Depth relation mining and graph neural network technology-based benefit association detection method

Abstract

The invention discloses a benefit association detection method based on deep relationship mining and a graph neural network technology, and belongs to the field of data analysis and mining. The method specifically comprises the following steps: firstly, establishing a cluster architecture comprising a review system main server, a cloud database and an object storage resource packet; then, a review identity card and a unified social credit code of the competing enterprise are imported to carry out unique identification verification; then, in combination with associated enterprises which are stored in a review database and have a benefit relationship with the review database, a DFS multi-level traversal algorithm is adopted to calculate a benefit association path between the review and the associated enterprises, and optimization and combination with participating enterprises are carried out to form benefit relationship graph structure data: review-associated enterprises-participating enterprises; and finally, processing graph structure data by using a graph neural network, and judging whether benefit association exists between review and competing enterprises or not. The method aims at accurately selecting judges without interest intersection for scenes such as competition and item selection, and the fairness and justice of the competition or item selection process are ensured.

Description

TECHNICAL FIELD

[0002] The application belongs to the field of data analysis and mining, and specifically relates to a benefit correlation detection method based on deep relationship mining and graph neural network technology. BACKGROUND

[0004] In the operation process of the event platform, how to effectively avoid the potential correlation between the contestants and the judges has always been a big challenge for platform managers. These correlation relationships may involve multiple aspects such as the equity structure and the interest relationship of both parties, and if not handled properly, it will seriously affect the reputation and credibility of the event.

[0005] At present, there is no interest penetration system specially for various review scenes, and the existing general tools are scattered and difficult to focus on the needs of various event review scenes. There are the following schemes about interest penetration:

[0006] Manual review: through manual review of the information of judges and participating enterprises, it is judged whether there is a benefit correlation. But the manual review method is inefficient and cannot meet the needs of large-scale events.

[0007] Database matching: through database matching of the information of judges and participating enterprises, it is judged whether there is a benefit correlation. For example, commercial software such as Aiqicha and Qicha can search for associated enterprises by searching for people or enterprises, but such platforms are scattered and not focused, and it is difficult to meet the needs of system lightweight and review special scenes.

[0008] Depth-first search algorithm: a single depth-first search algorithm can only present the benefit correlation relationship, and cannot realize the benefit correlation judgment. SUMMARY

[0010] The application proposes a benefit correlation detection method based on deep relationship mining and graph neural network technology, which aims to accurately select judges without benefit intersection for events, project selection and other scenes, and ensure the fairness and justice of the event or project selection process.

[0011] The benefit correlation detection method based on deep relationship mining and graph neural network technology has the following specific steps:

[0012] Step 1: build a cluster architecture including an evaluation system main server, a cloud database and an object storage resource package.

[0013] The evaluation system main server is 1, the cloud database is 2, and the object storage resource package is 1. The main servers are linked and accessed through the intranet, and the extranet is accessed through the domain name.

[0014] This cluster architecture adopts a public cloud server architecture, based on the HTTP / HTTPS protocol, and supported by underlying services such as the Zend engine, Redis caching, and load balancing. It supports both distributed and shared clusters, and the cluster servers can automatically synchronize files. The cluster supports disaster recovery solutions; if the primary host fails, the secondary host automatically takes over.

[0015] The cloud database is connected to the API of the listed company business information database, the review expert database, and the participating company database.

[0016] Step 2: Import the ID cards of the reviewers and the unified social credit codes of the participating companies for unique identification verification;

[0017] The unique identifier verification for participating companies aims to ensure the uniqueness of their identities, prevent identity duplication or impersonation, and avoid a decrease in the accuracy of conflict of interest warnings due to incorrect company names, thus ensuring the fairness and transparency of the competition. Based on the company's unified social credit code, the system automatically verifies the correct company name and adds it to the company database.

[0018] The unique identifier verification for judges aims to ensure the uniqueness of judges' identities, prevent identity duplication or impersonation, and avoid a decrease in the accuracy of conflict of interest warnings due to incorrect information about the name of the judge's organization during the import process, thus ensuring the fairness and transparency of the competition.

[0019] Step 3: Using the Depth-First Search (DFS) algorithm, combined with the business registration information database API of participating companies, obtain the interest linkage path between the reviewers and related companies; and based on a custom penetration depth threshold... and correlation strength threshold Optimize the path of interest linkage;

[0020] Specifically:

[0021] First, set a custom penetration depth threshold. and correlation strength threshold Limit the search scope;

[0022] Then, for review A, we search for related companies with interests stored in the review database and use the DFS multi-level traversal algorithm to calculate the path association between review and related companies.

[0023] For the currently visited vertex and search depth recursive process Judgment when or edge weight If the vertex is found to be a vertex, stop the search; otherwise, record the next vertex to be visited. , forming a path Continue the recursive process ;

[0024] Finally, to reduce unnecessary computations, pruning rules are introduced to extract the critical path:

[0025] If path Cumulative weight Then pruning is carried out. A user-defined minimum cumulative association strength; simultaneously, if the path If duplicate nodes or cycles exist, prune them.

[0026] The final output benefit-related path must satisfy:

[0027] ,and Acyclic.

[0028] Step 4: Combine the interest relationship path between the reviewers and related companies with that of the participating companies to form an interest relationship diagram structure: Reviewers - Related Companies - Participating Companies;

[0029] By treating the reviewers, related companies, and participating companies as nodes and the fund flows or guarantee relationships between them as edges, a benefit relationship graph structure is constructed.

[0030] The mathematical expression is ,in It is the set of edges in the path of shared interests. Indicates related enterprises and related companies Between, or related companies and review The connection relationship between them; It is the set of nodes in the path of shared interests; Represents the number of nodes;

[0031] The interest relationship diagram structure is represented by three matrices:

[0032] Adjacency matrix A: Matrix elements Represents a node and nodes Check if there is an edge between them; if so, return 1, otherwise return 0.

[0033] Node feature matrix Let n represent the feature vector of each node, where n represents the number of nodes. It is the feature dimension;

[0034] The degree matrix D is a diagonal matrix of the adjacency matrix A, with diagonal elements... Represents a node The degree (i.e., the number of neighboring nodes) is calculated using the following formula: ;

[0035] Step 5: Input the interest relationship graph structure into the pre-trained graph neural network model to determine the interest relationship and whether there is an interest relationship between the judges and the participating companies.

[0036] The specific process is as follows:

[0037] First, the pre-training process involves updating the node weights of the graph neural network by combining the interest relationship graph structure stored in the adjacency matrix with the degree matrix. ;

[0038] in In a graph neural network, the first... Layer nodes The node weights are initialized to the node feature matrix. ,Right now =X; Represents the adjacency matrix; Representing the adjacency matrix The degree matrix depicts the number of connections between each node and its neighboring nodes, or the total weight of the edges connected to it. In the graph neural network model, the first... The learnable weight matrix of the layer;

[0039] Then, the feature vectors of the interest relationship graph structure G are processed by a structured encoder to obtain the encoded feature vectors. ;

[0040] Finally, in practical applications, the trained graph neural network, i.e., the updated graph neural network node weights, will be used. Combined with the encoded feature vector Prediction is performed using a softmax classifier: ;

[0041] In the formula, L represents the probability of conflict of interest among participating companies. If there is a conflict of interest, then... If the value returns 0, then there is no conflict of interest. Return 1.

[0042] The advantages of this invention include:

[0043] 1) This invention is used in various competition review activities to penetrate the interests between reviewers and participating units, and realize the selection of reviewers without any conflict of interest. At present, there is no similar professional software system with a focused application area.

[0044] 2) This invention combines deep relationship mining with graph neural networks: the interest relationship path between the reviewer and related enterprises is obtained through the depth-first search algorithm. The interest relationship path is combined with the data of participating enterprises to form a graph structure data of reviewer-related enterprises-participating enterprises. The graph neural network is used to process the data to determine the interest relationship.

[0045] 3) This invention adopts enterprise unique identifier verification: the enterprise name and related information are verified by the unified social credit code to ensure the uniqueness of the enterprise identity; at the same time, the unique identifier verification of the reviewer identity is adopted to verify the information of the reviewer's unit to ensure the uniqueness of the expert's identity. Attached Figure Description

[0047] Figure 1 This is a flowchart of the interest association detection method based on deep relationship mining and graph neural network technology of the present invention;

[0048] Figure 2 This is a diagram of the interest-related early warning system architecture used in an embodiment of the present invention;

[0049] Figure 3 This is a flowchart illustrating the overall workflow of the interest-related early warning system used in this embodiment of the invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0052] To ensure the credibility and high quality of the competition, this invention proposes a method for detecting conflicts of interest based on deep relationship mining and graph neural network technology. This method provides a comprehensive and accurate service for penetrating the conflicts of interest of judges and contestants. It can penetrate the conflicts of interest of participating companies, contestants and judges, related companies and their affiliated units, automatically screen out judges with conflicts of interest, and match them with available judges without conflicts of interest.

[0053] The interest association detection method based on deep relationship mining and graph neural network technology, such as Figure 1 As shown, the specific steps are as follows:

[0054] Step 1: Build a cluster architecture that includes the main server of the review system, cloud database, and object storage resource package.

[0055] The cluster architecture is as follows: one main server for the review system, two cloud databases, and one object storage resource package. The main servers are interconnected via the intranet and accessed via domain names on the external network.

[0056] This cluster architecture adopts a public cloud server architecture, based on the HTTP / HTTPS protocol, and supported by underlying services such as the Zend engine, Redis caching, and load balancing. It supports both distributed and shared clusters, and the cluster servers can automatically synchronize files. The cluster supports disaster recovery solutions; if the primary host fails, the secondary host automatically takes over.

[0057] The cloud database is connected to the API of the listed company business information database, the review expert database, and the participating company database.

[0058] Step 2: Import the ID cards of the reviewers and the unified social credit codes of the participating companies for unique identification verification;

[0059] The unique identifier verification for participating companies aims to ensure the uniqueness of their identities, prevent identity duplication or impersonation, and avoid a decrease in the accuracy of conflict of interest warnings due to incorrect company names, thus ensuring the fairness and transparency of the competition. Based on the company's unified social credit code, the system automatically verifies the correct company name and adds it to the company database.

[0060] The unique identifier verification for judges aims to ensure the uniqueness of judges' identities, prevent identity duplication or impersonation, and avoid a decrease in the accuracy of conflict of interest warnings due to incorrect information about the name of the judge's organization during the import process, thus ensuring the fairness and transparency of the competition.

[0061] Step 3: Using the Depth-First Search (DFS) algorithm, combined with the business registration information database API of participating companies, obtain the interest linkage path between the reviewers and related companies; and based on a custom penetration depth threshold... and correlation strength threshold Optimize the path of interest linkage;

[0062] Specifically:

[0063] First, set a custom penetration depth threshold. and correlation strength threshold Limit the search scope;

[0064] Then, for review A, we search for related companies with interests stored in the review database and use the DFS multi-level traversal algorithm to calculate the path association between review and related companies.

[0065] For the currently visited vertex and search depth recursive process Judgment when or edge weight If the vertex is found to be a vertex, stop the search; otherwise, record the next vertex to be visited. , forming a path Continue the recursive process ;

[0066] Finally, to reduce unnecessary computations, pruning rules are introduced to extract the critical path:

[0067] If path Cumulative weight Then pruning is carried out. A user-defined minimum cumulative association strength; simultaneously, if the path If duplicate nodes or cycles exist, prune them.

[0068] The final output benefit-related path must satisfy: ,and Acyclic.

[0069] Step 4: Combine the interest relationship path between the reviewers and related companies with that of the participating companies to form an interest relationship diagram structure: Reviewers - Related Companies - Participating Companies;

[0070] By treating the reviewers, related companies, and participating companies as nodes and the fund flows or guarantee relationships between them as edges, a benefit relationship graph structure is constructed.

[0071] The mathematical expression is ,in It is the set of edges in the path of shared interests. Indicates related enterprises and related companies Between, or related companies and review The connection relationship between them; It is the set of nodes in the path of shared interests; Represents the number of nodes;

[0072] The interest relationship diagram structure is represented by three matrices:

[0073] Adjacency matrix A: Matrix elements Represents a node and nodes Check if there is an edge between them; if so, return 1, otherwise return 0.

[0074] Node feature matrix Let n represent the feature vector of each node, where n represents the number of nodes. It is the feature dimension;

[0075] The degree matrix D is a diagonal matrix of the adjacency matrix A, with diagonal elements... Represents a node The degree (i.e., the number of neighboring nodes) is calculated using the following formula: ;

[0076] Step 5: Input the interest relationship graph structure into the pre-trained graph neural network model to determine the interest relationship and whether there is an interest relationship between the judges and the participating companies.

[0077] The specific process is as follows:

[0078] First, the pre-training process involves updating the node weights of the graph neural network by combining the interest relationship graph structure stored in the adjacency matrix with the degree matrix:

[0079] in In a graph neural network, the first... Layer nodes The node weights are initialized to the node feature matrix. ,Right now =X; Represents the adjacency matrix; Representing the adjacency matrix The degree matrix depicts the number of connections between each node and its neighboring nodes, or the total weight of the edges connected to it. In the graph neural network model, the first... The learnable weight matrix of the layer;

[0080] Then, the feature vectors of the interest relationship graph structure G are processed by a structured encoder to obtain the encoded feature vectors. ;

[0081] Finally, in practical applications, the trained graph neural network, i.e., the updated graph neural network node weights, will be used. Combined with the encoded feature vector Prediction is performed using a softmax classifier: ;

[0082] In the formula, L represents the probability of conflict of interest among participating companies. If there is a conflict of interest, then... If the value returns 0, then there is no conflict of interest. Return 1.

[0083] Example

[0084] This embodiment employs a conflict of interest warning system that connects with a database of relevant business registration information. Using a depth-first search algorithm, it identifies the conflict of interest paths between the reviewers and their shareholdings, controlling stakes, and affiliated companies. These conflict of interest paths are then combined with participating companies to form a graph-structured data set. A graph neural network is used to automatically determine whether a conflict of interest exists between the reviewers and participating companies. Experts with no conflict of interest in the participating projects are selected for review, and the reviewers are grouped according to their respective track directions to ensure the fairness of the review results.

[0085] The aforementioned conflict of interest early warning system employs a distributed cloud computing architecture, combining deep relationship mining and graph neural network technologies. It is specifically designed to address the complex conflict of interest identification challenges between judges (experts) and participating entities (enterprises / individuals) in various review scenarios. This system automatically discovers, assesses, and warns of potential conflicts of interest from massive amounts of structured and unstructured data, significantly improving the fairness and transparency of review activities. Typical application scenarios include, but are not limited to: reviewing national-level innovation and entrepreneurship competitions, selecting experts for major government procurement projects, monitoring related-party transactions in the financial regulatory field, and reviewing the qualifications of independent directors of listed companies—areas requiring high-precision conflict of interest identification.

[0086] System architecture such as Figure 2 As shown, the conflict of interest warning system connects to the API of the listed company's business registration information database, the reviewer database, and the participating company database. Since the network of relationships between people and enterprises is generally composed of star topologies, the database uses the Neo4j graph database to better represent the relationships between enterprises and between enterprises and reviewers. Neo4j is an open-source graph database with high performance, scalability, and flexibility. Neo4j uses a node and relationship model to store data, a structure well-suited for handling data with complex relationships. Based on the characteristics of the data, the system stores reviewers' basic information, work units, and professional fields in hierarchical partitions, fully utilizing the advantages of storage components to achieve efficient data access. Simultaneously, based on a custom interest penetration depth principle and an embedded deep relationship mining algorithm, it can efficiently, accurately, and quickly identify enterprises with interests related to the reviewers, obtaining interest relationship data between reviewers and related enterprises. This interest relationship data is combined with the participating company data to form a graph structure data, which is then processed using a graph neural network to determine the interest relationship between reviewers and participating companies. This enables the review of the interest-relatedness of reviewers in the system, avoiding any potential interest-relatedness that could affect the judging results, and ensuring the credibility and high quality of the competition.

[0087] The process of the interest-related early warning system is as follows: Figure 3 As shown:

[0088] The first step is to import the review database and participating companies and verify their unique identifiers.

[0089] The second step involves the system performing a deep relationship mining function. Using a depth-first search algorithm and combined with the enterprise business information database API, the system obtains the interest relationship path between the review and related enterprises, and optimizes the interest relationship path based on the user-defined association level threshold and shareholding ratio threshold.

[0090] The third step is to combine the interest-related paths with the participating companies to form a graph-structured data: review experts - expert-related companies - participating companies.

[0091] The fourth step involves the system performing a conflict of interest determination function, using a graph neural network to process graph-structured data and determine whether there is a conflict of interest between the review experts and the participating companies.

[0092] In this embodiment, the custom penetration depth threshold dmax (e.g., a shareholding chain level of ≤5 is considered a valid association) and association strength threshold wmin (e.g., an expert's shareholding ratio of ≥5% is considered a valid association) limit the search range and avoid wasting computing power.

[0093] The Depth-First Search (DFS) multi-level traversal algorithm is then used to calculate path associations, where d is the current search depth (e.g., the level of the shareholding chain) and wij is the edge weight (e.g., the expert's shareholding percentage). To reduce invalid calculations, pruning rules are introduced to extract critical paths: where α is a user-defined minimum cumulative association strength (e.g., the total shareholding percentage must be ≥10%). An example of a critical interest association path is shown: Expert A → Company B (20% shareholding) → Company C (15% shareholding), which constitutes a valid path.

[0094] Graph Neural Networks (GNNs) are deep learning models specifically designed for processing graph-structured data. Their core principle is to update node feature representations by aggregating neighborhood information, thus capturing the topological structure and complex relationships between nodes in the graph. In this invention, entities such as reviewers, related companies, and participating companies are treated as nodes, and the relationships between them (e.g., fund flows, guarantee relationships) are treated as edges, constructing a graph structure of interest relationships. GNNs can then uncover hidden interest connections within these complex relationships based on this graph structure.

[0095] Graph convolutional networks fuse node embedding features and graph topology information, considering both node attribute features and local structural features. Through iterative node embedding updates, they gradually learn the overall representation of the graph. The weight expression for the graph neural network is shown below: ;

[0096] In the formula, the first Layer Nodes Let represent the trainable parameters in a graph convolutional neural network as follows: ,and The set of all adjacent nodes of a node is used To improve the efficiency of graph convolution operations, an adjacency matrix is ​​used to store the graph's structural information. This allows for convenient and efficient use of matrix operations to update node features. ;

[0097] in the formula Represents the adjacency matrix. The degree matrix is ​​then used To indicate, Adjacency matrices are used to depict the connectivity of each node in an adjacency matrix, representing the number of connections each node has with its neighbors or the sum of the weights of the edges connected to it. By transforming the graph structure into the form of an adjacency matrix, graph convolutional neural networks can be conveniently used to update the feature representations of each node.

[0098] The graph-structured data (reviewers - related companies - participating companies) is fed into a graph neural network, and a softmax classifier is used for prediction. ;

[0099] In the formula, L represents the probability of conflict of interest among the participating companies. This is the weight matrix of the graph neural network. These are the feature vectors of the interest relationship graph structure after passing through the structured encoder.

[0100] This invention can be applied to various scenarios that require conflict of interest checks, such as:

[0101] Competition Judging: During the competition judging process, we ensure that there is no conflict of interest between the judges and the participating organizations or contestants, and guarantee the fairness and transparency of the judging process.

[0102] Bidding and tendering: During the bidding and tendering process, ensure that there is no conflict of interest between the evaluation experts and the bidding companies, and guarantee the fairness and transparency of the bidding and tendering process.

[0103] Academic review: During the academic review process, it is ensured that there is no conflict of interest between the review experts and the authors of the papers being reviewed, thus guaranteeing the fairness and transparency of the academic review process.

[0104] The system architecture of this invention includes:

[0105] Public cloud servers: support load balancing, distributed clusters and shared clusters, and can automatically synchronize files between cluster servers.

[0106] Cloud database: Stores information about experts and participating companies, supporting high-concurrency access and data analysis.

[0107] Object storage resource package: Stores reports and files generated by the storage system, supporting high-concurrency access and data analysis.

[0108] The products involved in this invention include:

[0109] Conflict of interest warning system software: provides functions such as in-depth relationship mining, enterprise unique identifier verification, review unique identifier verification, and interest-related penetration.

[0110] Conflict of interest warning system service: Provides services based on the conflict of interest warning system, including review import, review pre-selection, review and enterprise group management, review screening, reviewer database entry, review data statistical analysis, etc.

[0111] Possible sales formats for this invention include:

[0112] Software License: Authorize users to use the conflict of interest warning system software, charged on an annual or per-use basis.

[0113] Service subscription: Provides services based on a conflict of interest warning system, charged monthly or annually.

[0114] Methods of obtaining evidence of infringement include:

[0115] Software code audit: Determine whether there is any infringement through software code audit.

[0116] System Function Comparison: By comparing the system functions, it can be determined whether any infringement has occurred.

[0117] User feedback: Determine whether any infringement has occurred based on user feedback.

[0118] Specific embodiments of the technical implementation of this invention include:

[0119] Enterprise Unique Identifier Verification: Enterprise names and related information are verified using the Unified Social Credit Code to ensure the uniqueness of enterprise identities. Verified enterprise information is added to the enterprise database to support subsequent checks for conflicts of interest.

[0120] Expert Uniqueness Verification: Verify the information of the expert's affiliated institution to ensure the uniqueness of the expert's identity. Add the verified expert information to the expert database to support subsequent checks for conflicts of interest.

[0121] Conflict of Interest Penetration: Based on an in-depth analysis of the relationships between participating companies and expert organizations, the system automatically identifies reviewers with potential conflicts of interest. A detailed conflict of interest report is generated, listing all discovered relationships between reviewers and participating organizations, including relationship type, degree of association, relationship hierarchy, and source of evidence.

[0122] This invention addresses the pain points of fragmented functions and complex operations in general-purpose tools, improving user experience and decision-making efficiency. Specifically for scenarios involving potential conflicts of interest, such as competition judging, project selection, bidding, and academic review, it innovatively implements a conflict-of-interest warning system to ensure the selection of judges without conflict of interest, guaranteeing fairness and impartiality in these scenarios.

[0123] This invention employs multi-source data fusion and real-time updates. The system connects in real-time with external data sources such as enterprise business registration information databases, enabling dynamic updates of relationships and ensuring the timeliness and reliability of the interest relationship mining results.

[0124] This invention presents a fully automated interest penetration and early warning mechanism system. Combining a depth-first search algorithm with a graph neural network, it automatically extracts the interest relationship paths between the reviewer and related companies from structured data (such as business registration information, equity relationships, and transaction records), and automatically identifies the interest relationships between the reviewer and participating companies without manual intervention. This solves the problems of low efficiency and narrow coverage in traditional manual review, achieving full-process automation.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting interest associations based on deep relationship mining and graph neural network technology, characterized in that, The specific steps are as follows: Step 1: Build a cluster architecture that includes the main server of the review system, cloud database, and object storage resource package; The cloud database is connected to the API of the listed company's business information database, the review database, and the participating company database; Step 2: Import the ID cards of the reviewers and the unified social credit codes of the participating companies for unique identification verification; Step 3: Using the Depth-First Search (DFS) algorithm, combined with the business registration information database API of participating companies, obtain the interest linkage path between the reviewers and related companies; and based on a custom penetration depth threshold... and correlation strength threshold Optimize the path of interest linkage; Step 4: Combine the interest relationship path between the reviewers and related companies with that of the participating companies to form an interest relationship diagram structure: Reviewers - Related Companies - Participating Companies; By treating the reviewers, related companies, and participating companies as nodes and the fund flows or guarantee relationships between them as edges, a benefit relationship graph structure is constructed. The mathematical expression is ,in It is the set of edges in the path of shared interests. Indicates related enterprises and related companies Between, or related companies and review The connection relationship between them; It is the set of nodes in the path of shared interests; Represents the number of nodes; The interest relationship diagram structure is represented by three matrices: 1) Adjacency matrix A: Matrix elements Represents a node and nodes Check if there is an edge between them; if so, return 1, otherwise return 0. 2) Node feature matrix Represents the feature vector of each node, where It is the feature dimension; 3) The degree matrix D is a diagonal matrix of the adjacency matrix A, and the diagonal elements... Represents a node The degree is calculated using the following formula: ； Step 5: Input the interest relationship graph structure into the pre-trained graph neural network model to determine the interest relationship and whether there is an interest relationship between the judges and the participating companies. The specific process is as follows: First, the pre-training process involves updating the node weights of the graph neural network by combining the interest relationship graph structure stored in the adjacency matrix with the degree matrix. The updated formula is: ； in In a graph neural network, the first... Layer nodes The node weights are initialized to the node feature matrix. ,Right now =X; Represents the adjacency matrix; Representing the adjacency matrix The degree matrix depicts the number of connections between each node and its neighboring nodes, or the total weight of the edges connected to it. In the graph neural network model, the first... The learnable weight matrix of the layer; Then, the feature vectors of the interest relationship graph structure G are processed by a structured encoder to obtain the encoded feature vectors. ; Finally, in practical applications, the trained graph neural network, i.e., the updated graph neural network node weights, will be used. Combined with the encoded feature vector Prediction is performed using a softmax classifier: ； In the formula, L represents the probability of conflict of interest among participating companies. If there is a conflict of interest, then... If the value returns 0, then there is no conflict of interest. Return 1.

2. The method as described in claim 1, characterized in that, In step one, the review system has one main server, two cloud databases, and one object storage resource package; the main servers are interconnected via the intranet and accessed via domain names on the external network.

3. The method as described in claim 1 or 2, characterized in that, In step one, the cluster architecture adopts a public cloud server architecture, based on the HTTP / HTTPS protocol, and supports distributed clusters and shared clusters through the Zend engine, Redis high-speed cache and load balancing underlying support services, with automatic file synchronization between cluster servers. The cluster supports disaster recovery solutions; if the primary host fails, the secondary host will automatically take over.

4. The method as described in claim 1, characterized in that, Step three specifically involves: First, set a custom penetration depth threshold. and correlation strength threshold Limit the search scope; Then, for review A, we search for related companies with interests stored in the review database and use the DFS multi-level traversal algorithm to calculate the path association between review and related companies. For the currently visited vertex and search depth recursive process Judgment when or edge weight If the vertex is found to be a vertex, stop the search; otherwise, record the next vertex to be visited. , forming a path Continue the recursive process ; Finally, to reduce unnecessary computations, pruning rules are introduced to extract the critical path: If path Cumulative weight Then pruning is carried out. A user-defined minimum cumulative association strength; simultaneously, if the path If duplicate nodes or cycles exist, prune them. The final output benefit-related path must satisfy: ,and Acyclic.

Images