rate is the ratio of the number of enterprise information data successfully identified by the related party identification model to the total amount of enterprise information data, multiplied by 100%.

[0100] Specifically, in this embodiment of the invention, when it is determined that the identification accuracy of the related party identification model is unqualified, the acquisition period for adjusting equity relationship data and employment relationship data is determined based on the comparison result of the ratio of the identification accuracy rate to the preset identification accuracy rate and the preset ratio.

[0101] If the ratio is less than or equal to the preset ratio, then the acquisition period duration is increased to the corresponding value by adjusting the first preset duration coefficient of 1.2.

[0102] If the ratio is greater than the preset ratio, then it is determined that the acquisition period duration will be increased to the corresponding value by the second preset duration adjustment coefficient of 1.5;

[0103] The ratio is the ratio of the recognition accuracy to the preset recognition accuracy.

[0104] In this embodiment of the invention, the preset ratio is 0.55, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0105] In this embodiment of the invention, the increased acquisition period duration is the product of the acquisition period duration and the preset duration adjustment coefficient. The preset duration adjustment coefficient includes a first preset duration adjustment coefficient with a value of 1.2 and a second preset duration adjustment coefficient with a value of 1.5. It can be understood that the acquisition period duration should be an integer.

[0106] Specifically, the adjusted related party identification results are used as the optimized related party identification data set to train the reinforcement learning model, and the related parties of the enterprise to be identified are identified based on the trained reinforcement learning model. It can be understood that the reinforcement learning model is the Actor-Critic algorithm model.

[0107] Specifically, this invention identifies enterprise information data acquired in real time through a related party identification model, and judges whether the model's identification accuracy is qualified by comparing the identification accuracy rate with the preset identification accuracy rate, thereby adjusting the acquisition cycle time, dynamically optimizing data acquisition, adapting to changes in actual needs, improving system performance and resource utilization efficiency, and enhancing model identification accuracy, thereby improving the accuracy of related party identification.

[0108] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a model for identifying related parties based on reinforcement learning, characterized in that, include: Acquire equity relationship data and employment relationship data of several enterprises, and construct a heterogeneous enterprise relationship graph; The exploration strategy for determining related parties is based on the directional center representation parameter of the enterprise relationship graph, wherein the directional center representation parameter is the product of out-degree centrality and weight 0.5 plus the product of in-degree centrality and weight 0.

5. The association complexity of the path nodes based on the corresponding exploration strategy determines whether there are potential related parties in the exploration path, and the association strength of the exploration path with potential related parties determines whether the proportion of potential related parties is qualified. The association complexity is the result of the sum of several hyper-edge relationships divided by the number of equity relationships, multiplied by the result of the sum of several hyper-edge relationships divided by the number of appointment relationships, and the association strength is the average value of several hyper-edge relationships in the exploration path with potential related parties. The existence of abnormal related parties is determined based on the path connectivity representation values ​​of the heterogeneous enterprise relationship graph where no potential related parties exist. The type of abnormal related party is determined based on the absolute value of the difference between the path connectivity representation value and a preset path connectivity representation value. Among them, under the condition that the proportion of potential related parties is qualified and there are abnormal related parties, the abnormal related parties are determined to be structural fracture abnormal related parties based on the comparison result that the absolute value of the difference between the path connectivity characterization value and the preset path connectivity characterization value is less than or equal to the preset absolute value of the difference. The abnormal related parties are determined to be closed loop abnormal related parties based on the comparison result that the absolute value of the difference between the path connectivity characterization value and the preset path connectivity characterization value is greater than the preset absolute value of the difference. The path connectivity characterization value is the ratio of the average value of the hyperedge relationship of the exploration path without potential related parties to the number of nodes. The related party identification results in the related party identification process are used as related party identification data to establish a related party identification model. Based on the identification accuracy of the related party identification model on the real-time acquired enterprise information data, it is determined whether the identification accuracy of the related party identification model is qualified. The acquisition cycle length of equity relationship data and employment relationship data is determined according to the ratio of the identification accuracy to the preset identification accuracy. The related party identification results after adjusting the acquisition period of equity relationship data and employment relationship data are used as the optimized related party identification dataset to train the reinforcement learning model, and the related parties of the enterprise to be identified are identified based on the trained reinforcement learning model.

2. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 1, characterized in that, Under the condition of constructing the heterogeneous enterprise relationship graph, the exploration strategy is determined to be an upward exploration strategy based on the comparison result that the direction center representation parameter of the heterogeneous enterprise relationship graph is less than or equal to the first preset direction center representation parameter. The upward exploration strategy is to explore upwards from the nodes in the heterogeneous enterprise relationship graph.

3. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 2, characterized in that, Based on the comparison results of the directional center representation parameter of the heterogeneous enterprise relationship graph being greater than the first preset directional center representation parameter and less than or equal to the second preset directional center representation parameter, the exploration strategy is determined to be a downward exploration strategy, which is to explore downwards from the nodes in the heterogeneous enterprise relationship graph.

4. The method for constructing a model for association identification based on reinforcement learning according to claim 3, characterized in that, Based on the comparison result that the directional center representation parameter of the heterogeneous enterprise relationship graph is greater than the second preset directional center representation parameter, the exploration strategy is determined to be a hybrid exploration strategy, which is to explore both above and below the nodes in the enterprise relationship graph at the same time.

5. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 1, characterized in that, Under the condition of determining the exploration strategy, the comparison results based on the path node association complexity of the corresponding exploration strategy being less than or equal to the preset association complexity determine that there are no potential related parties, and the comparison results based on the path connectivity representation value of the heterogeneous enterprise association relationship graph being less than or equal to the preset path connectivity representation value determine that there are abnormal related parties.

6. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 1, characterized in that, Given a defined exploration strategy, the existence of potential related parties is determined by comparing the path nodes of the corresponding exploration strategy with the path nodes with a correlation complexity greater than the preset correlation complexity. The proportion of potential related parties in the exploration path is determined to be qualified based on the comparison of the path nodes with a correlation strength greater than the preset correlation strength.

7. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 1, characterized in that, Based on the comparison result that the recognition accuracy of the related party identification model for real-time acquired enterprise information data is less than the preset recognition accuracy, it is determined that the recognition accuracy of the related party identification model is unqualified. Based on the comparison result that the ratio of the recognition accuracy to the preset recognition accuracy is less than or equal to the preset ratio, it is determined to increase the acquisition cycle duration by a first preset duration adjustment coefficient.

8. The method for constructing a model for identifying related parties based on reinforcement learning according to claim 7, characterized in that, Based on the comparison result where the ratio of the recognition accuracy to the preset recognition accuracy is greater than the preset ratio, the acquisition cycle duration is increased by a second preset duration adjustment coefficient.

Images