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

Abstract

The application relates to the field of data evaluation, and discloses an enterprise risk evaluation method based on a multi-agent risk evaluation algorithm MA-ERC, which comprises the following steps: S1, constructing an enterprise multi-dimensional risk knowledge graph; S2, designing a risk evaluation algorithm MA-ERC to evaluate the initial risk of an enterprise and quantifying the propagation risk value between enterprises; S3, constructing an MNF-GNN model to evaluate the risk of an enterprise; and S4, constructing an enterprise risk evaluation explainable graph EG. The application combines a multi-agent architecture based on a large language model with a graph-based risk propagation algorithm, more accurately and reasonably simulates the risk propagation process between enterprises, the calculated propagation risk value can effectively improve the evaluation accuracy of a risk evaluation model, and provides a new idea for combining a multi-agent based on a large language model with a knowledge graph and a graph neural network and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data evaluation, and particularly relates to an enterprise risk evaluation method based on a multi-agent risk evaluation algorithm MA-ERC. BACKGROUND

[0002] Most of the current mainstream enterprise risk evaluation technologies construct classification models based on structured information such as enterprise financial data, however, such methods are limited to the internal perspective of the enterprise and fail to consider the extensive connections between enterprises. In this context, knowledge graphs, as an effective tool 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 the following technical difficulties. First, how to most reasonably quantify the self-risk of an enterprise before using a graph propagation algorithm to simulate the risk propagation process between enterprises? The risk factors of an enterprise are multi-source and multi-modal, and a reasonable initial risk value of an enterprise is a prerequisite for correctly simulating risk propagation. Second, traditional graph neural networks often focus on the relationship between nodes and compress node attributes into fixed vectors, relying too much on flattened node features. Finally, although graph neural networks (GNN) perform excellently in graph embedding in the financial field, GNN is still essentially a black box model, and the evaluation results of the model often lack interpretability, making it difficult to be directly applied in the financial industry. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art. The present application provides an enterprise risk evaluation method based on a multi-agent risk evaluation algorithm MA-ERC.

[0004] To achieve the above-mentioned purpose, the present application provides an enterprise risk evaluation method based on a multi-agent risk evaluation algorithm MA-ERC, comprising the following steps:

[0005] S1, constructing an enterprise multi-dimensional risk knowledge graph:

[0006] at least comprising designing a schema layer, the schema layer defining at least core entities, risk propagation analysis entities and risk evaluation analysis entities and defining corresponding relationship types between the entities;

[0007] S2, designing a risk evaluation algorithm MA-ERC to evaluate the initial risk of an enterprise and quantify the propagation risk value between enterprises:

[0008] The risk evaluation algorithm MA-ERC evaluates the initial risk of an enterprise through a multi-agent evaluation architecture, simulates risk propagation by combining a personalized Pagerank algorithm to quantify the propagation risk value between enterprises, and the multi-agent evaluation architecture sets up different categories of expert agents for risk evaluation through prompt words according to different enterprise risk factors.

[0009] S3, constructing an MNF-GNN model for enterprise risk assessment:

[0010] Embedding the multi-modal enterprise data output by the multi-agent evaluation architecture in S2 as features, using a hybrid expert model for feature fusion to convert the multi-modal enterprise data into a feature vector of the node; using a multi-layer RGAT model to model the heterogeneous relationship and dynamically learn the importance weight of the neighbor node under different relationship types; obtaining the final node features through multi-layer RGAT stacking, extracting the features corresponding to the enterprise nodes, and mapping the enterprise propagation risk value into a vector, and splicing the features corresponding to the enterprise nodes and the enterprise propagation risk value to obtain the enterprise comprehensive features, and obtaining the evaluation result of the model through the linear layer;

[0011] S4, constructing an enterprise risk assessment explainable graph EG:

[0012] By obtaining the enterprise multi-dimensional risk score, the enterprise multi-dimensional risk analysis report, and the risk propagation probability between enterprises, an enterprise risk assessment explainable graph EG is constructed to provide an explainable explanation for the final evaluation result.

[0013] Optionally, the schema layer in S1 defines at least core entities, risk propagation analysis entities, and risk assessment analysis entities and defines corresponding relationship types between the entities, including:

[0014] The core entities include enterprise entities, the risk propagation analysis entities include three types of entities, namely, shareholders, senior managers, and controllers, which are used to analyze the enterprise external structure and the equity control path to simulate risk propagation, the risk assessment analysis entities include two types of entities, namely, industries and cities, which are used to mine the influence of macro-environmental policy factors on enterprise risk, and the symptom risk event entity is used to describe the risk symptoms disclosed by the enterprise in the current year.

[0015] Optionally, the multi-agent evaluation architecture in S2 sets different types of expert agents for risk assessment through prompt words for different enterprise risk factors, including:

[0016] Each type of expert agent includes multiple analysis agents, and each type of risk factor score is completed by cooperation of multiple analysis agents, wherein the analysis agent is composed of one confident agent and two moderate agents, the confident agent tends to persuade other agents, and the moderate agent focuses on listening to the answers of other agents.

[0017] Optionally, the risk assessment algorithm MA-ERC in S2 evaluates the initial risk of the enterprise through the multi-agent evaluation architecture, including:

[0018] In the first stage, multiple analytical agents receive enterprise risk data, which includes enterprise financial characteristics, CEO characteristics, and symptom risk characteristics, and generate a risk analysis report containing an initial risk score for the enterprise.

[0019] In the second stage, after obtaining the risk analysis report containing the initial risk score of the enterprise, multiple analytical agents interact through debate to obtain the analysis process and results of other analytical agents. Each analytical agent shares its own analysis results based on the enterprise risk data. The debate process revolves around the differences, and each analytical agent decides in the end whether to modify its own score.

[0020] 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.

[0021] 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 a risk analysis report containing the company's initial risk score.

[0022] The risk analysis report, which includes the company's initial risk score, extracts a multi-dimensional risk score for the company. Each dimension's risk score comprises scores from multiple expert agents. These multi-dimensional risk scores are then linearly weighted and fused to obtain the company's comprehensive initial risk score. .

[0023] Optionally, the risk assessment algorithm MA-ERC assesses the initial risk of enterprises 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:

[0024] Phase 1: Constructing a multi-dimensional enterprise risk knowledge graph In the context of [the search engine], the set of cross-capital relationships between any given enterprise node is used to find the relevant enterprise node. Any other enterprise node that is directly or indirectly connected to any other enterprise node in the relationship is defined as the relationship between the two enterprise nodes. Stored in a set of relations In the middle; through the set of relationships The risk weights of individual relationships are calculated, and then the risk weights of different types of relationships are merged to obtain the final risk propagation probability among all enterprises. ;

[0025] Phase Two: Assign initial risk values ​​to all enterprise nodes and then integrate the weights. Dynamically adjust the enterprise's overall initial risk value output by the multi-agent evaluation architecture The initial risk value for each company was obtained. After obtaining the initial risk of each enterprise and the risk propagation probability , the propagation process between risks is simulated by a personalized Pagerank algorithm, and multiple rounds of risk iteration are performed to obtain the final propagation risk value .

[0026] Optionally, the multi-modal enterprise data embedding output by the multi-agent evaluation architecture in S2 is taken as a feature, a hybrid expert model is used for feature fusion, and the multi-modal enterprise data is converted into a feature vector of a node, including:

[0027] The multi-modal enterprise data includes enterprise multi-dimensional risk scores and enterprise multi-dimensional risk analysis reports , wherein the enterprise multi-dimensional risk analysis report is obtained by fusing the risk analysis reports output by the multiple expert agents under each data type and containing the enterprise initial risk score;

[0028] The enterprise multi-dimensional risk scores are mapped to features through a feedforward neural network, and the semantic features of the text in the enterprise multi-dimensional risk analysis report are extracted using a Finbert pre-trained language model;

[0029] The hybrid expert model is used to fuse the features of the two different modalities, wherein the expert output is dynamically weighted and fused through a gating weight, the output vector of the hybrid expert model for different modalities is spliced, the multi-modal enterprise data is converted into a feature vector of a node, and the reconstruction of the enterprise node feature is completed by replacing the row vector corresponding to the enterprise node in the original node embedding matrix.

[0030] Optionally, for the enterprise multi-dimensional risk scores , a feedforward neural network composed of two linear layers is used to convert them into enterprise risk score features , which are represented as:

[0031] ;

[0032] wherein, and are weight matrices of hidden layers and output layers, and are bias vectors, is an activation function, and the enterprise risk score features are obtained through the FFN;

[0033] In Finbert, the CLS vector strategy is combined with the average pooling strategy, which is represented as:

[0034] ;

[0035] wherein, represents the final text feature, is the output vector of Finbert, is the CLS feature vector, is the full sequence average vector, is the actual length of the sequence, represents the token position index in the sequence;

[0036] A multi-dimensional risk analysis report of an enterprise Input the enterprise risk analysis report feature into Finbert , which is represented as:

[0037] ;

[0038] wherein, represents the Finbert model;

[0039] Using attention-based feature fusion can dynamically learn the importance weight of different features, realize adaptive feature fusion, which is represented as:

[0040] ;

[0041] ;

[0042] ;

[0043] wherein, is the original attention score, is the normalized attention weight, is the fused enterprise risk analysis report feature with data type , , represents the weight matrix, , represents the bias vector, represents the natural exponential function, represents the attention score of the risk analysis report feature of the th class;

[0044] For the fused enterprise risk analysis report feature and the enterprise risk score feature of two different modalities of features, a hybrid expert model is used for fusion, and the gating weight is calculated by the Sofemax function, which is represented as:

[0045] ;

[0046] wherein, is a set containing two different modalities of features, denotes the th feature of the th modality, is a weight matrix of the th modality feature with dimensionality of times the size of the expert subset, and the number of experts selected for each input is half of the total number of experts;

[0047] For each input , the fusion output is a weighted sum of the outputs of a randomly selected subset of experts weighted by gating weights, denoted as:

[0048] ;

[0049] wherein, is a randomly selected subset of expert indices with size , is the th expert network, each of which consists of two linear transformations and a Relu activation function, denotes the weight assigned to the th expert network by the gating mechanism, is the fusion output vector of the th modality of features, including the integrated risk score feature and the integrated risk analysis report feature , obtained by a weighted sum of the selected expert networks;

[0050] By concatenating the output vectors of different modalities of features, the multi-modal data of enterprises is converted into feature vectors of nodes ; finally, the feature vectors are used to replace the row vectors corresponding to the enterprise nodes in the original node embedding matrix , completing the reconstruction of the features of the enterprise nodes.

[0051] Optionally, a multi-layer RGAT model is used to model the heterogeneous relationships and dynamically learn the importance weights of the neighbor nodes under different relationship types, including:

[0052] The additive attention score between the head node and the tail node in different types of relationships is calculated to quantify the association strength of the two nodes under a specific relationship, denoted as:

[0053] ;

[0054] in, Represents the head node With tail node In relationship Attention score in , and For a trainable weight matrix, Represents a non-linear activation function. The eigenvectors of the head node. Represents the feature vector of adjacent nodes. and For relationship The query matrix and key matrix, Represents the query vector. Represents the key vector. For the edge eigenvectors;

[0055] After calculating attention scores for all different types of relations, cross-relation attention normalization is performed, and the importance of different relations is compared globally, as shown below:

[0056] ;

[0057] in, A set of relations in a knowledge graph. express A kind of relationship, express One of the adjacent nodes, For nodes The type is The set of adjacent nodes in a relation, The normalized attention coefficient. Indicates a relationship Next, node With adjacent nodes Attention score;

[0058] Information from neighboring nodes is aggregated using attention coefficient weighting, and multi-head attention aggregation for different relationship types is completed, as shown below:

[0059] ;

[0060] in, Indicates the attention head index. Indicates the number of attention heads. express A kind of relationship, Indicates adjacent nodes. Indicates a relationship With attention The specific weight matrix, Represents a node The type is The set of adjacent nodes in a relation, Indicates the head node With adjacent nodes Regarding the relationship With attention Attention coefficient The feature vectors of adjacent nodes;

[0061] Bullish attention The average aggregation is used to obtain the final output of node features. , is represented as:

[0062] ;

[0063] in, This represents the final node feature representation obtained after fusing information from multiple attention heads. Represents a non-linear activation function. Indicates the number of attention heads. Represents a node The aggregation result of the m-th attention head.

[0064] Optionally, the final node features are obtained through multi-layer RGAT stacking. 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 concatenated to obtain the enterprise's comprehensive features. The model evaluation results are then obtained through a linear layer, including:

[0065] After passing through the first layer of RGAT, ReLU activation is used and fed into the next layer of RGAT; the final node features are obtained by stacking multiple layers of RGAT. ;

[0066] final node features Enterprise nodes corresponding to Extract and measure the enterprise's risk of transmission. Map to vectors and concatenate them and To obtain the comprehensive characteristics of the enterprise , The final node characteristics of an enterprise node are represented using comprehensive enterprise characteristics. The evaluation results of the model obtained through the linear layer are expressed as follows:

[0067] ;

[0068] in, evaluation labels representing the model, linear layer weights representing the model, enterprise comprehensive features representing the model, bias vectors representing the model.

[0069] Compared with the prior art, the present application has the following beneficial effects:

[0070] In the present application, the 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 the present application can effectively improve the evaluation accuracy of the risk assessment model. At the same time, the present application provides a new idea for combining the multi-agent based on a large language model with knowledge graphs and graph neural networks and other technologies.

[0071] In the present application, by introducing expert risk scores, risk reports of multi-agents, and designing multi-modal node feature fusion, the enterprise nodes are given more realistic and differentiated features, effectively enhancing the discrimination ability of the model.

[0072] In the present application, an explainable graph EG for enterprise risk assessment is constructed, thereby solving the black box problem existing in traditional machine learning. Securities regulators and investors can intuitively view the correlation between enterprise risk and irregular behavior, greatly improving the transparency and credibility of the model, thereby strengthening the efficiency of compliance management and risk prevention.

[0073] The specific embodiments of the present application will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0074] The accompanying drawings, which are part of the present application, serve to provide a further understanding of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application, but do not constitute an improper limitation on the present application. Obviously, the drawings described below are only some embodiments, and other drawings can be obtained from these drawings by those of ordinary skill in the art without creating creative labor.

[0075] In the drawings:

[0076] Figure 1 a flowchart of the enterprise risk assessment method of the multi-agent-based risk assessment algorithm MA-ERC in the present specific embodiment;

[0077] Figure 2 a technical framework diagram of the enterprise risk assessment method of the multi-agent-based risk assessment algorithm MA-ERC in the present specific embodiment;

[0078] Figure 3A schematic diagram of an enterprise multi-dimensional risk knowledge graph framework of an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC in the embodiment;

[0079] Figure 4 An enterprise multi-modal multi-source risk data processing diagram of an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC in the embodiment;

[0080] Figure 5 A flowchart of the MA-ERC algorithm of an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC in the embodiment;

[0081] Figure 6 A MNF-GNN model framework diagram of an enterprise risk assessment method based on a multi-agent risk assessment algorithm MA-ERC in the embodiment. DETAILED DESCRIPTION

[0082] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The following embodiments are used to explain the present application, but are not used to limit the scope of the present application.

[0083] Please refer to 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:

[0084] S1, constructing an enterprise multi-dimensional risk knowledge graph:

[0085] At least including designing a schema layer, the schema layer at least defining core entities, risk propagation analysis entities and risk assessment analysis entities and defining corresponding relationship types between entities;

[0086] S2, designing a risk assessment algorithm MA-ERC to assess the initial risk of enterprises and quantifying the propagation risk value between enterprises:

[0087] The risk assessment algorithm MA-ERC assesses the initial risk of enterprises through a multi-agent evaluation architecture, simulates risk propagation by combining a personalized Pagerank algorithm to quantify the propagation risk value between enterprises, and sets different categories of expert agents for risk assessment through prompt words for different enterprise risk factors in the multi-agent evaluation architecture;

[0088] S3, constructing a MNF-GNN model for enterprise risk assessment:

[0089] The multi-modal enterprise data output by the multi-agent evaluation architecture in S2 is embedded as features, a hybrid expert model is used for feature fusion, and the multi-modal enterprise data is converted into a feature vector of a node; a multi-layer RGAT model is used to model the heterogeneous relationship and dynamically learn the importance weight of the neighbor node under different relationship types; the final node features are obtained through multi-layer RGAT stacking, the features corresponding to the enterprise nodes are extracted, and the enterprise propagation risk value is mapped into a vector; the features corresponding to the enterprise nodes and the enterprise propagation risk value are spliced to obtain the comprehensive features of the enterprise, and the evaluation result of the model is obtained through a linear layer;

[0090] S4, constructing an enterprise risk evaluation explainable graph EG:

[0091] By obtaining the enterprise multi-dimensional risk score, the enterprise multi-dimensional risk analysis report, and the risk propagation probability between enterprises, an enterprise risk evaluation explainable graph EG is constructed to provide explainability for the final evaluation result.

[0092] 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. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, a vehicle-mounted electronic device, a wearable device, etc., and the non-mobile electronic device can be a server and a personal computer, etc., which are not limited in the present application. The evaluation method in this embodiment is described below taking the server as an example.

[0093] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0094] It should be noted that the enterprise knowledge graph refers to a vertical field knowledge graph focusing on enterprise information and relationships, which has wide commercial applications. In the field of enterprise risk analysis, constructing an enterprise knowledge graph has become a common method. At the same time, the composition of the knowledge graph will be different depending on the purpose to be achieved. The main purposes of constructing the knowledge graph in this embodiment are as follows: on the one hand, it is hoped that by constructing the knowledge graph, the association between listed companies can be modeled, and the flow process of risk among listed companies can be simulated to quantify the concept of propagation risk. On the other hand, it is hoped that by constructing the knowledge graph, the complex association characteristics between nodes such as enterprises, shareholders, and cities can be mined and coded to provide more effective information for risk assessment.

[0095] Please see Figure 3 Based on the two purposes, and referring to the existing knowledge graph related research on enterprise risk analysis. The mode layer of the knowledge graph is designed in this embodiment, which abstracts and models entities and relationships around the two cores of risk transmission and risk assessment. In addition to enterprise entities, three types of entities, namely shareholders, senior managers and controlling persons, are defined to analyze the external structure and equity control path of the enterprise to simulate risk transmission; two types of entities, industry and city, are defined to mine the influence of macro-environmental policy factors on enterprise violations; and the entity of risk symptom event is defined to describe the risk symptoms disclosed by the enterprise in the current year. The corresponding relationship types between entities are defined. The multi-dimensional risk knowledge graph of the enterprise designed in this embodiment is shown in Figure 3 As shown in Table 1, the entities and relationship types of the embodiment are shown in Table 1:

[0096] Table 1: Entity-relationship table of multi-dimensional risk knowledge graph of enterprise

[0097]

[0098] Please continue to see Figure 3 According to the designed mode layer, the information of listed companies in 2022 is collected from the CSMAR database in this embodiment, including the information of shareholders, senior managers, cities and industries of listed companies, as well as the investment situation of listed companies, and the risk symptom reports of enterprises are crawled from the Sina Finance website. In this way, triples are formed and a multi-dimensional risk knowledge graph of enterprises is constructed. The constructed graph is saved to neo4j. The number of different types of nodes and relationships in the multi-dimensional risk knowledge graph of enterprises is shown in Table 2:

[0099] Table 2: Size table of multi-dimensional risk knowledge graph of enterprise

[0100]

[0101] It's worth noting that risk propagation has always been a hot topic in enterprise risk assessment. With the widespread application of knowledge graphs, risk propagation paths between enterprises through various means such as executives and equity have been revealed. Building on this, the personalized PageRank algorithm, a variant of the original PageRank algorithm, is widely used in enterprise risk propagation analysis due to its biased random walk mechanism, which can more accurately simulate the risk propagation process in enterprise networks. It assesses the influence of each node in risk propagation by simulating the process of risk propagation in the network, thereby quantifying the enterprise's propagation risk value. However, current methods focus on optimizing the propagation mechanism of the original PPR algorithm, but the initial risk value of nodes is generally assigned using the mean or based on a single risk factor, which obviously cannot fully reflect the true risk of the enterprise. In fact, for the personalized PageRank algorithm, the initial value is more important than that of the original PPR algorithm. In the PPR algorithm, the initial value is the source of risk, and in each subsequent random jump, there is 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 affects the distribution of the risk's influence in the network. If a company's initial risk value fails to represent its true risk, the more rational the propagation mechanism, the more the final result deviates from reality, amplifying erroneous risk signals. Therefore, the reasonable setting and optimization of the initial risk value becomes a key factor affecting the effectiveness of the PPR algorithm. Traditional methods, which assign values ​​based on the mean or a single dimension, ignore the multidimensionality and complexity of corporate risk.

[0102] Based on this, this embodiment proposes the MA-ERC algorithm to assess the initial risk of enterprises through a multi-agent evaluation architecture, and combines it with a personalized Pagerank algorithm to simulate risk propagation and quantify the risk value of propagation among enterprises. The multi-agent evaluation architecture introduces multimodal enterprise risk data to fuse and quantify the initial risk value of enterprises, making it closer to the actual risk status of enterprises. The input of the multi-agent evaluation architecture includes three types of enterprise risk factors: financial characteristics, CEO characteristics, and symptom risk characteristics. For different risk factors, three types of expert agents are set up using prompt words to assess the degree of enterprise risk.

[0103] Please see Figure 2 , Figure 4 and Figure 5The MA-ERC algorithm takes as input the company's financial characteristics, CEO characteristics, and symptom risk characteristics. Financial characteristics primarily include structured data composed of economic indicators reflecting the company's profitability (cost-to-profit ratio), growth capacity (total asset growth rate), and solvency (retained earnings-to-asset ratio, cash ratio, tangible asset-liability ratio). For CEO characteristics, five dimensions of data with the greatest impact on corporate misconduct are selected, including age, gender, education background, salary, and whether the CEO holds concurrent positions. For symptom risk characteristics, an LLM-based summarizing agent is used to process corporate risk event text crawled from Sina Finance website, including extracting high-frequency risks and long-term impacts. Specifically, because corporate risk events involve basic corporate information, this embodiment uses LLM to analyze historical corporate data; therefore, event anonymization is performed to ensure that LLM does not introduce premature information. Using a summarizing agent for summarization and inference can filter out noisy information, highlight core trend changes, and uncover potential information to extract high-value features.

[0104] Specifically, in the configuration of expert agents, each type of expert agent includes three analytical agents, which collaborate to complete the scoring of that type of risk factor. The three analytical agents consist of one confident agent and two moderate agents; the former tends to persuade other agents to believe in it, while the latter focus on listening to the answers from other agents.

[0105] The multi-agent initial risk assessment process mainly includes four stages. In the first stage, the three analytical agents receive enterprise risk data, namely the enterprise's financial characteristics, CEO characteristics, and symptom 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.

[0106] In the second phase, after obtaining the risk analysis report containing the company's initial risk score, multiple analytical agents interact through debate to obtain the analysis processes and results of other analytical agents. The purpose of the debate among multiple analytical agents is to achieve analysis sharing and consensus, allowing each analytical agent to share its analysis results based on the company's risk data. The analysis results of different analytical agents often differ, and the debate process revolves around these differences. Each agent ultimately decides whether to modify its own score.

[0107] In the third stage, each analytical agent reflects on its own analysis based on the analysis process in the first stage and the debate process in the second stage, judging whether its analysis report still has shortcomings that need to be revised. The reflection mechanism mimics the human way of thinking. By re-examining its own analysis process, iteratively raising questions and verifying them with the knowledge base, the agent can locate and correct errors, which can effectively solve the potential illusion problem of large language models.

[0108] 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. The purpose of the second round of debate is for each agent to share the process by which the others corrected errors and filled logical loopholes based on the debate and reflection, thereby forming a cross-validation network. This allows them to discover hidden errors that they themselves had not yet detected, and ultimately reach a group consensus.

[0109] Specifically, in the first stage, the enterprise's risk data is input, and a risk analysis report containing the enterprise's initial risk score is generated by the analysis agent. , is represented as:

[0110] ;

[0111] in, For the company being analyzed, As input features, A set of data types including financial risk, CEO risk, and symptom risk can be represented as follows: ,in, This represents a data type representing financial risk. This indicates the data type of CEO risk. This indicates a data type representing signs of risk. It is one of the data types. To analyze a set of intelligent agents, To enable expert agents to perform data analysis and specify the prompts that represent the risk level of the risk value, This represents the i-th analytical agent;

[0112] In the second phase, multiple analytical agents engage in the first round of debate, represented as follows:

[0113] ;

[0114] in, This is a risk analysis report, modified by the analytical agent after debate, which includes the company's initial risk score. The input to the analytical agent is the cue words that enable it to engage in debate. , and These represent the risk analysis reports from the other two analytical agents in the first phase, which include the initial risk score of the enterprise.

[0115] In the third stage, multiple analytical agents reflect on their respective actions, as shown below:

[0116] ;

[0117] in, This indicates a risk analysis report, including the initial risk score of the enterprise, that has been revised by the intelligent agent after reflection. Cue words that indicate that the analytical agent should engage in reflective behavior;

[0118] In the fourth stage, multiple analytical agents engage in a second round of debate, represented as follows:

[0119] ;

[0120] in, The final output of the expert agent is an initial risk analysis report for the enterprise, including a financial risk analysis report. CEO Risk Analysis Report Symptom and Risk Analysis Report ; and This refers to the initial risk analysis report of the enterprise by the other two analytical agents after the third phase. These are prompts used to induce the analytical agent to engage in debating behavior.

[0121] The final report output can extract multi-dimensional risk scores for enterprises from multiple analytical agents. Including financial risk scoring CEO risk score Symptom Risk Score Each risk factor score includes scores from three expert agents. These three risk scores are then linearly weighted and fused to obtain the company's overall initial risk score. , is represented as:

[0122] ;

[0123] in, , , For linear fusion weights, For the number of expert agents, here The value is 3. To obtain the final overall initial risk value for the enterprise, , , These represent the scores of individual financial, CEO, and symptom risk analysis agents, respectively.

[0124] Calculate the probability of risk propagation between enterprises in the constructed multidimensional enterprise risk knowledge graph. middle. exist Figure 5The graph is represented as a heterogeneous graph, using seven different colors—cyan (enterprises), green (shareholders), pink (executives), blue (controllers), yellow (industries), purple (cities), and red (risk events)—to represent seven different types of nodes. The set of cross-capital relationships between nodes is defined as... ,in, This represents different types of relationships between enterprise nodes and other nodes. (For enterprise multidimensional risk knowledge graphs) Any enterprise node in Search for it in the graph. Any other enterprise node directly or indirectly connected to it. Define the relationship between nodes as Stored in a collection middle.

[0125] For the three types of relationships—investment, shareholding, and controlling stake—if the relationship is directly related, the shareholding ratio / investment amount is used as the relationship weight. If the relationship is indirect, the weight is defined by merging the investment amounts of the two companies or by finding the largest shareholding percentage in the relationship path between the two companies. Set threshold and According to weight Perform filtering to identify relationships with all weights exceeding a threshold and store them in a relationship set. middle.

[0126] After obtaining the set of relations Next, the risk weight of a single relationship is calculated. The purpose of this step is to quantify the impact of different relationships on the strength of risk transmission between firms, and then to integrate cross-capital relationships with different risk weights to obtain the probability of risk propagation between firms. The formula for calculating the risk weight of a single relationship is shown below:

[0127] ;

[0128] in, This refers to Company 1 in the annual Discovering risky behavior, This refers to Company 2 in the year Discovering risky behavior, This is an indicator function; its value is 1 if the condition is true, and 0 otherwise. The relation type is The relationship is the total number of records after filtering. This is one of the records. If the company is in a relationship... If a two-way pattern of "risk in the same year → risk of related enterprises in the following year" exists, it indicates that there is a possibility of risk transmission in this type of relationship. Calculate the probability of this pattern occurring in this type of relationship and obtain the relationship type. Risk weight .

[0129] Subsequently, since there are often various different capital cross-linking relationships between enterprises, the risk weights of different types of relationships are combined to obtain the final risk propagation probability among all enterprises. The formula is shown below:

[0130] ;

[0131] in, For relation type Risk weight of the relationship for The total number of relationships in the equation.

[0132] 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 for each enterprise is as follows:

[0133] ;

[0134] in, For enterprises The initial risk value of the enterprise obtained by multiple analytical agents is divided by 100 to scale it down to the range of 0-1. This indicates whether the company engaged in any risky activities in the previous year; a value of 1 indicates such activities, and a value of 0 indicates otherwise. The fusion weights for the two initial risks are determined by... The fusion ratio of the two initial risks is dynamically adjusted to ultimately obtain the company's initial risk value. .

[0135] After obtaining the initial risk value and the probability of risk transmission Then, the propagation process between risks is simulated, and multiple rounds of risk iteration are performed. The calculation formula is shown below:

[0136] ;

[0137] in, As a risk dynamic adjustment factor, unlike the original damping factor, it dynamically adjusts the decay ratio of risk weight with each risk iteration round to prevent the risk of high-risk nodes from being excessively diluted during the iteration process. For iteration rounds, Indicates the maximum number of iterations. express Index of adjacent nodes in the middle, For iteration The final transmission risk value obtained is Iteration +1 times to obtain the final transmission risk value.

[0138] The obtained propagation risk value was used as an additional one-dimensional feature and input into the machine learning model. This model was then compared with a machine learning model whose input only included the company's economic characteristics. The experimental results showing the increase in machine learning performance after adding the propagation risk value are shown in Table 3.

[0139] Table 3: Increase in Machine Learning Performance After Adding Risk Values

[0140]

[0141] Please see Figure 1 , Figure 4 and Figure 6 To construct an MNF-GNN model for enterprise risk assessment, multimodal enterprise data output by multiple agents in S2 is embedded as features. A hybrid expert model (MOE) is used for feature fusion, transforming the multimodal enterprise data into a comprehensive feature vector of multimodal enterprise nodes. Then, in... Figure 6 Heterogeneous diagrams In, with Figure Five The same colors correspond to different types of nodes. Also, using... This represents the different types of relationships between enterprise nodes and other nodes, as shown in Table 1. For multiple types of relationships, a multi-layer RGAT model is used to model them, dynamically learning the importance weights of neighboring nodes under different relationship types. The final node features are obtained by stacking multiple layers of RGAT, the features corresponding to the enterprise nodes are extracted, and the enterprise's propagation risk value is mapped into a vector. These two features are concatenated to obtain the comprehensive features of the enterprise, and the model evaluation results are obtained through a linear layer.

[0142] Specifically, this involves multi-dimensional risk scoring for enterprises. Multi-dimensional risk analysis report for enterprises This is achieved by embedding the features into the multimodal node feature encoder involved. This is used for enterprise multidimensional risk scoring. The risk score is transformed into enterprise risk scoring features through a feedforward neural network consisting of two linear layers, as shown in the following formula:

[0143] ;

[0144] in, and This is the weight matrix for the hidden layer and the output layer. and For bias vectors, As the activation function, the enterprise risk score features are obtained through this FFN. .

[0145] Due to the enterprise's multi-dimensional risk analysis report For text data containing a large amount of financial terminology, the Finbert model is used to extract semantic features from the text. Furthermore, to fully exploit the global and local semantic information within the analysis report, a combination of the CLS vector strategy and average pooling strategy is used in Finbert, as shown in the following formula:

[0146] ;

[0147] in, This is Finbert's output vector. For CLS feature vectors, This is the average vector of the entire sequence. This represents the actual length of the sequence.

[0148] Enterprise multi-dimensional risk analysis report enter Features of Enterprise Risk Analysis Report As shown in the following formula:

[0149] ;

[0150] in, Represents the Finbert model. This refers to a multi-dimensional risk analysis report for enterprises.

[0151] because Includes and These two different types of characteristics need to be combined to represent the characteristics of a corporate risk analysis report. Attention-based feature fusion can dynamically learn the importance weights of different features, achieving adaptive feature fusion, as shown in the following equation:

[0152] ;

[0153] ;

[0154] ;

[0155] in, For the original attention score, The attention weights are normalized. Features of the integrated enterprise risk analysis report , Represents the weight matrix. , This represents the bias vector. This represents the natural exponential function. Indicates the first Attention score for risk analysis report features.

[0156] Characteristics of the integrated enterprise risk analysis report and the characteristics of enterprise risk scores obtained by numerical scoring mapping. Features from two different modalities are fused using a hybrid expert model (MOE). MOE has multiple expert networks and a modality-independent gating mechanism. For features from different modalities, the expert outputs need to be dynamically weighted and fused using gating weights. The result is calculated using the Sofemax function, as shown in the following formula:

[0157] ;

[0158] in, It is a set containing three different modal features. Indicates the first Characteristics of class modality For the first The weight matrix of the modality features, with dimension equal to the input size. Size of expert subset The product of these factors, where the number of experts selected for each input is half the total number of experts.

[0159] For each input The fusion output is weighted and summed using gated weights on the outputs of a randomly selected subset of experts, as shown in the following formula:

[0160] ;

[0161] in, For size A randomly selected subset of expert indexes, For the first A network of experts, each expert consisting of two layers of linear transformation and ReLU activation function. For the first The fusion output vector of modal features, including comprehensive risk score features. Features of comprehensive risk analysis reports It is obtained by weighted summation of the selected experts.

[0162] By concatenating the output vectors of different modal features, enterprise multimodal data is transformed into a comprehensive feature vector of multimodal enterprise nodes. ; Use last Replace the original node embedding matrix The row vectors corresponding to the enterprise nodes are used to reconstruct the features of the enterprise nodes.

[0163] Specifically, based on the constructed enterprise multi-dimensional risk knowledge graph With the original node embedding matrix Using a multi-layered RGAT model to analyze the enterprise's multi-dimensional risk knowledge graph 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 node in different relationship types is calculated. With tail node The additive attention score between two nodes quantifies the strength of their association under a specific relationship, as shown in the following formula:

[0164] ;

[0165] in, , and For a trainable weight matrix, Represents a non-linear activation function. The eigenvectors of the head node. Represents the feature vector of adjacent nodes. and For relationship The query matrix and key matrix, Represents the query vector. Represents the key vector. For the edge eigenvectors.

[0166] After calculating the attention scores for all different types of relations, cross-relation attention normalization is performed, and the importance of different relations is compared globally, as shown in the following formula:

[0167] ;

[0168] in, A set of relations in a knowledge graph. express A kind of relationship, express One of the adjacent nodes, For nodes The type is The set of adjacent nodes in a relation, This is the normalized attention coefficient.

[0169] Then, the information of neighboring nodes is aggregated by weighting the attention coefficients, and multi-head attention aggregation of different relationship types is completed, as shown in the following formula:

[0170] ;

[0171] in, Indicates the attention head index. Indicates the number of attention heads. express A kind of relationship, Indicates adjacent nodes. Indicates a relationship With attention The specific weight matrix, Indicates the head node With adjacent nodes Regarding the relationship With attention Attention coefficient is the feature vector of the adjacent node.

[0172] Bullish attention The average aggregation is used to obtain the final output of node features. As shown in the following formula:

[0173] ;

[0174] in, Represents a non-linear activation function. Indicates the number of attention heads. Represents a node The aggregation result of the m-th attention head.

[0175] To further aggregate the global features of multi-hop neighbor nodes, after passing through the first layer of RGAT, ReLU activation is used and fed into the next layer of RGAT; the final node features are obtained by stacking multiple layers of RGAT. .Will Extract the features corresponding to the enterprise nodes and assign the enterprise's propagation risk value. Mapping to vectors, these two features are concatenated to obtain the comprehensive characteristics of the enterprise. , Representing the final node characteristics of an enterprise node, using The evaluation results of the model are obtained through the linear layer, as shown in the following equation.

[0176] ;

[0177] in, Indicates the evaluation label of the model. Represents the linear layer weights. Indicates the comprehensive characteristics of the enterprise. This represents the bias vector.

[0178] After obtaining the evaluation results, the binary cross-entropy loss function is used to calculate the loss at a length of... Evaluation results With the company's real label The loss between them is used to continuously iterate and optimize the model, as shown in the following equation:

[0179] ;

[0180] in, This represents the calculated loss value. Indicates the number of samples. Indicates the first The true label of each sample Indicates the first Evaluation labels for each sample Represents a logarithmic function.

[0181] As a specific implementation method, the proposed method in this embodiment is compared with existing mainstream enterprise risk assessment methods to comprehensively evaluate its performance under different feature perspectives and modeling strategies. This mainly includes representative models from the following three categories: the first is a machine learning model that evaluates enterprise structured data; the second is a graph neural network model that evaluates enterprise multidimensional risk knowledge graphs; and the third is an enterprise risk analysis model that evaluates through multi-source data fusion. The comparative experimental results are shown in Table 4 below:

[0182] Table 4: Comparative Experiment Results

[0183]

[0184] An interpretable graph (EG) for enterprise risk assessment is constructed: Each enterprise node is connected to its own financial, CEO, and symptom risk scoring nodes through scoring relationships. Each scoring node is connected to its corresponding cause node through causal relationships, reflecting the enterprise's own risks. The content of the cause nodes constitutes the conclusions in the final risk analysis report of the multi-agent system, i.e., the reasons for the agent's risk scoring, as shown in Table 5.

[0185] Table 5: Examples of Reasons for Agent Risk Scoring

[0186]

[0187] Furthermore, each enterprise is connected to nodes with risk propagation channels through risk propagation relationships, with the risk propagation probability used as the edge weight. Each node has an attribute evaluation value, which is the evaluation result of the risk assessment model. By analyzing the node attributes, edge weights, and causal relationships along these paths, an interpretable risk propagation chain can be constructed, revealing the key influencing factors behind the model's evaluation results. For example, by querying the upstream propagation path of a high-risk enterprise, external enterprise nodes that significantly affect its score can be identified, determining whether the enterprise is classified as high-risk by the model due to its association with multiple high-risk enterprises. Simultaneously, the focus can be placed on the enterprise's own scoring nodes to identify specific risk points such as financial anomalies, CEO changes, and negative public opinion, thereby achieving a retrospective explanation from "assessment results" to "risk sources."

[0188] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A business risk assessment method based on the multi-agent risk assessment algorithm MA-ERC, characterized in that, Includes the following steps: S1. Construct a multi-dimensional risk knowledge graph for enterprises: At least one schema layer is designed, which defines at least a core entity, a risk propagation analysis entity, and a risk assessment analysis entity, and defines the corresponding relationship types between each entity; S2. Design the MA-ERC risk assessment algorithm to assess the initial risk of enterprises and quantify the risk of inter-enterprise propagation: The risk assessment algorithm MA-ERC assesses the initial risk of enterprises through a multi-agent assessment architecture, and combines a personalized Pagerank algorithm to simulate risk propagation in order to quantify the risk value of propagation between enterprises. The multi-agent assessment architecture sets up different categories of expert agents for risk assessment based on prompt words for different enterprise risk factors. S3. Construct an MNF-GNN model for enterprise risk assessment: The multimodal enterprise data output from the multi-agent evaluation architecture in S2 is embedded as features, and a hybrid expert model is used for feature fusion to transform the multimodal enterprise data into feature vectors of nodes. A multi-layer RGAT model is used to model the 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, the features corresponding to the enterprise nodes are extracted, and the enterprise's propagation risk value is mapped into a vector. The features corresponding to the enterprise nodes and the enterprise's propagation risk value are concatenated to obtain the comprehensive enterprise features. The model evaluation results are obtained through a linear layer. S4. Construct an interpretable diagram for enterprise risk assessment (EG): By obtaining multi-dimensional risk scores, multi-dimensional risk analysis reports, and risk propagation probabilities between enterprises, an interpretable graph (EG) for enterprise risk assessment is constructed to provide interpretability for the final assessment results.

2. The method according to claim 1, characterized in that, The schema layer described in S1 at least defines a core entity, a risk propagation analysis entity, and a risk assessment analysis entity, and defines the corresponding relationship types between each entity, including: The core entities include corporate entities, the risk propagation analysis entities include three types of entities: shareholders, senior executives, and controllers, which are used to analyze the external structure and equity control path of enterprises 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 macroeconomic policy factors on corporate risks, and also include symptom risk event entities, which are used to describe the risk symptoms that are disclosed to the enterprise in the current year.

3. The method according to claim 1, characterized in that, The S2 multi-agent evaluation architecture uses prompts to assign different categories of expert agents to conduct risk assessments for different enterprise risk factors, including: Each type of expert agent includes multiple analytical agents that collaborate to complete risk factor scoring for their respective categories. Each analytical agent consists of one confident agent and two moderate agents. The confident agent tends to persuade the other agents, while the moderate agents focus on listening to the answers of the other agents.

4. The method according to claim 3, characterized in that, The risk assessment algorithm MA-ERC described in S2 assesses a firm's initial risk through a multi-agent assessment architecture, including: In the first stage, multiple analytical agents receive enterprise risk data, which includes enterprise financial characteristics, CEO characteristics, and symptom risk characteristics, and generate a risk analysis report containing an initial risk score for the enterprise. In the second stage, after obtaining the risk analysis report containing the initial risk score of the enterprise, multiple analytical agents interact through debate to obtain the analysis process and results of other analytical agents. Each analytical agent shares its own analysis results based on the enterprise risk data. The debate process revolves around the differences, and each analytical agent decides in the end 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 a risk analysis report containing the company's initial risk score. The risk analysis report, which includes the company's initial risk score, extracts a multi-dimensional risk score for the company. Each dimension's risk score comprises scores from multiple expert agents. These multi-dimensional risk scores are then linearly weighted and fused to obtain the company's comprehensive initial risk score. .

5. The method according to claim 4, characterized in that, The risk assessment algorithm MA-ERC assesses the initial risk of enterprises 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: Constructing a multi-dimensional enterprise risk knowledge graph In the context of [the search engine], the set of cross-capital relationships between any given enterprise node is used to find the relevant enterprise node. Any other enterprise node that is directly or indirectly connected to any other enterprise node in the relationship is defined as the relationship between the two enterprise nodes. Stored in a set of relations In the middle; through the set of relationships The risk weights of individual relationships are calculated, and then the risk weights of different types of relationships are merged to obtain the final risk propagation probability among all enterprises. ; Phase Two: Assign initial risk values ​​to all enterprise nodes and then integrate the weights. Dynamically adjust the enterprise's overall initial risk value output by the multi-agent evaluation architecture The initial risk value for each company was obtained. After obtaining the initial risk of each enterprise and the probability of risk transmission Then, the propagation process between risks is simulated using a personalized PageRank algorithm, and multiple rounds of risk iteration are performed to finally obtain the propagation risk value. .

6. The method according to claim 4, characterized in that, The multimodal enterprise data output from 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 for nodes, including: The multimodal enterprise data includes multi-dimensional risk scores for enterprises. Multi-dimensional risk analysis report for enterprises Among them, the enterprise's multi-dimensional risk analysis report It is obtained by merging risk analysis reports containing the company's initial risk score output by multiple expert agents under each data type; The enterprise's multi-dimensional risk score is obtained through a feedforward neural network. Using mapping as features, a Finbert pre-trained language model is used to extract multi-dimensional risk analysis reports for enterprises. Semantic features in text; A hybrid expert model is used to fuse features from two different modalities. The expert outputs are dynamically weighted and fused using gating weights. The output vectors of the hybrid expert model with different modal features are concatenated to transform multimodal enterprise data into feature vectors of nodes. The enterprise node features are reconstructed by replacing the row vectors corresponding to the enterprise nodes in the original node embedding matrix.

7. The method according to claim 6, characterized in that, For enterprise multi-dimensional risk scoring The risk score features are transformed into enterprise risk scoring features through a feedforward neural network consisting of two linear layers. , is represented as: ； in, and This is the weight matrix for the hidden layer and the output layer. and For bias vectors, Using FFN as the activation function, the enterprise risk score features are obtained. ; The combination of CLS vector strategy and average pooling strategy in Finbert is represented as follows: ； in, This represents the final text features obtained. This is Finbert's output vector. For CLS feature vectors, This is the average vector of the entire sequence. This is the actual length of the sequence. Indicates the position index of the token in the sequence; Enterprise multi-dimensional risk analysis report Inputting Finbert yields features of the enterprise risk analysis report. , is represented as: ； in, Represents the Finbert model; Attention-based feature fusion can dynamically learn the importance weights of different features, achieving adaptive feature fusion, as shown below: ； ； ； in, For the original attention score, The attention weights are normalized. For data type Features of the integrated enterprise risk analysis report , Represents the weight matrix. , This represents the bias vector. This represents the natural exponential function. Indicates the first Attention score of risk analysis report features; Characteristics of the integrated enterprise risk analysis report and Enterprise Risk Scoring Characteristics Features from two different modalities are fused using a hybrid expert model with gated weights. Calculated using the Sofemax function, it is expressed as: ； in, It is a set containing two different modal features. Indicates the first Characteristics of class modality For the first The weight matrix of the modality features, with dimension equal to the input size. Size of expert subset The product of these factors, where the number of experts selected for each input is half the total number of experts; For each input The fusion output is a weighted sum of the outputs of a randomly selected subset of experts, obtained by gating weights, and is expressed as: ； in, For size A randomly selected subset of expert indexes, For the first There are several expert networks, each consisting of two layers of linear transformation and a ReLU activation function. This indicates that the gating mechanism is assigned to the first... The weights of an expert network, For the first The fusion output vector of modal features, including comprehensive risk score features. Features of comprehensive risk analysis reports This is obtained by weighted summation of the selected expert network; By concatenating the output vectors of different modal features, enterprise multimodal data is transformed into feature vectors for nodes. Finally, the feature vectors are used. Replace the original node embedding matrix The row vectors corresponding to the enterprise nodes are used to reconstruct the features of the enterprise nodes.

8. The method according to claim 5, characterized in that, A multi-layer RGAT model is used to model heterogeneous relationships, dynamically learning the importance weights of neighbor nodes under different relationship types, including: Calculate the head node in different types of relations With tail node The additive attention score between two nodes quantifies the strength of their association under a specific relationship, and is expressed as: ； in, Represents the head node With tail node In relation Attention score in , and For a trainable weight matrix, Represents a non-linear activation function. The eigenvectors of the head node. Represents the feature vector of adjacent nodes. and For relationship The query matrix and key matrix, Represents the query vector. Represents the key vector. For the edge eigenvectors; After calculating attention scores for all different types of relations, cross-relation attention normalization is performed, and the importance of different relations is compared globally, as shown below: ； in, A set of relations in a knowledge graph. express A kind of relationship, express One of the adjacent nodes, For nodes The type is The set of adjacent nodes in a relation, The normalized attention coefficient. Indicates a relationship Next, node With adjacent nodes Attention score; Information from neighboring nodes is aggregated using attention coefficient weighting, and multi-head attention aggregation for different relationship types is completed, as shown below: ； in, Indicates the attention head index. Indicates the number of attention heads. express A kind of relationship, Indicates adjacent nodes. Indicates a relationship With attention The specific weight matrix, Represents a node The type is The set of adjacent nodes in a relation, Represents the head node With adjacent nodes Regarding relationships With attention Attention coefficient The feature vectors of adjacent nodes; Bullish attention The average aggregation is used to obtain the final output of node features. , is represented as: ； in, This represents the final node feature representation obtained after fusing information from multiple attention heads. Represents a non-linear activation function. Indicates the number of attention heads. Represents a node The aggregation result of the m-th attention head.

9. The method according to claim 8, characterized in that, The final node features are obtained by stacking multiple layers of RGAT. The features corresponding to the enterprise nodes are extracted, and the enterprise's propagation risk value is mapped into a vector. The features corresponding to the enterprise nodes and the enterprise's propagation risk value are concatenated to obtain the comprehensive enterprise features. The model evaluation results are obtained through a linear layer, including: After passing through the first layer of RGAT, ReLU activation is used and fed into the next layer of RGAT; the final node features are obtained by stacking multiple layers of RGAT. ; final node features Enterprise nodes corresponding to Extract and measure the enterprise's risk of transmission. Map to vectors and concatenate them and To obtain the comprehensive characteristics of the enterprise , The final node characteristics of an enterprise node are represented using comprehensive enterprise characteristics. The evaluation results of the model obtained through the linear layer are expressed as follows: ； in, Indicates the evaluation label of the model. Represents the linear layer weights. Indicates the comprehensive characteristics of the enterprise. This represents the bias vector.

Images