Enterprise risk assessment method based on multi-agent risk assessment algorithm MA-ERC

Abstract

The invention relates to the field of data assessment, and discloses an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC, and the method comprises the following steps: S1, constructing an enterprise multi-dimensional risk knowledge graph; s2, designing a risk assessment algorithm MA-ERC to assess initial risks of the enterprises and quantify propagation risk values among the enterprises; s3, constructing an MNF-GNN model to carry out enterprise risk assessment; s4, constructing an enterprise risk assessment interpretable graph EG; according to the method, a multi-agent architecture based on a large language model and a risk propagation algorithm based on a graph are combined, the risk propagation process between enterprises is more accurately and reasonably simulated, and the assessment precision of a risk assessment model can be effectively improved through the calculated propagation risk value; and a new thought is provided for the combination of the multi-agent based on the large language model, the knowledge graph, the graph neural network and other technologies.

Description

Technical Field

[0001] The present application belongs to the field of data evaluation, and in particular relates to an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC. Background Art

[0002] Current mainstream enterprise risk assessment technologies mostly build classification models based on structured information such as corporate financial data. However, these approaches are limited to an internal enterprise perspective and fail to consider the extensive interconnectedness between enterprises. Against this backdrop, knowledge graphs, as effective tools for integrating multi-source information and expressing complex entity relationships, have gradually been introduced into the field of enterprise risk modeling. However, existing methods still face several major technical difficulties. First, before using graph propagation algorithms to simulate the risk propagation process between enterprises, how can we most effectively quantify an enterprise's inherent risk? Enterprise risk factors are multi-source and multimodal, and a reasonable initial enterprise risk value is a prerequisite for accurately simulating risk propagation. Second, traditional graph neural networks often focus on relationships between nodes, compressing node attributes into fixed vectors and relying too heavily on flattened node features. Finally, while graph neural networks (GNNs) have shown promise in graph embedding in the financial sector, they are still essentially black-box models. Their evaluation results often lack interpretability, making them difficult to directly apply in the financial industry. Summary of the Invention

[0003] The technical problem to be solved by this application is to overcome the shortcomings of the existing technology. This application provides an enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC.

[0004] To achieve the above objectives, this application provides an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC, comprising the following steps: S1. Constructing a multi-dimensional enterprise risk knowledge graph: At least one model layer is designed, wherein the model layer defines at least a core entity, a risk propagation analysis entity, and a risk assessment analysis entity and defines corresponding relationship types between the entities; S2. Design the risk assessment algorithm MA-ERC to assess the initial risk of enterprises and quantify the risk value of transmission between enterprises: The risk assessment algorithm MA-ERC uses a multi-agent evaluation framework to assess the initial risk of an enterprise, and combines it with a personalized Pagerank algorithm to simulate risk propagation to quantify the risk value propagated between enterprises. The multi-agent evaluation framework uses prompt words to set up different categories of expert agents to perform risk assessments based on different enterprise risk factors. S3. Constructing MNF-GNN model for enterprise risk assessment: The multimodal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features, and a hybrid expert model is used for feature fusion to convert the multimodal enterprise data into node feature vectors. A multi-layer RGAT model is used to model heterogeneous relationships and dynamically learn the importance weights of neighbor nodes under different relationship types. The final node features are obtained by stacking multiple layers of RGAT, and the features corresponding to the enterprise nodes are extracted. The enterprise's propagation risk value is mapped to a vector, and the features corresponding to the enterprise nodes and the enterprise's propagation risk value are spliced ​​to obtain the enterprise's comprehensive features. The model's evaluation results are obtained through a linear layer. S4. Construct an interpretable graph EG for enterprise risk assessment: By obtaining the enterprise multi-dimensional risk scores, enterprise multi-dimensional risk analysis reports and the risk transmission probability between enterprises, an enterprise risk assessment interpretable graph EG is constructed to provide an explainable description for the final assessment results.

[0005] Optionally, the model layer in S1 defines at least a core entity, a risk propagation analysis entity, and a risk assessment analysis entity, and defines corresponding relationship types between the entities, including: The core entities include enterprise entities, the risk propagation analysis entities include three types of entities: shareholders, executives and controllers, which are used to analyze the external structure and equity control path of the enterprise to simulate risk propagation, the risk assessment analysis entities include two types of entities: industry and city, which are used to explore the impact of macro-environmental policy factors on enterprise risks, and also include sign risk event entities, which are used to describe the risk signs disclosed by the enterprise in the current year.

[0006] Optionally, the multi-agent assessment architecture in S2 sets up different categories of expert agents to perform risk assessment based on different enterprise risk factors through prompt words, including: Each type of expert agent includes multiple analysis agents, which collaborate to complete the risk factor scoring of their respective categories. The analysis agents are composed of a confident agent and two gentle agents. The confident agent tends to persuade other agents, and the gentle agents listen attentively to the answers of other agents.

[0007] Optionally, the risk assessment algorithm MA-ERC described in S2 assesses the initial risk of the enterprise through a multi-agent assessment architecture, including: In the first stage, the plurality of analysis agents respectively receive enterprise risk data, the enterprise risk data including enterprise financial characteristics, CEO characteristics and symptom risk characteristics, and generate a risk analysis report including an initial risk score for the enterprise; In the second phase, after receiving a risk analysis report containing the company's initial risk score, multiple analytical agents interact through debate, obtaining the analysis process and results of other analytical agents. Each analytical agent then shares its own analysis results based on the company's risk data. The debate focuses on the differences, and each analytical agent ultimately decides whether to modify its own score. In the third stage, each analytical agent conducts self-reflection based on the analysis process in the first stage and the debate process in the second stage; In the fourth stage, after the first round of debate and self-reflection, multiple analytical agents engage in a second round of debate and draw a final conclusion based on the results of this debate, which is the final output of the risk analysis report containing the initial risk score of the enterprise; Extract the enterprise's multi-dimensional risk score from the final output risk analysis report containing the enterprise's initial risk score, where the risk score of each dimension includes the scores of multiple expert agents. The risk scores of multiple dimensions are linearly weighted and fused to obtain the enterprise's comprehensive initial risk score. .

[0008] Optionally, the risk assessment algorithm MA-ERC assesses the initial risk of an enterprise through a multi-agent assessment architecture, and combines a personalized Pagerank algorithm to simulate risk propagation to quantify the propagation risk value between enterprises, including: Phase 1: In the constructed enterprise multi-dimensional risk knowledge map In the search, any enterprise node is found through the cross-capital relationship set between nodes Any other enterprise node directly or indirectly connected by any relationship in the relationship, the relationship between two enterprise nodes is defined as , stored in the relationship collection In; through the relationship set Calculate the risk weight of a single relationship, and then combine the risk weights of different types of relationships to obtain the final risk transmission probability between all enterprises ; The second stage: assign initial risk values ​​to all enterprise nodes and integrate weights Dynamically adjust the enterprise's comprehensive initial risk value output by the multi-agent assessment framework , get the initial risk value of each enterprise , after obtaining the initial risk of each enterprise and the probability of risk transmission Finally, the personalized Pagerank algorithm is used to simulate the risk propagation process, and multiple rounds of risk iteration are performed to obtain the final propagation risk value. .

[0009] Optionally, the multimodal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features, and a hybrid expert model is used for feature fusion to convert the multimodal enterprise data into feature vectors of nodes, including: The multimodal enterprise data includes enterprise multi-dimensional risk scores and enterprise multi-dimensional risk analysis report , wherein the enterprise multi-dimensional risk analysis report It is obtained by fusing the risk analysis reports containing the initial risk scores of enterprises output by multiple expert agents under each data type; The enterprise's multi-dimensional risk score is calculated by a feedforward neural network Mapping to features, using Finbert pre-trained language model to extract enterprise multi-dimensional risk analysis report Semantic features in text; A hybrid expert model is used to fuse features of two different modalities. The gating weights are used to dynamically weight the fusion expert outputs, and the output vectors of the hybrid expert model with different modal features are spliced ​​to convert multimodal enterprise data into feature vectors of nodes. The row vectors corresponding to the enterprise nodes in the original node embedding matrix are replaced to complete the reconstruction of enterprise node features.

[0010] Optional, for enterprise multi-dimensional risk scoring , which is converted into enterprise risk score features through a feedforward neural network consisting of two linear layers , expressed as: ; in, and is the weight matrix of the hidden layer and the output layer, and is the bias vector, is the activation function, and the enterprise risk score feature is obtained through FFN ; The combination of CLS vector strategy and average pooling strategy in Finbert is expressed as: ; in, Represents the final text features, is the output vector of Finbert, is the CLS eigenvector, is the mean vector of the entire sequence, is the actual length of the sequence, Represents the token position index in the sequence; Enterprise multi-dimensional risk analysis report Input Finbert to get the enterprise risk analysis report features , expressed as: ; in, represents the Finbert model; Attention-based feature fusion can dynamically learn the importance weights of different features and achieve adaptive feature fusion, which is expressed as: ; ; ; in, is the raw attention score, is the normalized attention weight, The data type is The characteristics of the integrated enterprise risk analysis report, 、 represents the weight matrix, 、 represents the bias vector, represents the natural exponential function, Indicates the Attention score for risk analysis report features; Characteristics of the enterprise risk analysis report after integration and enterprise risk scoring characteristics Features of two different modalities are fused using a hybrid expert model and gated weights It is calculated by Sofemax function and expressed as: ; in, is a set of features containing two different modalities, Indicates the The characteristics of the modal class, For the The weight matrix of the modal feature, the dimension is the input size and expert subset size The product of , the number of experts selected for each input is half of the total number of experts; For each input , the fusion output is weighted summed by the gate weights on the randomly selected expert subset outputs, expressed as: ; in, For size A randomly selected subset of expert indices, For the Expert networks, each of which consists of two layers of linear transformation and ReLU activation function. Indicates that the gating mechanism is assigned to The weight of the expert network, For the The fusion output vector of modal features, including comprehensive risk score features and comprehensive risk analysis report features , obtained by weighted summation of the selected expert networks; By splicing the output vectors of different modal features, the enterprise multimodal data is converted into the feature vector of the node. ; Finally, use the feature vector Replace the original node embedding matrix The row vector corresponding to the enterprise node in the data is used to complete the reconstruction of the enterprise node features.

[0011] Optionally, a multi-layer RGAT model is used to model heterogeneous relationships and dynamically learn the importance weights of neighbor nodes under different relationship types, including: Calculate the head nodes of different types of relationships With the tail node The additive attention score between quantifies the association strength between two nodes under a specific relationship and is expressed as: ; in, Represents the head node With the tail node In relationship The attention score in 、 and is the trainable weight matrix, represents a nonlinear activation function, represents the feature vector of the head node, represents the feature vector of the adjacent nodes, and For the relationship The query matrix and key matrix of represents the query vector, represents the key vector, For the edge The eigenvector of After calculating the attention scores of all different types of relations, we perform cross-relation attention normalization to globally compare the importance of different relations, expressed as: ; in, is the set of relations in the knowledge graph, express A relationship in express An adjacent node in For nodes The type is The set of adjacent nodes of the relationship, is the normalized attention coefficient, Indicates that in the relationship Next, node and adjacent nodes Attention score; The information of neighbor nodes is weighted and aggregated by the attention coefficient, and multi-head attention aggregation of different relationship types is completed, which is expressed as: ; in, represents the attention head index, represents the number of attention heads, express A relationship in represents adjacent nodes, Indicates the relationship With attention head A specific weight matrix, Representation node The type is The set of adjacent nodes of the relationship, Represents the head node and adjacent nodes Targeted relationships With attention head The attention coefficient, is the feature vector of the adjacent node; Focus on multiple Average aggregation and get the final output of node features , expressed as: ; in, It represents the final node feature representation obtained by fusing multiple attention head information. represents a nonlinear activation function, represents the number of attention heads, Representation node Aggregation result of the mth attention head.

[0012] Optionally, multiple layers of RGAT are stacked to obtain the final node features, the features corresponding to the enterprise nodes are extracted, and the enterprise's propagation risk value is mapped to a vector. The features corresponding to the enterprise nodes and the enterprise's propagation risk value are combined to obtain the enterprise's comprehensive features, and the model evaluation results are obtained through the linear layer, including: After passing the first level of RGAT, Use ReLU activation and input to the next layer of RGAT, and obtain the final node features by stacking multiple layers of RGAT ; The final node feature The corresponding enterprise node Extract it and calculate the enterprise's transmission risk value Map to vector, concatenate and To obtain comprehensive characteristics of the enterprise , Represents the final node features of the enterprise node, using the enterprise comprehensive features The evaluation result of the model is obtained through the linear layer, which is expressed as: ; in, represents the evaluation label of the model, represents the linear layer weight, Represents the comprehensive characteristics of the enterprise, Represents the bias vector.

[0013] After adopting the above technical solution, this application has the following beneficial effects compared with the prior art: In this application, a multi-agent architecture based on a large language model is combined with a graph-based risk propagation algorithm to more accurately and reasonably simulate the risk propagation process between enterprises. The propagation risk value calculated by this application can effectively improve the assessment accuracy of the risk assessment model; at the same time, this application provides new ideas for the combination of multi-agent based on a large language model with knowledge graphs and graph neural networks and other technologies.

[0014] In this application, by introducing multi-agent expert risk scoring, risk reporting and designing multimodal node feature fusion, enterprise nodes are given more realistic and differentiated features, effectively enhancing the model's discriminative ability.

[0015] In this application, an enterprise risk assessment interpretable graph EG is constructed to solve the black box problem existing in traditional machine learning, enabling securities regulators and investors to intuitively view the relationship between enterprise risks and violations, greatly improving the transparency and credibility of the model, and thus enhancing compliance management and risk prevention efficiency.

[0016] The specific implementation methods of the present application are further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are part of this application and are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application but do not constitute an undue limitation of this application. Obviously, the drawings described below are only some embodiments. For those of ordinary skill in the art, other drawings can be derived from these drawings without inventive effort.

[0018] In the attached figure: Figure 1 Schematic diagram of the process of enterprise risk assessment method based on multi-agent risk assessment algorithm MA-ERC in this specific implementation; Figure 2 Schematic diagram of the technical framework of the enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC in this specific implementation; Figure 3 Schematic diagram of the enterprise multi-dimensional risk knowledge graph framework of the enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC in this specific implementation; Figure 4 This is a diagram of enterprise multi-modal and multi-source risk data processing of an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC in this specific implementation; Figure 5 This is a flowchart of the MA-ERC algorithm of the enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC in this specific embodiment; Figure 6 This is the MNF-GNN model framework diagram of the enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC in this specific implementation. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application but are not used to limit the scope of the present application.

[0020] See Figure 1 and Figure 2 The present application provides an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC, comprising the following steps: S1. Constructing a multi-dimensional enterprise risk knowledge graph: At least one model layer is designed, wherein the model layer defines at least a core entity, a risk propagation analysis entity, and a risk assessment analysis entity and defines corresponding relationship types between the entities; S2. Design the risk assessment algorithm MA-ERC to assess the initial risk of enterprises and quantify the risk value of transmission between enterprises: The risk assessment algorithm MA-ERC uses a multi-agent evaluation framework to assess the initial risk of an enterprise, and combines it with a personalized Pagerank algorithm to simulate risk propagation to quantify the risk value propagated between enterprises. The multi-agent evaluation framework uses prompt words to set up different categories of expert agents to perform risk assessments based on different enterprise risk factors. S3. Constructing MNF-GNN model for enterprise risk assessment: The multimodal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features, and a hybrid expert model is used for feature fusion to convert the multimodal enterprise data into node feature vectors. A multi-layer RGAT model is used to model heterogeneous relationships and dynamically learn the importance weights of neighbor nodes under different relationship types. The final node features are obtained by stacking multiple layers of RGAT, and the features corresponding to the enterprise nodes are extracted. The enterprise's propagation risk value is mapped to a vector, and the features corresponding to the enterprise nodes and the enterprise's propagation risk value are spliced ​​to obtain the enterprise's comprehensive features. The model's evaluation results are obtained through a linear layer. S4. Construct an interpretable graph EG for enterprise risk assessment: By obtaining the enterprise multi-dimensional risk scores, enterprise multi-dimensional risk analysis reports and the risk transmission probability between enterprises, an enterprise risk assessment interpretable graph EG is constructed to provide an explainable description for the final assessment results.

[0021] It should be noted that the execution subject of the method in this embodiment is an evaluation device, which can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, etc., and the non-mobile electronic device can be a server and a personal computer, etc., which are not specifically limited in this application. The following describes the evaluation method in this embodiment using the execution subject as an example of a server.

[0022] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, the meaning of "plurality" is two or more, unless otherwise clearly specified.

[0023] It should be noted that an enterprise knowledge graph refers to a vertical field knowledge graph that focuses on enterprise information and relationships, and has a wide range of commercial applications. In the field of enterprise risk analysis, constructing an enterprise knowledge graph has become a commonly used method. At the same time, the composition of the knowledge graph will be different depending on the purpose to be achieved. There are two main purposes for constructing a knowledge graph in this embodiment. On the one hand, it is hoped that by constructing a knowledge graph to model the relationship between listed companies and simulate the flow of risks between listed companies, in order to quantify the concept of transmission risk. On the other hand, it is hoped that by constructing a knowledge graph to mine and encode the complex correlation characteristics between nodes such as enterprises, shareholders and cities, more effective information can be provided for risk assessment.

[0024] See Figure 3 , based on these two purposes, and referring to the existing research on knowledge graphs for enterprise risk analysis. This embodiment designs the model layer of the knowledge graph, abstracting and modeling entities and relationships around the two core elements of risk propagation and risk assessment. In addition to the enterprise entity, three types of entities, namely shareholders, executives, and controllers, are defined to analyze the external structure of the enterprise and the equity control path to simulate risk propagation; two types of entities, namely industries and cities, are defined to explore the impact of macro-environmental policy factors on corporate violations; and the entity type of sign risk events is defined to describe the risk signs disclosed by the enterprise in the current year, and the corresponding relationship types between entities are defined. The enterprise multi-dimensional risk knowledge graph designed in this embodiment is as follows: Figure 3 The entity and relationship types of this embodiment are shown in Table 1: Table 1: Entity-relationship table of enterprise multidimensional risk knowledge graph

[0025] Please continue to see Figure 3 According to the designed model layer, this embodiment collects the required information of listed companies in 2022 from the CSMAR database, including information on shareholders, executives, cities, industries, and overseas investments of listed companies. It also crawls the risk signs reports of companies from the Sina Finance website to form triples and construct a multi-dimensional risk knowledge graph for companies. The constructed graph is saved in neo4j. The number of different types of nodes and relationships in the multi-dimensional risk knowledge graph of companies is shown in Table 2: Table 2: Enterprise multi-dimensional risk knowledge graph scale

[0026] It's important to note that risk propagation has long been a hot topic in enterprise risk assessment. With the widespread adoption of knowledge graphs, they have revealed the risk propagation pathways between enterprises, through various channels such as executive management and equity. Based on this, the Personalized PageRank algorithm, a variant of the PageRank algorithm, leverages its biased random walk mechanism to more accurately simulate the risk propagation process within enterprise networks and has been widely used in enterprise risk propagation analysis. By simulating the risk propagation process within a network, it assesses the influence of each node in risk propagation and, consequently, quantifies the enterprise's propagation risk. However, existing methods focus on optimizing the propagation mechanism of the original PPR algorithm. Initial node risk values ​​are generally assigned using the mean or a single risk factor, which clearly does not fully reflect the enterprise's true risk. In fact, the initial value is more important for the Personalized PageRank algorithm than for the original PPR algorithm. In the PPR algorithm, the initial value is the source of risk. Each subsequent random jump has a probability of returning to the preset initial node distribution. Therefore, it not only determines the set of nodes the algorithm focuses on but also influences the distribution of risk influence within the network. If a company's initial risk value doesn't represent its true risk, the more rational the propagation mechanism, the further the final result deviates from reality, amplifying false risk signals. Therefore, the proper setting and optimization of the initial risk value becomes a key factor influencing the effectiveness of the PPR algorithm. Traditional methods, which rely on average or single-dimensional assignments, ignore the multidimensionality and complexity of corporate risk.

[0027] Based on this, this embodiment proposes the MA-ERC algorithm to assess the initial risk of an enterprise through a multi-agent evaluation framework, and combines it with a personalized Pagerank algorithm to simulate risk propagation to quantify the risk value propagated between enterprises. The multi-agent evaluation framework introduces multimodal enterprise risk data to integrate and quantify the initial risk value of the enterprise, making it closer to the enterprise's actual risk status. The input of the multi-agent evaluation framework includes three enterprise risk factors: financial characteristics, CEO characteristics, and symptomatic risk characteristics. Three types of expert agents are set up for different risk factors through prompt words to assess the risk level of the enterprise.

[0028] See Figure 2 、 Figure 4 and Figure 5The MA-ERC algorithm inputs a company's financial characteristics, CEO characteristics, and symptomatic risk characteristics. Financial characteristics primarily include structured data consisting of economic indicators reflecting a company's profitability (cost-to-profit ratio), growth potential (total asset growth rate), and solvency (retained earnings to assets ratio, cash ratio, and tangible assets-to-liabilities ratio). For CEO characteristics, five dimensions of data, including age, gender, education background, salary, and whether a CEO holds a part-time job, are selected to identify the CEO's greatest influence on corporate violations. For symptomatic risk characteristics, an LLM-based summary agent is used to process enterprise risk event text crawled from Sina Finance, extracting high-frequency risks and long-term impacts. Because enterprise risk events involve basic enterprise information, this example uses an LLM to analyze historical enterprise data. Therefore, to ensure that the LLM does not introduce advanced information, event desensitization is performed. Using a summary agent for summarization and reasoning can filter out noisy information, highlight core trend changes, and mine potential information to extract high-value features.

[0029] Specifically, each type of expert agent consists of three analytical agents, which collaborate to complete the risk factor scoring for that type of expert agent. The three analytical agents consist of one confident agent and two moderate agents. The former tends to persuade the other agents to believe in themselves, while the latter focuses on listening to the other agents' answers.

[0030] The multi-agent initial risk assessment process consists of four main stages. In the first stage, the three analytical agents receive enterprise risk data, namely the enterprise's financial characteristics, CEO characteristics, and symptomatic risk characteristics, and generate a risk analysis report containing the enterprise's initial risk score. The higher the score, the higher the risk level of the enterprise in that dimension. In the second phase, after receiving the risk analysis report containing the company's initial risk score, multiple analytical agents interact through debate to learn about the analysis processes and results of other analytical agents. The purpose of this debate is to achieve analysis sharing and reach consensus, allowing each analytical agent to share its own analysis results based on the company's risk data. The results of different analytical agents often differ, and the debate process revolves around these differences. Ultimately, each agent decides whether to modify its own score. In the third stage, each analytical agent conducts self-reflection based on the analysis process in the first stage and the debate process in the second stage, determining whether its analysis report still has deficiencies that require revision. This reflection mechanism mimics human thinking. By re-examining its own analysis process, iteratively raising questions and verifying them against the knowledge base, the agent can locate and correct errors, effectively addressing the potential hallucination problem of large language models. In the fourth stage, after the first round of debate and self-reflection, multiple analytical agents engage in a second round of debate, drawing a final conclusion based on the results of this debate. The goal of this second round of debate is for each agent to share their own process of correcting errors and filling logical gaps based on debate and reflection. This cross-validation network allows agents to discover hidden errors that they themselves may have missed, ultimately leading to a collective consensus.

[0031] Specifically, in the first stage, the enterprise risk data is input and a risk analysis report containing the enterprise's initial risk score is generated by the analysis agent. , expressed as: ; in, For the company being analyzed, is the input feature, It is a data type set and includes three data types: financial risk, CEO risk, and symptom risk, which can be expressed as ,in, Represents the financial risk data type, Indicates CEO risk data type, Indicates the symptom risk data type, It is one of the data types. To analyze the collection of agents, In order to enable the expert agent to analyze the data and specify the risk level prompt words represented by the risk value, represents the i-th analysis agent; In the second stage, multiple analysis agents conduct the first round of debate, which can be expressed as: ; in, The risk analysis report containing the initial risk score of the enterprise is modified by the analysis agent after the debate. In order to prompt the analysis agent to conduct debate, the input of the analysis agent is , and They represent the risk analysis reports of the other two analysis agents in the first phase, which contain the initial risk scores of the enterprises; In the third stage, multiple analytical agents reflect separately, which can be expressed as: ; in, It represents the risk analysis report containing the initial risk score of the enterprise, which has been revised by the analysis agent after reflection. Indicates the prompt words that make the analytical agent reflective behavior; In the fourth stage, multiple analysis agents conduct a second round of debate, which can be expressed as: ; in, The final output of the expert agent is the initial enterprise risk analysis report, including the financial risk analysis report , CEO Risk Analysis Report and Symptom Risk Analysis Report ; and represents the initial risk analysis report of the other two analysis agents after the third stage. Represents the prompt words that enable the analysis agent to perform debate behavior; The final output report can be used to extract the enterprise multi-dimensional risk scores of multiple analytical agents. , including financial risk scores , CEO Risk Score Symptom risk score , where each dimension risk factor score includes the scores of three expert agents, and the risk scores of these three dimensions are linearly weighted and fused to obtain the comprehensive initial risk score of the enterprise , expressed as: ; in, 、 、 is the linear fusion weight, is the number of expert agents, where The value is 3, is the final comprehensive initial risk value of the enterprise, 、 、 Represent the scores of individual financial, CEO and symptom risk analysis agents respectively.

[0032] Calculate the probability of risk transmission between enterprises and build a multi-dimensional risk knowledge graph for enterprises middle. exist Figure 5 It is represented as a heterogeneous graph, with seven different colors (cyan-enterprise, green-shareholder, pink-executive, blue-actual controller, yellow-industry, purple-city, red-risk event) and seven different types of nodes. The cross-capital relationship set between nodes is defined as ,in, Represents different types of relationships between enterprise nodes and other nodes. For enterprise multidimensional risk knowledge graph Any enterprise node in , find the corresponding Other enterprise nodes directly or indirectly connected by any relationship , the relationship between nodes is defined as , stored in the collection middle.

[0033] For the three types of relationships, investment, shareholding and controlling, if it is a direct relationship, the shareholding ratio / investment amount is used as the relationship weight If it is an indirect relationship, the weight is defined by combining the investment amounts of the two companies or finding the largest shareholding ratio in the relationship path between the two companies. . Set the threshold and , according to the weight Filter out all relationships with weights above the threshold and save them in the relationship collection middle.

[0034] Getting the relationship set Finally, the risk weight of a single relationship is calculated. The purpose of this step is to quantify the impact of different relationships on the intensity of risk transmission between enterprises, and then integrate the cross-capital relationships with different risk weights to obtain the probability of risk transmission between enterprises. The formula for calculating the risk weight of a single relationship is as follows: ; in, It refers to the annual Identify risky behavior, It refers to the annual Identify risky behavior, is an indicator function, the value is 1 when the condition is met, otherwise it is 0. Is the relationship type The total number of records after filtering. For one of the records. If there is a bidirectional "risk in the same year → risk of related enterprises in the next year" pattern in the relationship, it indicates that there is a possibility of risk transmission in this type of relationship. The probability of occurrence of this pattern in this type of relationship is calculated, and the relationship type is obtained. Risk weight .

[0035] Afterwards, since there are often multiple different capital cross-correlations between enterprises, the risk weights of different types of relationships are combined to obtain the final risk transmission probability between all enterprises. , the formula is as follows: ; in, For relationship type The risk weight of the relationship, for The total number of relationships in .

[0036] After completing the first phase, the next step is to assign initial risk values ​​to all enterprise nodes. The formula for calculating the initial risk value of each enterprise is as follows: ; in, For enterprises The initial risk value of the enterprise obtained by multiple analytical agents is divided by 100 to scale it to the range of 0-1; Indicates whether the enterprise has engaged in risky behavior in the previous year. If yes, the value is 1, otherwise it is 0. is the fusion weight of the two initial risks, Dynamically adjust the fusion ratio of the two initial risks and ultimately obtain the initial risk value of the enterprise .

[0037] Getting the initial risk value and the probability of risk transmission Finally, the risk propagation process is simulated and multiple rounds of risk iteration are performed. The calculation formula is as follows: ; in, It is a dynamic risk adjustment factor. Unlike the original damping factor, it dynamically adjusts the attenuation ratio of the risk weight with each risk iteration round to prevent the risk of high-risk nodes from being overly diluted during the iteration process. is the iteration round, represents the maximum number of iterations, express The index of the adjacent node in , For iteration The final transmission risk value is Represents iteration +1 times the final transmission risk value.

[0038] The obtained transmission risk value is input into the machine learning model as an additional one-dimensional feature, and compared with the machine learning model whose input only contains the economic characteristics of the enterprise. The experimental results of the machine learning effect enhancement after adding the transmission risk value are shown in Table 3: Table 3: Increase in machine learning effect after adding risk value

[0039] See Figure 1 、 Figure 4 and Figure 6, construct the MNF-GNN model for enterprise risk assessment: embed the multimodal enterprise data output by multiple agents in S2 into features, use the hybrid expert model MOE for feature fusion, and transform the multimodal enterprise data into a comprehensive feature vector of multimodal enterprise nodes. Figure 6 Heterogeneous graph In, with Figure 5 The same, different colors correspond to different types of nodes. At the same time, use Represents different types of relationships between enterprise nodes and other nodes. See Table 1 for details. A multi-layer RGAT model is used to model these relationships. The importance weights of neighbor nodes under different relationship types are dynamically learned. The final node features are obtained by stacking multiple layers of RGAT. The features corresponding to the enterprise node are extracted, and the enterprise's propagation risk value is mapped to a vector. These two features are concatenated to obtain the enterprise's comprehensive features, and the model's evaluation results are obtained through a linear layer.

[0040] Specifically, for the enterprise's multi-dimensional risk scoring and enterprise multi-dimensional risk analysis report , which is embedded as a feature through the multimodal node feature encoder involved. For enterprise multi-dimensional risk scoring , through a feedforward neural network consisting of two linear layers, it is converted into enterprise risk score features, as shown in the following formula: ; in, and is the weight matrix of the hidden layer and the output layer. and is the bias vector, is the activation function, and the enterprise risk score feature is obtained through this FFN .

[0041] Due to the enterprise's multi-dimensional risk analysis report For text data containing a large amount of financial terms, the Finbert model is used to extract semantic features from the text. At the same time, in order to fully explore the global semantic information and local semantic features in the analysis report, the CLS vector strategy and the average pooling strategy are combined in Finbert, as shown in the following formula: ; in, is the output vector of Finbert, is the CLS eigenvector, is the mean vector of the entire sequence, is the actual length of the sequence.

[0042] Enterprise multi-dimensional risk analysis report enter Get the characteristics of enterprise risk analysis report , as shown below: ; in, represents the Finbert model, Represents a multi-dimensional risk analysis report for an enterprise.

[0043] because Contains and Two different types of features need to be integrated to represent the characteristics of the enterprise risk analysis report Attention-based feature fusion can dynamically learn the importance weights of different features and achieve adaptive feature fusion, as shown in the following formula: ; ; ; in, is the raw attention score, is the normalized attention weight, The characteristics of the enterprise risk analysis report after integration, 、 represents the weight matrix, 、 represents the bias vector, represents the natural exponential function, Indicates the Attention scores for class risk analysis report features.

[0044] Characteristics of the enterprise risk analysis report after integration And the enterprise risk score characteristics obtained by numerical score mapping The features of two different modalities are fused using the hybrid expert model MOE. MOE has multiple expert networks and a modality-independent gating mechanism. For features of different modalities, it is necessary to dynamically weight the fusion expert output through gating weights. It is calculated by the Sofemax function, as shown below: ; in, is a set of three different modal features, Indicates the The characteristics of the modal class, For the The weight matrix of the modal feature, the dimension is the input size and expert subset size The number of experts selected for each input is half of the total number of experts.

[0045] For each input , the fusion output is weighted summed by the gated weights on the randomly selected expert subset outputs, as shown in the following formula: ; in, For size A randomly selected subset of expert indices, For the Expert networks, each expert consists of two layers of linear transformation and ReLU activation function, For the The fusion output vector of modal features, including comprehensive risk score features and comprehensive risk analysis report features , obtained by weighted summation of the selected experts.

[0046] By splicing the output vectors of different modal features, the enterprise multimodal data is converted into a comprehensive feature vector of a multimodal enterprise node. ; Last used Replace the original node embedding matrix The row vector corresponding to the enterprise node in the data is used to complete the reconstruction of the enterprise node features.

[0047] Specifically, based on the constructed enterprise multi-dimensional risk knowledge graph and the original node embedding matrix , using the multi-layer RGAT model to map the enterprise's multi-dimensional risk knowledge The heterogeneous relationships in the model are modeled, and the importance weights of neighbor nodes under different relationship types are dynamically learned to obtain more accurate and comprehensive node features. First, the head nodes in different types of relationships are calculated. With the tail node The additive attention score between quantifies the association strength between two nodes under a specific relationship, as shown in the following formula: ; in, 、 and is the trainable weight matrix, represents a nonlinear activation function, represents the feature vector of the head node, represents the feature vector of the adjacent nodes, and For the relationship The query matrix and key matrix of represents the query vector, represents the key vector, For the edge The eigenvector of .

[0048] After calculating the attention scores of all different types of relations, cross-relationship attention normalization is performed to globally compare the importance of different relations, as shown in the following formula: ; in, is the set of relations in the knowledge graph, express A relationship in express An adjacent node in For nodes The type is The set of adjacent nodes of the relationship, is the normalized attention coefficient.

[0049] Then, the information of neighbor nodes is weighted and aggregated by the attention coefficient, and multi-head attention aggregation of different relationship types is completed, as shown in the following formula: ; in, represents the attention head index, represents the number of attention heads, express A relationship in represents adjacent nodes, Indicates the relationship With attention head A specific weight matrix, Represents the head node and adjacent nodes Targeted relationships With attention head The attention coefficient, is the feature vector of the adjacent node.

[0050] Focus on multiple Average aggregation and get the final output of node features , as shown below: ; in, represents a nonlinear activation function, represents the number of attention heads, Representation node Aggregation result of the mth attention head.

[0051] In order to further aggregate the global features of multi-hop neighbor nodes, after passing through the first layer of RGAT, Use ReLU activation and input to the next layer of RGAT, and obtain the final node features by stacking multiple layers of RGAT .Will The features corresponding to the enterprise nodes are extracted and the enterprise's propagation risk value is calculated. Mapped into vectors, these two features are concatenated to obtain the comprehensive characteristics of the enterprise , Represents the final node characteristics of the enterprise node, using The evaluation result of the model is obtained through the linear layer, as shown below ; in, represents the evaluation label of the model, represents the linear layer weight, Represents the comprehensive characteristics of the enterprise, Represents the bias vector.

[0052] After obtaining the evaluation results, the binary cross entropy loss function is used to calculate the length of Evaluation results With the real label of the enterprise The loss between , so that the model can be continuously optimized iteratively, as shown in the following formula: ; in, Represents the calculated loss value, represents the number of samples, Indicates the The true labels of samples, Indicates the The evaluation labels of samples, Represents the logarithmic function.

[0053] As a specific implementation, the method proposed in this example was compared with existing mainstream enterprise risk assessment methods to comprehensively evaluate their performance under different feature perspectives and modeling strategies. These models primarily include representatives from the following three categories: a machine learning model that evaluates using the enterprise's structured data; a graph neural network model that evaluates using the enterprise's multidimensional risk knowledge graph; and an enterprise risk analysis model that evaluates using multi-source data fusion. The comparative experimental results are shown in Table 4 below: Table 4: Comparative experimental results

[0054] Constructing an interpretable graph (EG) for enterprise risk assessment: Each enterprise node is connected to its own financial, CEO, and symptom risk score nodes via a score relationship. Each score node is connected to its corresponding cause node via a cause relationship, reflecting the enterprise's risk. The content of the cause node represents the conclusion of the multi-agent's final risk analysis report, i.e., the agent's risk score rationale, as shown in Table 5.

[0055] Table 5: Example of agent risk scoring reasons

[0056] Furthermore, each enterprise is connected to nodes with risk transmission channels through a risk transmission relationship, with the risk transmission probability serving as the edge weight. Each node has an attribute evaluation value, representing the risk assessment model's assessment result. By analyzing the node attributes, edge weights, and causal relationships along these paths, an interpretable risk transmission chain can be constructed, revealing the key factors influencing the model's assessment results. For example, by querying the upstream transmission path of a high-risk enterprise, external enterprise nodes that significantly influence its score can be identified to determine whether the model's high-risk rating is due to its association with multiple high-risk enterprises. Furthermore, the model can focus on the enterprise's own scoring nodes to identify specific risk points, such as financial anomalies, CEO changes, and negative public opinion, thereby achieving an abductive explanation from "assessment results" to "risk sources."

[0057] The above is only a preferred embodiment of the present application and does not constitute any form of limitation to the present application. Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the present application can make some changes or modifications to equivalent embodiments with equivalent changes using the technical content suggested above without departing from the scope of the technical solution of the present application. The implementation schemes in the above embodiments can also be further combined or replaced. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present application that do not depart from the content of the technical solution of the present application still fall within the scope of the solution of the present application.

Claims

1. The enterprise risk assessment method based on the multi-agent risk assessment algorithm MA-ERC is characterized by: The following steps are involved: S1. Constructing a multi-dimensional enterprise risk knowledge graph: At least one model layer is designed, wherein the model layer defines at least a core entity, a risk propagation analysis entity, and a risk assessment analysis entity and defines corresponding relationship types between the entities; S2. Design the risk assessment algorithm MA-ERC to assess the initial risk of enterprises and quantify the risk value of transmission between enterprises: The risk assessment algorithm MA-ERC uses a multi-agent evaluation framework to assess the initial risk of an enterprise, and combines it with a personalized Pagerank algorithm to simulate risk propagation to quantify the risk value propagated between enterprises. The multi-agent evaluation framework uses prompt words to set up different categories of expert agents to perform risk assessments based on different enterprise risk factors. S3. Constructing MNF-GNN model for enterprise risk assessment: The multimodal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features. A hybrid expert model is used for feature fusion to convert the multimodal enterprise data into feature vectors of nodes. A multi-layer RGAT model is used to model heterogeneous relationships and dynamically learn the importance weights of neighbor nodes under different relationship types. The final node features are obtained by stacking multiple layers of RGAT, and the features corresponding to the enterprise nodes are extracted. The enterprise's propagation risk value is mapped to a vector. The features corresponding to the enterprise nodes and the enterprise's propagation risk value are concatenated to obtain the enterprise's comprehensive features, and the model evaluation results are obtained through the linear layer. S4. Construct an interpretable graph EG for enterprise risk assessment: By obtaining the enterprise multi-dimensional risk scores, enterprise multi-dimensional risk analysis reports and the risk transmission probability between enterprises, an enterprise risk assessment interpretable graph EG is constructed to provide an explainable description for the final assessment results.

2. The method according to claim 1, characterized in that The model layer described in S1 at least defines the core entity, risk propagation analysis entity, and risk assessment analysis entity and defines the corresponding relationship types between each entity, including: The core entities include enterprise entities, the risk propagation analysis entities include three types of entities: shareholders, executives and controllers, which are used to analyze the external structure and equity control path of the enterprise to simulate risk propagation, the risk assessment analysis entities include two types of entities: industry and city, which are used to explore the impact of macro-environmental policy factors on enterprise risks, and also include sign risk event entities, which are used to describe the risk signs disclosed by the enterprise in the current year.

3. The method according to claim 1, characterized in that The multi-agent assessment architecture in S2 uses different categories of expert agents to perform risk assessments based on different enterprise risk factors using prompts, including: Each type of expert agent includes multiple analysis agents, which collaborate to complete the risk factor scoring of their respective categories. The analysis agents are composed of a confident agent and two gentle agents. The confident agent tends to persuade other agents, and the gentle agents listen attentively to the answers of other agents.

4. The method according to claim 3, characterized in that The risk assessment algorithm MA-ERC described in S2 assesses the initial risk of an enterprise through a multi-agent assessment architecture, including: In the first stage, the plurality of analysis agents respectively receive enterprise risk data, the enterprise risk data including enterprise financial characteristics, CEO characteristics and symptom risk characteristics, and generate a risk analysis report including an initial risk score for the enterprise; In the second phase, after receiving a risk analysis report containing the company's initial risk score, multiple analytical agents interact through debate, obtaining the analysis process and results of other analytical agents. Each analytical agent then shares its own analysis results based on the company's risk data. The debate focuses on the differences, and each analytical agent ultimately decides whether to modify its own score. In the third stage, each analytical agent conducts self-reflection based on the analysis process in the first stage and the debate process in the second stage; In the fourth stage, after the first round of debate and self-reflection, multiple analytical agents engage in a second round of debate and draw a final conclusion based on the results of this debate, which is the final output of the risk analysis report containing the initial risk score of the enterprise; Extract the enterprise's multi-dimensional risk score from the final output risk analysis report containing the enterprise's initial risk score, where the risk score of each dimension includes the scores of multiple expert agents. The risk scores of multiple dimensions are linearly weighted and fused to obtain the enterprise's comprehensive initial risk score. .

5. The method according to claim 4, characterized in that The risk assessment algorithm MA-ERC uses a multi-agent evaluation framework to assess the initial risk of an enterprise and combines it with a personalized Pagerank algorithm to simulate risk propagation to quantify the risk value propagated between enterprises, including: Phase 1: In the constructed enterprise multi-dimensional risk knowledge map In the search, any enterprise node is found through the cross-capital relationship set between nodes Any other enterprise node directly or indirectly connected by any relationship in the relationship, the relationship between two enterprise nodes is defined as , stored in the relationship collection In; through the relationship set Calculate the risk weight of a single relationship, and then combine the risk weights of different types of relationships to obtain the final risk transmission probability between all enterprises ; The second stage: assign initial risk values ​​to all enterprise nodes and integrate weights Dynamically adjust the enterprise's comprehensive initial risk value output by the multi-agent assessment framework , get the initial risk value of each enterprise , after obtaining the initial risk of each enterprise and the probability of risk transmission Finally, the personalized Pagerank algorithm is used to simulate the risk propagation process, and multiple rounds of risk iteration are performed to obtain the final propagation risk value. .

6. The method according to claim 4, characterized in that The multimodal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features. A hybrid expert model is used for feature fusion to transform the multimodal enterprise data into feature vectors of nodes, including: The multimodal enterprise data includes enterprise multi-dimensional risk scores and enterprise multi-dimensional risk analysis report , wherein the enterprise multi-dimensional risk analysis report It is obtained by fusing the risk analysis reports containing the initial risk scores of enterprises output by multiple expert agents under each data type; The enterprise's multi-dimensional risk score is calculated by a feedforward neural network Mapping to features, using Finbert pre-trained language model to extract enterprise multi-dimensional risk analysis report Semantic features in text; A hybrid expert model is used to fuse features of two different modalities. The gating weights are used to dynamically weight the fusion expert outputs, and the output vectors of the hybrid expert model with different modal features are spliced ​​to convert multimodal enterprise data into feature vectors of nodes. The row vectors corresponding to the enterprise nodes in the original node embedding matrix are replaced to complete the reconstruction of enterprise node features.

7. The method according to claim 6, characterized in that Multi-dimensional risk scoring for enterprises , which is converted into enterprise risk score features through a feedforward neural network consisting of two linear layers , expressed as: ； in, and is the weight matrix of the hidden layer and the output layer, and is the bias vector, is the activation function, and the enterprise risk score feature is obtained through FFN ; The combination of CLS vector strategy and average pooling strategy in Finbert is expressed as: ； in, Represents the final text features, is the output vector of Finbert, is the CLS eigenvector, is the mean vector of the entire sequence, is the actual length of the sequence, Represents the token position index in the sequence; Enterprise multi-dimensional risk analysis report Input Finbert to get the enterprise risk analysis report features , expressed as: ； in, represents the Finbert model; Attention-based feature fusion can dynamically learn the importance weights of different features and achieve adaptive feature fusion, which is expressed as: ； ； ； in, is the raw attention score, is the normalized attention weight, The data type is The characteristics of the integrated enterprise risk analysis report, 、 represents the weight matrix, 、 represents the bias vector, represents the natural exponential function, Indicates the Attention score for risk analysis report features; Characteristics of the enterprise risk analysis report after integration and enterprise risk scoring characteristics Features of two different modalities are fused using a hybrid expert model and gated weights It is calculated by Sofemax function and expressed as: ； in, is a set of features containing two different modalities, Indicates the The characteristics of the class modality, For the The weight matrix of the modal feature, the dimension is the input size and expert subset size The product of , the number of experts selected for each input is half of the total number of experts; For each input , the fusion output is weighted summed by the gate weights on the randomly selected expert subset outputs, expressed as: ； in, For size A randomly selected subset of expert indices, For the Expert networks, each of which consists of two layers of linear transformation and Relu activation function. Indicates that the gating mechanism is assigned to The weight of the expert network, For the The fusion output vector of modal features, including comprehensive risk score features and comprehensive risk analysis report features , obtained by weighted summation of the selected expert networks; By splicing the output vectors of different modal features, the enterprise multimodal data is converted into the feature vector of the node. ; Finally, use the feature vector Replace the original node embedding matrix The row vector corresponding to the enterprise node in the data is used to complete the reconstruction of the enterprise node features.

8. The method according to claim 5, characterized in that A multi-layer RGAT model is used to model heterogeneous relationships and dynamically learn the importance weights of neighbor nodes under different relationship types, including: Calculate the head nodes of different types of relationships With the tail node The additive attention score between quantifies the association strength between two nodes under a specific relationship and is expressed as: ； in, Indicates the head node With the tail node In relationship The attention score in 、 and is the trainable weight matrix, represents a nonlinear activation function, represents the feature vector of the head node, represents the feature vector of the adjacent nodes, and For the relationship The query matrix and key matrix of represents the query vector, represents the key vector, For the edge The eigenvector of After calculating the attention scores of all different types of relations, we perform cross-relation attention normalization to globally compare the importance of different relations, expressed as: ； in, is the set of relations in the knowledge graph, express A relationship in express An adjacent node in For nodes The type is The set of adjacent nodes of the relationship, is the normalized attention coefficient, Indicates that in the relationship Next, node and adjacent nodes Attention score; The information of neighbor nodes is weighted and aggregated by the attention coefficient, and multi-head attention aggregation of different relationship types is completed, which is expressed as: ； in, represents the attention head index, represents the number of attention heads, express A relationship in represents adjacent nodes, Indicates a relationship With attention head A specific weight matrix, Representation node The type is The set of adjacent nodes of the relationship, Represents the head node and adjacent nodes Targeted relationships With attention head The attention coefficient, is the feature vector of the adjacent node; Focus on multiple Average aggregation and get the final output of node features , expressed as: ； in, It represents the final node feature representation obtained by fusing multiple attention head information. represents a nonlinear activation function, represents the number of attention heads, Representation node Aggregation result of the mth attention head.

9. The method according to claim 8, characterized in that By stacking multiple layers of RGAT, we obtain the final node features, extract the features corresponding to the enterprise nodes, and map the enterprise's propagation risk value into a vector. The features corresponding to the enterprise nodes and the enterprise's propagation risk value are combined to obtain the enterprise's comprehensive features. The model's evaluation results are obtained through the linear layer, including: After passing the first level of RGAT, Use ReLU activation and input to the next layer of RGAT, and obtain the final node features by stacking multiple layers of RGAT ; The final node feature The corresponding enterprise node Extract it and calculate the enterprise's transmission risk value Map to vector, concatenate and To obtain comprehensive characteristics of the enterprise , Represents the final node features of the enterprise node, using the enterprise comprehensive features The evaluation result of the model is obtained through the linear layer, which is expressed as: ； in, represents the evaluation label of the model, represents the linear layer weight, Represents the comprehensive characteristics of the enterprise, Represents the bias vector.

Images