An explainable enterprise transaction risk intelligent assessment method and system

Abstract

This invention discloses an interpretable intelligent assessment method and system for enterprise transaction risks. The method includes the following steps: Step 1: Constructing and evolving a dynamic heterogeneous knowledge graph to obtain a dynamic heterogeneous knowledge graph; Step 2: Using a graph attention network to extract deep features from each enterprise node in the graph, and by aggregating the features of neighboring nodes, mapping the complex topological relationships into a high-dimensional feature vector to obtain the final layer feature vector of the enterprise node graph; Step 3: Performing interpretable risk assessment and path backtracking to obtain a quantified risk score and Top-K interpretable evidence paths; Step 4: For abnormal transactions that deliberately evade regulation, calculating the isomorphism difference of the subgraph, and analyzing the entropy value of the node degree distribution to identify anomalies in the topological structure, obtaining anomaly entropy values.

Description

Technical Field

[0001] This invention relates to the field of internet finance security technology, and in particular to an interpretable intelligent assessment method and system for enterprise transaction risks. Background Technology

[0002] With the increasing complexity of globalized trade and capital markets, the scale and frequency of corporate transactions are constantly rising. Corporate background checks and risk assessments have become core components of risk control, anti-money laundering, and anti-fraud efforts. However, in the current business environment, inter-corporate relationships are characterized by increasingly concealed structures, dynamic methods, and cross-domain scope. Through the establishment of multiple shell companies, cross-shareholding agreements, and nominee agreements, malicious entities can effectively conceal actual control and potential debt burdens. Traditional assessment methods heavily rely on manual verification and simple rule engines. Faced with terabytes of heterogeneous business, judicial, and financial data, they struggle to identify deeply nested risk transmission chains, resulting in inefficient assessments and a high risk of false positives or omissions, seriously threatening financial security and market integrity.

[0003] To address the aforementioned challenges, existing research primarily employs query analysis based on relational databases and risk scoring techniques based on simple machine learning models. While these methods can handle structured data, they have significant limitations in complex relationship modeling and dynamic prediction. Furthermore, existing systems are mostly static models, unable to perceive the real-time impact of dynamic events such as changes in corporate equity or legal representatives on risk topology, exhibiting severe lag. Moreover, existing risk assessment models are mostly black-box structures, only outputting risk probability values ​​and lacking logical explanations and evidence backtracking for review conclusions, failing to meet regulatory agencies' interpretability requirements for compliance review results. Finally, existing methods lack depth in mining unstructured public opinion information, making it difficult to capture non-financial related risks hidden within news events.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an interpretable intelligent assessment method and system for enterprise transaction risks. The invention aims to overcome the aforementioned technical deficiencies by constructing a dynamically evolving enterprise panoramic knowledge graph and combining deep graph learning with interpretable reasoning mechanisms.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An interpretable intelligent assessment method for enterprise transaction risk includes the following steps: Step 1: Obtain structured and unstructured data from heterogeneous data sources such as the industrial and commercial system, court announcements, and news media; construct and evolve a dynamic heterogeneous knowledge graph to obtain a real-time updated dynamic heterogeneous knowledge graph. ; Step 2: Obtain a dynamically updated heterogeneous knowledge graph in real time. Using a graph attention network to focus on each enterprise node in the graph Deep feature extraction is performed, and by aggregating the features of neighboring nodes, the complex topological relationships are mapped into high-dimensional feature vectors to obtain enterprise nodes. The final layer feature vector ; Step 3: Obtain the feature vectors of both parties in the transaction. and The feature vector of the current transaction and the attention coefficient between each side Explainable risk assessment and path backtracking are conducted to obtain a quantified risk score. and Top-K paths of interpretable evidence; Step 4: Obtain the local subgraph To identify anomalies in the topology by calculating the isomorphism differences of the subgraph and analyzing the entropy values ​​of the node degree distribution for abnormal transactions that deliberately evade regulation, anomaly entropy values ​​are obtained. It outputs a warning signal for manual review and intervention.

[0007] Furthermore, step 1 includes the following steps: Step 1.1: Use the pre-trained NLP model BERT to extract enterprise entities E, attributes A, and relationships R from heterogeneous data sources to construct a dynamic graph. Where T is the time axis; extract the triple relationship including the defendant, plaintiff and cause of action from the court announcement text; obtain the shareholding changes from the real-time data stream of the industry and commerce bureau; Step 1.3: Use a message queue to receive business registration change events. Execute the evolution function Incremental updates to the graph structure: ; in, For graph evolution operators, responsible for adding, deleting, and modifying the attributes of nodes and edges. The graph structure at the current moment. This is a newly triggered change event.

[0008] Furthermore, step 2 includes the following steps: Step 2.1: Use the company's operating indicators as the initial vector. Through multi-layer W matrix learning, different types of nodes are mapped to a unified feature space. These different types of nodes include natural persons, shell companies, and group enterprises. Step 2.2: Use a graph attention network to obtain the deep topological features of each enterprise node; for each node... In the eigenvectors of the layer Aggregation is achieved through the following formula: ; in, () is the sigmoid function. Let i be the set of neighbors of node i. Here, W represents the attention coefficients, and W is the learnable shared weight matrix. Let j be a feature of the l-th layer neighbor node j, and let j be the risk contribution weight of the neighbor node i.

[0009] Furthermore, step 3 includes the following steps: Step 3.1: Combine the topological vectors of both trading parties with the feature vector of the current transaction. Feature concatenation, outputting both parties to the transaction through a fully connected neural network. and Risk scores between Specifically, risk score The calculation method is as follows: ; Among them, the transaction feature vector The transaction amount, transaction type, region, and timestamp of the current auditable transaction are extracted in real time, and then concatenated after normalization and one-hot encoding; if Enter backtracking mode and extract the evidence chain path P; Step 3.1: Based on The formula identifies the path with the largest product in the graph as the salient path; specifically, the salient score is calculated using the following formula: ; in, The significance score of path P is used; a higher score indicates a more significant impact of the associated path on risk. The top-K paths serve as the evidence chain for background checks, and the specific logic of risk transmission is visualized.

[0010] Furthermore, step 4 includes the following steps: Step 4.1: Taking the transacting party as the center, extract all related entities within its 3rd-order neighborhood to form a subgraph; Step 4.2: Calculate the anomaly entropy value of the local subgraph The formula is: ; in, Degree of nodes in the subgraph The probability distribution, The larger the value, the more complex and disordered the relational structure of the region, which may pose a risk of adversarial forgery. Step 4.3: Perform degree distribution statistics on the target company's network of connections. When the entropy value is lower than the preset normal enterprise benchmark, a warning will pop up on the review interface.

[0011] To achieve the above objectives, the present invention also employs the following technical solution: An interpretable intelligent assessment system for enterprise transaction risk includes: Data source layer: responsible for the real-time access and preprocessing of all heterogeneous data, providing multi-dimensional information support for background checks; Graph Engine Layer: Responsible for the construction and dynamic evolution of the graph, including entity and relation extraction, dynamic evolution and incremental updates; Core algorithm layer: Utilizes deep graph learning techniques to extract deep topological features from enterprise interconnected networks, including graph attention networks, risk representation learning, and risk assessment models; Application Output Layer: Provides visual and interpretable review conclusions for the final review process.

[0012] Furthermore, the data source layer includes: business registration data: providing information on equity investments, legal entity relationships, and senior management appointments between corporate entities; judicial documents: using pre-trained NLP models to extract entity and risk event triples from legal proceedings and enforcement announcements; news media: capturing unstructured negative public opinion through sentiment analysis and entity linking to identify non-financial related risks; and transaction data: accessing the dynamic characteristics of transactions under review.

[0013] Furthermore, in the graph engine layer, entity and relation extraction: mapping multi-source heterogeneous data into nodes and edges in a dynamic heterogeneous knowledge graph; dynamic evolution and incremental update: listening for change events through a message queue and executing evolution functions.

[0014] Furthermore, in the core algorithm layer, the graph attention network calculates the attention coefficient using the graph attention mechanism to identify the risk contribution of neighboring nodes to the current subject; risk representation learning aggregates nodes into feature vectors in a high-dimensional space through weight matrices and multi-layer nonlinear transformations, achieving penetrating feature extraction of multi-layer nested structures; and the risk assessment model uses a multilayer perceptron to fuse the topological vectors of both parties in a transaction to quantify potential transaction risks.

[0015] Furthermore, in the output layer, the risk score is output: the quantitative score is output. Path saliency analysis: Extracting evidence paths based on the product of attention coefficients to achieve risk traceability; Anomaly entropy detection: Targeting structured abnormal enterprise transactions, quantifying the subgraph topological entropy value. Achieve early warning of adversarial anomalies.

[0016] Compared with the prior art, the beneficial effects of this invention are as follows: This invention not only enables millisecond-level updates to the graph structure through an event-driven mechanism to capture instantaneous risks, but also utilizes an attention-based graph neural network to automatically extract critical paths leading to risks as review criteria while performing high-dimensional relationship representation. Furthermore, through deep fusion of multimodal information and adversarial anomaly detection, this invention significantly improves the accuracy and reliability of risk identification in complex transaction scenarios, providing an intelligent and transparent technical means for enterprise transaction security. Attached Figure Description

[0017] Figure 1 A flowchart of an interpretable intelligent assessment method for enterprise transaction risks; Figure 2 This is an architecture diagram of an interpretable intelligent assessment system for enterprise transaction risks. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1: An interpretable intelligent assessment method for enterprise transaction risk, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain structured and unstructured data from heterogeneous data sources such as the industrial and commercial system, court announcements, and news media; construct and evolve a dynamic heterogeneous knowledge graph to obtain a real-time updated dynamic heterogeneous knowledge graph. .

[0020] In this embodiment, step 1 is the dynamic heterogeneous knowledge graph construction and evolution step, which is responsible for the real-time ingestion of multi-source data and the dynamic updating of the graph structure.

[0021] In this embodiment, step 1 includes the following steps: Step 1.1: Use the pre-trained NLP model BERT to extract enterprise entities E, attributes A, and relationships R from heterogeneous data sources to construct a dynamic graph. Where T is the time axis; extract the ternary relationship including the defendant, plaintiff and cause of action from the court announcement text; obtain the shareholding changes from the real-time data stream of industry and commerce.

[0022] Step 1.3: Use a message queue to receive business registration change events. Execute the evolution function Incremental updates to the graph structure: ; in, For graph evolution operators, responsible for adding, deleting, and modifying the attributes of nodes and edges. The graph structure at the current moment. This is a newly triggered change event.

[0023] In this embodiment, when the event occurs that "Company A pledges 40% of its equity to Company B", the function... Instead of reconstructing the entire graph, only the nodes are reconstructed. and Modify its adjacency list.

[0024] In this embodiment, the validity period is marked for each relationship using the time axis T, ensuring that the graph structure is invoked when reviewing the current transaction context. It provides the latest legal and business status, effectively avoiding the risk of misjudgment due to information lag.

[0025] Step 2: Obtain a dynamically updated heterogeneous knowledge graph in real time. Using a graph attention network (GAT) to focus on each enterprise node in the graph Deep feature extraction is performed, and by aggregating the features of neighboring nodes, the complex topological relationships are mapped into high-dimensional feature vectors to obtain enterprise nodes. The final layer feature vector .

[0026] In this embodiment, step 2 is a relation representation learning step based on graph neural networks, which is responsible for transforming complex enterprise relationships into topological feature vectors in a high-dimensional space.

[0027] In this embodiment, step 2 includes the following steps: Step 2.1: Use the company's operating indicators (registered capital, tax level) as the initial vector. Through multi-layer W matrix learning, different types of nodes are mapped to a unified feature space. These different types of nodes include natural persons, shell companies, and group enterprises.

[0028] Step 2.2: Use a graph attention network to obtain the deep topological features of each enterprise node; for each node... In the eigenvectors of the layer Aggregation is achieved through the following formula: ; in, () is the sigmoid function. Let i be the set of neighbors of node i. Here, W represents the attention coefficients, and W is the learnable shared weight matrix. Let j be a feature of the l-th layer neighbor node j, and let j be the risk contribution weight of the neighbor node i.

[0029] In this embodiment, through the formula Multiple iterations (usually taking the number of iterations as 1) (to achieve deep penetration), thereby gradually aggregating the risk characteristics of the remote actual controller into the feature vector of the current trading entity.

[0030] In this embodiment, The calculation will incorporate business logic for biasing. For example, if there is a default execution relationship between node j and node i, the bias of this edge will be... The value will be automatically increased by the algorithm.

[0031] Step 3: Obtain the feature vectors of both parties in the transaction. and The feature vector of the current transaction and the attention coefficient between each side Explainable risk assessment and path backtracking are conducted to obtain a quantified risk score. And the Top-K paths of interpretable evidence.

[0032] In this embodiment, step 3 is the interpretability risk assessment and path backtracking step, which is responsible for quantifying transaction risks and extracting a visual evidence chain.

[0033] In this embodiment, step 3 includes the following steps: Step 3.1: Combine the topological vectors of both trading parties with the feature vector of the current transaction. Feature concatenation, followed by output of both parties to the transaction via a fully connected neural network (MLP). and Risk scores between Specifically, risk score The calculation method is as follows: ; Among them, the transaction feature vector The transaction amount, transaction type, region, and timestamp of the current auditable transaction are extracted in real time, and then concatenated after normalization and one-hot encoding; if Enter backtracking mode and extract the evidence chain path P.

[0034] Step 3.1: Based on The formula identifies the path with the largest product in the graph as the salient path; specifically, the salient score is calculated using the following formula: ; in, The significance score of path P is used; a higher score indicates a more significant impact of the associated path on risk. The top-K paths serve as the evidence chain for background checks, and the specific logic of risk transmission is visualized.

[0035] In this embodiment, step 3 transforms the mathematical path into a business graph, highlights the risk transmission chain in red, and automatically generates natural language review suggestions such as "due to the discovery of suspected transfer of benefits with related parties after penetrating through the second-tier subsidiary."

[0036] Step 4: Obtain the local subgraph For abnormal transactions that deliberately evade regulation (such as fictitious circular transactions), the isomorphism difference of the subgraph is calculated, and the entropy value of the node degree distribution is analyzed to identify anomalies in the topological structure, thus obtaining anomaly entropy values. It outputs a warning signal for manual review and intervention.

[0037] In this embodiment, step 4 is an adversarial anomaly detection and early warning step, which is responsible for identifying structurally abnormal corporate transaction backgrounds that evade review.

[0038] In this embodiment, step 4 includes the following steps: Step 4.1: Taking the transacting party as the center, extract all related entities within its 3rd-order neighborhood to form a subgraph.

[0039] Step 4.2: Calculate the anomaly entropy value of the local subgraph The formula is: ; in, Degree of nodes in the subgraph The probability distribution, The larger the value, the more complex and disordered the relational structure of the region, which may pose a risk of adversarial forgery.

[0040] In this embodiment, the network of connections of a normal enterprise typically follows a power-law distribution, meaning that a few nodes have high degree and most nodes have low degree.

[0041] Step 4.3: Perform degree distribution statistics on the target company's network of connections. When the entropy value is lower than the preset normal enterprise benchmark, a warning will pop up on the review interface.

[0042] In this embodiment, normal enterprise networks have natural density differences, while abnormal enterprise transaction backgrounds often exhibit highly regular and isomorphic topologies in order to disperse funds or conceal paths, which leads to a significantly lower entropy value H.

[0043] This embodiment of an interpretable intelligent assessment method for enterprise transaction risk firstly involves receiving a review request containing the identity identifiers and transaction characteristics of the transaction initiator and receiver, and verifying the timeliness of the current graph in real time through an evolutionary function. Listen for underlying change events To achieve the change in the map state from Towards The incremental advancement proceeds. Steps 2 and 3 then initiate L-wheel graph neural network iterations, penetrating and aggregating the complex and hidden risk transmission chains into a high-dimensional topological feature vector. This vector is then concatenated with transaction features and used by a risk assessment model to calculate the violation probability score within the [0, 1] interval. When the risk score exceeds a preset threshold, it is judged as high-risk and the path backtracking mechanism is automatically triggered to extract significant evidence chains. Simultaneously, step 4 quantifies the subgraph topological entropy value to identify structured abnormal enterprise transactions. If the score is below the threshold, it is marked as low-risk and a regular review path is generated. Finally, various indicators, visualized evidence chains, and review suggestions can be summarized to output a detailed intelligent review report and end the process.

[0044] This embodiment presents an interpretable intelligent assessment method for enterprise transaction risks. On one hand, it constructs a dynamically evolving panoramic knowledge graph. This enables real-time perception and in-depth analysis of enterprise background information. Furthermore, by integrating multi-source heterogeneous data and utilizing graph attention networks to perform high-dimensional representation of complex relationships, it can penetrate multi-layered nested structures and accurately identify hidden risks. Moreover, it not only uses risk probability... To provide early warnings, a more innovative approach is to introduce path saliency analysis based on attention coefficients. This enables visualized evidence backtracking for risk conclusions, significantly improving the credibility of review decisions. Furthermore, it utilizes abnormal entropy values ​​to target complex topological fraud deliberately created by malicious actors. An adversarial defense mechanism has been constructed, filling the gap in existing technologies for identifying structurally abnormal corporate transactions. In summary, using the method of this embodiment, the review system can be upgraded from passive information querying to proactive, penetrating investigation and logical reasoning. This significantly improves the efficiency of background investigations while effectively reducing financial compliance risks, providing a solid technical guarantee for the construction of a credit system and transaction security in modern business society.

[0045] Regarding the problems existing in the current technology: To address the issues of information lag and inability to perceive real-time changes in enterprise business and risk status in existing technologies, this paper proposes a dynamic heterogeneous knowledge graph construction and evolution process, utilizing evolutionary functions to capture incremental events. By transforming static industrial and commercial and judicial data into a dynamic graph structure that automatically updates over time T. This ensures the timeliness of risk assessment benchmarks and solves the problem of data expiration associated with traditional manual verification.

[0046] To address the difficulty in identifying complex penetration relationships and deep nested risk transmission in existing technologies, this paper proposes a relation representation learning step based on graph neural networks. It employs graph attention networks to perform high-dimensional modeling of enterprise nodes and evolves through multi-layered networks. This enables the model to automatically identify deep risks hidden behind multiple layers of equity or related relationships, thereby improving the accuracy of identifying concealed fraudulent methods.

[0047] To address the issues of existing technologies where models appear as black boxes and risk conclusions lack logical basis and interpretability, this paper proposes an interpretable risk assessment and path backtracking step. While outputting a risk score, it uses the path saliency formula Score(P) to inversely retrieve the evolutionary path with the largest attention weight product in the graph. This step transforms abstract feature vectors into a visualized chain of risk evidence, providing intuitive logical support for manual auditing and meeting the compliance requirements of financial regulation.

[0048] To address the challenge of preventing deliberate, regulatory-evasive, adversarial fraudulent transactions in existing technologies, this paper introduces subgraph sampling and entropy calculation formulas as part of an adversarial anomaly detection and early warning process. By analyzing the topological complexity and degree distribution anomalies of the transaction subgraph, it proactively identifies adversarial fraudulent activities disguised as legitimate business transactions through complex links. This step can discover abnormal patterns that are difficult for rule engines to access from the perspective of structural integrity, significantly enhancing the system's defensive capabilities.

[0049] Example 2: An interpretable intelligent assessment system for enterprise transaction risk, such as Figure 2 As shown, it includes: Data source layer: responsible for the real-time access and preprocessing of all heterogeneous data, providing multi-dimensional information support for background review.

[0050] Graph Engine Layer: Responsible for the construction and dynamic evolution of the graph, including entity and relation extraction, dynamic evolution, and incremental updates.

[0051] Core algorithm layer: Utilizes deep graph learning technology to extract deep topological features from enterprise interconnected networks, including graph attention networks, risk representation learning, and risk assessment models.

[0052] Application Output Layer: Provides visual and interpretable review conclusions for the final review process.

[0053] In this embodiment, the data source layer includes: business registration data: providing information on equity investments, legal entity relationships, and senior management appointments between corporate entities; judicial documents: using pre-trained NLP models (such as BERT) to extract entity and risk event triples from legal proceedings and enforcement announcements; news media: capturing unstructured negative public opinion through sentiment analysis and entity linking to identify non-financial related risks; and transaction data: accessing the dynamic characteristics of the transactions to be reviewed, such as amount, frequency, region, and industry attributes.

[0054] In this embodiment, the graph engine layer performs entity and relation extraction: mapping multi-source heterogeneous data into nodes and edges in a dynamic heterogeneous knowledge graph; dynamic evolution and incremental updates: listening for change events through a message queue and executing evolution functions. A local incremental operator is used to update only the affected adjacency list nodes, ensuring sub-second response capability of the graph structure and solving the lag problem of traditional modeling.

[0055] In this embodiment, the core algorithm layer includes: a graph attention network, which uses a graph attention mechanism to calculate attention coefficients and identify the risk contribution of neighboring nodes to the current subject; a risk representation learning model, which aggregates nodes into feature vectors in a high-dimensional space through a weight matrix and multi-layer nonlinear transformation, thereby achieving penetrating feature extraction of multi-layer nested structures; and a risk assessment model, which uses a multilayer perceptron to fuse the topological vectors of the two parties in a transaction to quantify potential transaction risks.

[0056] In this embodiment, the output layer outputs a risk score: a quantitative score. Path saliency analysis: Extracting evidence paths based on the product of attention coefficients to achieve risk traceability; Anomaly entropy detection: Targeting structured abnormal enterprise transactions, quantifying the subgraph topological entropy value. Achieve early warning of adversarial anomalies.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An interpretable intelligent assessment method for enterprise transaction risk, characterized in that, Includes the following steps: Step 1: Obtain structured and unstructured data from heterogeneous data sources such as the industrial and commercial system, court announcements, and news media; construct and evolve a dynamic heterogeneous knowledge graph to obtain a real-time updated dynamic heterogeneous knowledge graph. ; Step 2: Obtain a dynamically updated heterogeneous knowledge graph in real time. Using a graph attention network to focus on each enterprise node in the graph Deep feature extraction is performed, and by aggregating the features of neighboring nodes, the complex topological relationships are mapped into high-dimensional feature vectors to obtain enterprise nodes. The final layer feature vector ; Step 3: Obtain the feature vectors of both parties in the transaction. and The feature vector of the current transaction and the attention coefficient between each side Explainable risk assessment and path backtracking are conducted to obtain a quantified risk score. and Top-K paths of interpretable evidence; Step 4: Obtain the local subgraph To identify anomalies in the topology by calculating the isomorphism differences of the subgraph and analyzing the entropy values ​​of the node degree distribution for abnormal transactions that deliberately evade regulation, anomaly entropy values ​​are obtained. It outputs early warning signals for manual review and intervention.

2. The interpretable intelligent assessment method for enterprise transaction risk according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Use the pre-trained NLP model BERT to extract enterprise entities E, attributes A, and relationships R from heterogeneous data sources to construct a dynamic graph. Where T is the time axis; extract the triple relationship including the defendant, plaintiff and cause of action from the court announcement text; obtain the shareholding changes from the real-time data stream of the industry and commerce bureau; Step 1.3: Use a message queue to receive business registration change events. Execute the evolution function Incremental updates to the graph structure: ； in, For graph evolution operators, responsible for adding, deleting, and modifying the attributes of nodes and edges. The graph structure at the current moment. This is a newly triggered change event.

3. The interpretable intelligent assessment method for enterprise transaction risk according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Use the company's operating indicators as the initial vector. Through multi-layer W matrix learning, different types of nodes are mapped to a unified feature space. These different types of nodes include natural persons, shell companies, and group enterprises. Step 2.2: Use a graph attention network to obtain the deep topological features of each enterprise node; for each node... In the eigenvectors of the layer Aggregation is achieved through the following formula: ； in, () is the sigmoid function. Let i be the set of neighbors of node i. Here, W represents the attention coefficients, and W is the learnable shared weight matrix. Let j be a feature of the l-th layer neighbor node j, and let j be the risk contribution weight of the neighbor node i.

4. The interpretable intelligent assessment method for enterprise transaction risk according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Combine the topological vectors of both trading parties with the feature vector of the current transaction. Feature concatenation, outputting both parties to the transaction through a fully connected neural network. and Risk scores between Specifically, the risk score is calculated as follows: ； Among them, the transaction feature vector The transaction amount, transaction type, region, and timestamp of the current auditable transaction are extracted in real time and then concatenated after normalization and one-hot encoding; if Enter backtracking mode and extract the evidence chain path P; Step 3.1: Based on The formula identifies the path with the largest product in the graph as the salient path; specifically, the salient score is calculated using the following formula: ； in, The significance score of path P is used; a higher score indicates a more significant impact of the associated path on risk. The top-K paths serve as the evidence chain for background checks, and the specific logic of risk transmission is visualized.

5. The interpretable intelligent assessment method for enterprise transaction risk according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Taking the transacting party as the center, extract all related entities within its 3rd-order neighborhood to form a subgraph; Step 4.2: Calculate the anomaly entropy value of the local subgraph The formula is: ； in, Degree of nodes in the subgraph The probability distribution, The larger the value, the more complex and disordered the relational structure of the region, which may pose a risk of adversarial forgery. Step 4.3: Perform degree distribution statistics on the target company's network of connections. When the entropy value is lower than the preset normal enterprise benchmark, a warning will pop up on the review interface.

6. An interpretable intelligent assessment system for enterprise transaction risk, characterized in that, include: Data source layer: responsible for the real-time access and preprocessing of all heterogeneous data, providing multi-dimensional information support for background checks; Graph Engine Layer: Responsible for the construction and dynamic evolution of the graph, including entity and relation extraction, dynamic evolution and incremental updates; Core algorithm layer: Utilizes deep graph learning techniques to extract deep topological features from enterprise interconnected networks, including graph attention networks, risk representation learning, and risk assessment models; Application Output Layer: Provides visual and interpretable review conclusions for the final review process.

7. The interpretable intelligent assessment method for enterprise transaction risk according to claim 6, characterized in that, The data source layer includes: business registration data: providing information on equity investments, legal relationships, and senior management appointments between corporate entities; judicial documents: using pre-trained NLP models to extract entity and risk event triples from legal proceedings and enforcement announcements; news media: capturing unstructured negative public opinion through sentiment analysis and entity linking, and identifying non-financial related risks; and transaction data: accessing the dynamic characteristics of transactions under review.

8. The interpretable intelligent assessment method for enterprise transaction risk according to claim 6, characterized in that, In the graph engine layer, entity and relation extraction: mapping multi-source heterogeneous data into nodes and edges in a dynamic heterogeneous knowledge graph; dynamic evolution and incremental update: listening for change events through a message queue and executing evolution functions.

9. The interpretable intelligent assessment method for enterprise transaction risk according to claim 6, characterized in that, In the core algorithm layer, the graph attention network calculates attention coefficients using a graph attention mechanism to identify the risk contribution of neighboring nodes to the current subject. Risk representation learning: Through weight matrix and multi-layer nonlinear transformation, nodes are aggregated into feature vectors in high-dimensional space, realizing penetrating feature extraction of multi-layer nested structures; Risk assessment model: The topological vectors of both parties in a transaction are fused using a multilayer perceptron to quantify potential transaction risks.

10. The interpretable intelligent assessment method for enterprise transaction risk according to claim 6, characterized in that, In the application output layer, the risk score is output: the quantitative score is output. Path saliency analysis: Extracting evidence paths based on the product of attention coefficients to achieve risk traceability; Anomaly entropy detection: Targeting structured abnormal enterprise transactions, quantifying the subgraph topological entropy value. Achieve early warning of adversarial anomalies.

Images