Enterprise association conduction risk evaluation method based on fusion graph neural network

Abstract

The invention provides an enterprise association conduction risk evaluation method based on a fusion graph neural network, and belongs to the field of general finance, and the method mainly comprises the steps: 1) building an enterprise association relation graph; 2) labeling enterprise nodes in the enterprise relation graph based on default samples; 3) extracting enterprise characteristics and graph characteristics to establish an enterprise risk index assessment system; 4) training an enterprise association conduction risk evaluation model based on a graph neural network, 5) building an enterprise association conduction risk evaluation system, mining enterprise association relationship characteristics to evaluate the enterprise association conduction risk in real time, and carrying out real-time visual display on the enterprise association relationship, the self risk of the enterprise and the associated enterprise, and the association conduction risk.

Description

Technical Field

[0001] This invention relates to the field of inclusive finance, specifically to the application of financial technologies such as machine learning and big data in the field of corporate financial risk assessment, and more specifically to a method for assessing corporate correlation transmission risk by integrating graph neural networks. Background Technology

[0002] Currently, fintech technologies such as machine learning and graph databases are widely used in the field of financial risk assessment. Machine learning algorithms can extract key features from large amounts of financial data, build accurate risk assessment models, and achieve risk early warning and control. Through pattern recognition and predictive analysis, it effectively improves the accuracy and efficiency of risk judgment. Graph databases, with their unique data structure and efficient query capabilities, play an important role in processing complex relational data and achieving real-time risk assessment. They excel at expressing and traversing the relationships between entities, providing underlying support for risk transmission analysis. As an emerging technology, graph neural networks, combining graph structures and deep learning, can effectively capture complex nonlinear relationships between enterprises, providing a more powerful tool for analyzing risk transmission. Through node embedding and message passing mechanisms, it deeply mines potential risk signals in the relational network. Enterprise relational transmission risk assessment emphasizes network effects, which traditional methods struggle to address. Graph neural networks, by modeling relational paths, can accurately identify and quantify the transmission path and scope of risk, providing crucial support for systemic risk prevention and control.

[0003] Post-loan risk monitoring and assessment is a crucial part of the entire corporate financial lending process. Traditional monitoring often focuses on risks already incurred by the company itself, failing to effectively cover risks transmitted through related parties such as shareholders, investors, and supply chains. These related-party transmission risks are highly concealed and have a wide impact, posing a serious challenge to loan approval decisions and post-loan early warning management. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention provides a method for assessing enterprise-related risk transmission by integrating graph neural networks. This method deeply integrates machine learning, graph databases, and big data technologies, focusing on the accurate identification and quantitative evaluation of enterprise-related risk transmission. It can be applied not only to pre-loan relationship screening and risk penetration review during the loan process, but also to post-loan dynamic tracking and early warning of risk transmission paths. This expands the scope of risk monitoring from a single entity to the entire related network, providing core support for risk prevention and control throughout the entire credit process. By establishing a comprehensive risk assessment system, potential enterprise risks can be identified in a timely manner, their impact assessed, and corresponding prevention and control measures implemented to ensure the smooth operation of business activities.

[0005] The technical solution of this invention is: A method for assessing enterprise-related transmission risks by incorporating graph neural networks includes the following steps: 1) Based on the target customer group enterprises and the multi-source data of enterprises, extract the enterprise relationship to establish a relationship graph. The nodes on the graph represent enterprises, and the edges represent the relationship between enterprises. The relationship includes two categories: direct relationship and suspected relationship. The weight of the edge is the weight of each relationship.

[0006] 2) Organize the training samples of the supervised model, and based on the default and non-default samples in the target customer group, identify whether the enterprise nodes in the relationship graph are in default. Unidentified enterprise nodes are enterprises whose relationship transmission risks need to be predicted.

[0007] 3) Establish an enterprise risk indicator assessment system. The indicator system covers the enterprise's own characteristics and graph characteristics. The enterprise's own characteristics are extracted from its own basic attributes and risk attributes, and the enterprise's relational characteristics are extracted from the enterprise's relational graph as the enterprise's graph structure characteristics.

[0008] 4) The enterprise's own characteristics and relationship graph structure characteristics are fused to form enterprise association risk characteristics. Based on the association risk characteristics after fusion of training samples, a graph neural network algorithm is used to train a supervised machine learning model. Based on the trained supervised model, the probability of enterprise association transmission risk of unknown nodes is predicted.

[0009] 5) Establish an enterprise-related risk assessment system, including updating enterprise relationship graph data, extracting enterprise-related risk features, training and updating supervised machine learning models, and visualizing enterprise relationships and transmission risks.

[0010] Furthermore, Construct a corporate relationship graph, which defines the target customer group and the hierarchy and types of relationships among the enterprises: The database selected for building the enterprise relationship graph is not limited; only the extraction process is restricted. The enterprise relationship graph is built based on a specific target customer group. The types of enterprise relationships mainly include direct relationships and suspected relationships. Direct relationships include branches, external investments, corporate shareholders, individual shareholders, directors, senior management, supervisors and managers, and historical directors, senior management, supervisors and managers. Suspected relationships include relationships with the same actual controller, the same phone number, the same email address, the same mailing address, the same registered address, the same domain name information, the same court documents, the same company name, the same patent information, and the same software copyright.

[0011] Based on the annotation of default samples in the enterprise relationship graph, the annotation method for enterprise nodes in the enterprise relationship graph is as follows: 1) Organize the training samples of the supervised model. Based on the default situation of enterprises that have been lent by financial institutions in the past and the default risk indicators in the public data, organize the training samples in the target customer group. The samples mainly include three parts: default samples, non-default samples, and samples of default situations to be predicted. The sample data is mainly stored in a relational database, and each sample is mainly represented by a unique enterprise identifier. 2) Locate known defaulting enterprise samples in the relationship graph and clearly mark the risk status attributes of their nodes as "defaulted", "non-defaulted" or "unknown". Nodes marked as "defaulted" are the starting point of the entire risk transmission network. By directly marking them, all known risk outbreak points can be accurately anchored in the complex network. 3) Label the risk attributes of enterprise nodes in the relationship graph. Starting from each "defaulted" node, traverse all its first-degree related enterprises. For these first-degree related neighbors, assign the "related risk" attribute according to the type and strength of their association with the risk source.

[0012] Constructing a risk model for enterprise-related transmission, integrating enterprise-specific risks and enterprise graph characteristic risks, and the construction method of the enterprise-related transmission risk assessment model: 1) The enterprise risk indicator system consists of two parts: enterprise risk characteristics and enterprise graph characteristics. Enterprise characteristics are extracted from enterprise data in the Spark environment, mainly through a typical ETL process of extraction, transformation, and loading. Spark SQL is used to perform efficient distributed join and aggregation operations on the original, multi-source heterogeneous enterprise basic and risk data. Normalization is performed using StandardScaler or MinMaxScaler in the Spark MLlib library to process categorical features, numerical indexes, text and high-risk features respectively. Vectorization is performed using TF-IDF or Word2Vec in Spark MLlib. The core risk labels of each enterprise node, such as "whether it has defaulted" and "whether it has been executed", serve as supervision signals for model training. The technical framework used in the enterprise feature processing is not limited. 2) The GAT graph neural network algorithm is used to extract enterprise graph features and fuse enterprise features with enterprise graph features. Spark GraphFrames is mainly used to manage and manipulate the enterprise relationship graph, storing node and edge vectors separately. Spark's distributed in-memory computing model is used to process large-scale graphs with billions of nodes and edges. The core mechanism of GAT is used for attention-driven feature propagation. The propagation process allows the model to assign different importance to different neighbors to achieve differentiated modeling of risk transmission. In order to capture the relationship transmission effect over a longer distance, multiple layers of GAT are stacked to achieve multi-layer stacking and risk transmission perception. 3) A supervised enterprise association transmission risk assessment model was trained based on the GAT algorithm. After multiple rounds of GAT layer iterations, each enterprise node in the graph learned a new graph-enhanced feature vector that integrates its own attributes and surrounding association risks. This graph-enhanced feature vector was then used as the final feature input to a fully connected neural network for supervised learning. The output layer predicted a binary label and generated a transmission risk score based on the learned complex features. The training process was mainly carried out on Spark using a distributed gradient descent optimization algorithm. Cross-entropy loss function was used for classification. Spark's parallel capabilities were utilized to distribute large-scale graph data and model parameters across multiple nodes in the cluster for computation. To prevent overfitting, Dropout, L2 regularization, and early stopping techniques were widely used in the model.

[0013] Components and implementation methods of an enterprise-related risk assessment system: The enterprise-related risk assessment system mainly includes functions such as updating enterprise relationship graph data, extracting enterprise-related risk characteristics, enterprise relationship, real-time prediction and visualization of transmission risks.

[0014] 1) Real-time data pipeline and dynamic graph update: The system mainly builds a real-time data communication pipeline based on Kafka components. Various data sources are continuously pushed to different Kafka topics. Spark Streaming is used as a reliable consumer to continuously pull these streaming data in a micro-batch processing manner. Within the Spark cluster, the system first cleans, transforms, and parses the data relationships. The processed incremental graph data is updated in real time to a graph database (such as Neo4j) or distributed storage to ensure that the enterprise relationship graph is always up-to-date.

[0015] 2) Integrated Streaming and Batch Processing Risk Feature Engineering: The feature extraction process adopts an integrated streaming and batch processing design. For the enterprise's own static attribute system, Spark SQL is used to periodically (batch process) extract and calculate from the data warehouse. For graph structure features with high real-time requirements, when the graph is updated or a specific event is triggered, the system will start a real-time Spark job. This job uses the GraphFrames API and calculates the graph structure features of the enterprise nodes of interest in real time based on the latest graph snapshot. Finally, the static attributes and dynamic graph features are concatenated and merged in Spark memory to form a comprehensive, real-time enterprise-related risk feature vector.

[0016] 3) Online inference and visualization interaction of model service: The trained GNN model is packaged into an independent service through the model export function of Spark MLlib. When the real-time feature vector is ready, the system will send it to the model service for online inference to obtain the enterprise's correlation transmission risk score at the current moment. This score result, along with the correlation path behind it, will be pushed to the front end for visualization through real-time communication technology. The front end will intuitively present the complex correlation network and risk transmission path to risk control personnel through visualization, supporting their drill-down query and interactive exploration, thereby completing the closed loop from data to insight.

[0017] The beneficial effects of this invention are 1. Compared with traditional risk assessment models that only focus on the static characteristics of the enterprise itself, this invention proposes a method that integrates graph neural network algorithms to analyze the characteristics of the enterprise itself and the characteristics of the enterprise relationship graph to evaluate the risk of enterprise relationship transmission. This invention constructs an enterprise relationship graph and uses graph neural network (GNN) technology to deeply integrate the characteristics of the enterprise itself with the complex graph structure features, realizing a paradigm shift in risk assessment from isolated individuals to network relationships. This enables the model to accurately capture and quantify the transmission effect of risk along multi-dimensional paths such as equity, guarantees, and supply chains, realizing a leap from point-like risk assessment to networked risk insight, and greatly expanding the depth and breadth of risk monitoring. 2. Compared with traditional graph embedding methods, the core of this invention adopts the GAT graph neural network algorithm, which uses graph attention network to achieve differentiated risk transmission quantification. Its inherent attention mechanism can adaptively assign differentiated importance weights to different relationships. This means that the model can automatically identify and pay more attention to those relationships with high risk transmission capacity (such as large guarantees), while weakening the influence of secondary relationships, thereby achieving a more refined and accurate quantitative assessment of the risk transmission of relationships and improving the accuracy of risk assessment. 3. This invention builds a complete risk assessment system based on real-time big data frameworks such as Kafka and Spark, and constructs an end-to-end real-time big data processing and decision support system. It realizes integrated batch processing of the entire chain from data access, graph update, feature calculation to model inference. This not only ensures the real-time perception and assessment of enterprise risk status, but also intuitively presents the complex risk transmission path through visualization technology, providing risk control personnel with dynamic and interactive decision-making basis, and significantly improving the timeliness and effectiveness of risk prevention and control. 4. The system architecture of this invention is highly flexible in design, providing a highly scalable and flexible technical architecture. Based on the Spark distributed computing framework, the system can be horizontally scaled to cope with the ever-increasing massive graph data and computing needs. At the same time, the modular design allows for the flexible introduction of new association types and risk characteristics, and can be easily integrated with real-time computing components such as Spark Streaming, ensuring the system's strong adaptability in the face of future business changes and technological developments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This invention proposes a method for assessing enterprise-related risk transmission by integrating graph neural networks. First, a graph database technology is used to construct an enterprise relationship graph. Based on default and non-default samples from the target customer group, enterprise nodes in the relationship graph are identified. Enterprise-specific features are extracted from their own basic attributes and risk attributes, and graph structure features are extracted from the enterprise relationship graph. These enterprise-specific features and graph structure features are then fused to form enterprise-related risk features. A supervised machine learning model is trained using a graph neural network algorithm. Finally, an enterprise-related risk transmission assessment system is established. The main implementation steps are as follows: I. Construction of Enterprise Relationship Map 1. The types of corporate relationships mainly include two categories: direct corporate relationships and suspected corporate relationships. Direct corporate relationships include branch offices, external investments, corporate shareholders, natural person shareholders, directors, senior executives and supervisors, and historical directors, senior executives and supervisors, etc. Suspected corporate relationships include relationships with the same actual controller, the same phone number, the same email address, the same mailing address, the same registered address, the same domain name information, the same court documents, the same company name, the same patent information, the same software copyright, etc. 2. Enterprise Relationship Extraction: Using multi-source data of self-owned enterprises as the extraction source, the extraction rules for each relationship type are determined. The multi-source data of enterprises is mainly stored in the distributed relational database TiDB. The relationship type extraction rules mainly include the extraction source table, the relationship extraction logic, and the extraction fields. The extraction of various relationships forms a three-dimensional data structure of <enterprise, relationship, enterprise>. Each enterprise is identified by its unique enterprise identifier, and the relationship is identified by the relationship name. 3. Enterprise Relationship Graph Construction: Based on the target customer group enterprises and multi-source enterprise data, enterprise relationship graphs are extracted and constructed using NebulaGraph graph database technology. Nodes in the graph represent enterprises, and edges represent relationships between enterprises. There are a total of 12 types of enterprise relationships, including direct relationships and suspected relationships. The weight of each edge is the weight of the relationship, and the arrow of the edge indicates the direction of the relationship. Direct relationships such as branch offices, external investments, and enterprise shareholders are all unidirectional relationships, while suspected relationships such as the same phone number or email address are all bidirectional relationships.

[0021] II. Mark enterprise nodes in the enterprise relationship graph based on default samples. 1. Organize the training samples for the supervised model: The enterprise nodes in the enterprise relationship graph cover the target customer group within a specific range. Based on the default situation of enterprises that have been lent to by financial institutions such as banks and guarantee companies in the past, and the default risk indicators in the public data, we organize the training samples for the supervised model in the target customer group. The samples mainly include three parts: default samples, non-default samples, and samples of default situations to be predicted. The sample data is mainly stored in a relational database. Each sample is mainly represented by a unique identifier such as the enterprise's unified credit code. 2. Enterprise node labeling in the relationship graph: Locate the known defaulting enterprise samples in the graph and clearly mark the risk status attribute of their nodes as "defaulted", "non-defaulted" or "unknown". The nodes marked as "defaulted" are the starting point of the entire risk transmission network and the foundation point for all subsequent risk analysis. Through direct labeling, all known risk outbreak points are accurately anchored in the complex network. 3. Risk Attribute Labeling for Enterprise Nodes in the Relationship Graph: Risk is primarily transmitted through interconnected edges such as equity chains, guarantee circles, and supply chains. Starting from each "defaulted" node, the process iterates through all its first-degree related enterprises (i.e., directly connected enterprises). For these first-degree related neighbors, the "related risk" attribute is assigned based on the type and strength of their association with the risk source. For example, if enterprise A provides a large guarantee to enterprise B and A has defaulted, then enterprise B should be marked as "high risk"; if enterprise C is a core supplier of enterprise D and C has defaulted, then enterprise D may be marked as "supply chain risk." This step extends the scope of the risk's impact from a single point to its directly connected related enterprises.

[0022] III. Constructing an Enterprise-Related Transmission Risk Assessment Model Based on Graph Neural Network Algorithm This invention uses graph neural network algorithm to construct a risk assessment model for enterprise correlation transmission. It mainly includes two parts: establishing an enterprise risk indicator assessment system, which mainly involves extracting enterprise characteristics and enterprise graph features, and training a supervised enterprise correlation transmission risk assessment model based on the GAT algorithm.

[0023] 1. Enterprise-Specific Feature Extraction: The enterprise risk indicator assessment system mainly includes two parts: enterprise-specific features and enterprise graph features. Enterprise-specific features are extracted from enterprise-owned data such as basic business registration information, legal litigation records, operational data, intellectual property data, and financial data. In this invention, the extraction of enterprise-specific features is primarily implemented in the Spark environment, through a typical ETL process of extraction, transformation, and loading. Spark SQL is mainly used to perform efficient distributed joins and aggregation operations on the raw, multi-source, heterogeneous enterprise basic and risk data. For numerical features such as registered capital, years of establishment, and revenue scale, Spark... The StandardScaler or MinMaxScaler in the MLlib library are used for normalization to eliminate the influence of units and accelerate model convergence. For categorical features such as industry classification, region, and enterprise type, StringIndexer is used to convert them into numerical indices, and then OneHotEncoder is used for one-hot encoding so that they can be effectively learned by the model. For textual and high-risk label features such as key text information extracted from judicial documents and public opinion announcements, TF-IDF or Word2Vec in SparkMLlib are mainly used for vectorization. The core risk labels of each enterprise node, such as "whether it has defaulted" and "whether it has been executed", serve as supervision signals for model training. 2. Extraction of enterprise graph features and fusion with its own features: Traditional graph embedding methods (such as Node2Vec) cannot dynamically integrate node features, while the advantage of GAT is that it can learn node features and graph structure information end-to-end. In this invention, the graph neural network algorithm GAT is mainly used to extract enterprise graph features and realize the fusion of enterprise features with graph features.

[0024] (1) Graph structure definition and distributed storage: In this invention, Spark GraphFrames is mainly used to manage and operate enterprise association graphs. GraphFrames represents graph data as two Spark DataFrames, one for nodes (containing the enterprise's own feature vectors generated above) and one for edges (containing attributes such as association type, shareholding ratio, and guarantee amount), that is, the association edges between nodes. This form is naturally in line with Spark's distributed memory computing model and can easily handle large-scale graphs with billions of nodes and edges. (2) The core mechanism of GAT is attention-driven feature propagation: Each enterprise node initializes its node features using its own feature vector as the initial node representation; for each central node in the graph, GAT calculates the attention coefficient for all its first-order neighbor nodes. The calculation process does not use a simple weighted average, but instead uses a learnable weight matrix W to map the feature vector of each node (including itself) to a higher-dimensional latent space. Then, an attention mechanism a (usually a single-layer feedforward neural network) is used to calculate the correlation score between the central node and its neighbor nodes. The formula can be simplified to: e_ij = a(W * h_i, W * h_j), where h_i and h_j are the features of nodes i and j, respectively, and e_ij represents the importance of node j to node i; in order to make the attention coefficient easy to compare between different nodes, the attention weight is normalized using the softmax function, α_ij = softmax_j(e_ij) = exp(e_ij) / Σ_(k∈N_i) exp(e_ik), where α_ij This refers to the final, normalized attention weight, which clearly indicates the contribution of neighbor node j to the central node i in this information aggregation. For feature aggregation and updating, the new representation of the central node is obtained by weighting and summing the features of all its neighbors according to their attention weights, and then passing it through a non-linear activation function (such as ELU), h_i' = σ( Σ_(j∈N_i) α_ij * W * h_j ). This process allows the model to assign different importance to different neighbors. For example, an association "guaranteed by a severely defaulting company" will have a much higher attention weight than an association "with only a weak equity investment," thus achieving differentiated modeling of risk transmission. (3) Multi-layer stacking and risk transmission perception: A single-layer GAT can only aggregate information of first-order neighbors. In order to capture the correlation transmission effect over a longer distance (such as the risk brought by "friends of friends"), multiple layers of GAT will be stacked. In the second layer, each node will aggregate the feature representation of its first layer output, which already contains the information of its first-order neighbors. Thus, after K layers of GAT, the final representation of each node will contain the topological information and feature information of all its neighbors within K hops, thereby achieving deep perception of the risk transmission path.

[0025] 3. Supervised Enterprise Linkage Transmission Risk Assessment Model Based on GAT Algorithm: After multiple iterations of GAT layers, each enterprise node in the graph learns a novel graph-augmented feature vector that integrates its own attributes and surrounding associated risks. Compared to the initial static features, this vector provides a more profound and forward-looking characterization of risk. This graph-augmented feature vector is used as the final feature input to a fully connected neural network for supervised learning. The task of the output layer is to predict a binary label, such as whether a risk event will occur within the next 90 days, or a regression value, such as a transmission risk score, based on the learned complex features. In this invention, the training process is mainly carried out on Spark using a distributed gradient descent optimization algorithm (such as Adam), using the cross-entropy loss function for classification, and leveraging Spark's parallel capabilities to distribute large-scale graph data and model parameters across multiple nodes in the cluster for computation, greatly shortening the training time. To prevent overfitting, Dropout, L2 regularization, and early stopping techniques are widely used in the model.

[0026] IV. Establishment of Enterprise-Related Transmission Risk Assessment System The enterprise-related risk assessment system mainly includes functions such as updating enterprise relationship graph data, extracting enterprise-related risk features, training and updating supervised machine learning models, real-time prediction and visualization of enterprise relationships and transmission risks. This invention utilizes a real-time big data framework to build the enterprise-related risk assessment system, primarily using Kafka and Spark to achieve real-time data transmission and analysis of enterprise-related risk transmission within the assessment system.

[0027] 1. Real-time Data Pipeline and Dynamic Graph Update: This invention primarily builds a real-time data communication pipeline based on the Kafka component. Various data sources, such as changes in business registration information, equity pledges, legal proceedings, and public opinion events, are continuously pushed to different Kafka topics. Spark Streaming jobs, acting as reliable consumers, continuously pull these streaming data in a micro-batch processing manner. Within the Spark cluster, the system first cleans, transforms, and parses the data relationships. For example, from a set of equity change data, the investor and investee are extracted, and a new "investment" relationship edge is constructed. The processed incremental graph data is updated in real time to a graph database (such as Neo4j) or distributed storage to ensure that the enterprise relationship graph is always up-to-date, providing an accurate network topology foundation for risk transmission analysis. 2. Integrated Streaming and Batch Processing Risk Feature Engineering: The feature extraction process adopts an integrated streaming and batch processing design. For the static attributes of the enterprise itself (such as registered capital and industry category), the system extracts and calculates them from the data warehouse periodically (batch processing) using Spark SQL. For graph structure features with high real-time requirements, when the graph is updated or a specific event is triggered, the system will start a real-time Spark job. This job uses the GraphFrames API to calculate the network centrality index (such as PageRank value and degree) and other graph structure features of the enterprise nodes of interest in real time based on the latest graph snapshot, or uses an efficient graph traversal algorithm to identify the shortest transmission path and distance between the enterprise and known risk sources. Finally, the static attributes and dynamic graph features are spliced ​​and fused in Spark memory to form a comprehensive, real-time enterprise-related risk feature vector. 3. Online Inference and Visual Interaction of Model Services: The trained GNN model is encapsulated into an independent service through the model export function of Spark MLlib. When the real-time feature vector is ready, the system will send it to the model service for online inference to obtain the risk score of the enterprise at the current moment. This score, along with the underlying correlation path such as "due to the default of Company A, the risk is transmitted to the target enterprise through the guarantee chain", will be pushed to the front end for visualization through real-time communication technology. The front end will intuitively present the complex correlation network and risk transmission path to risk control personnel through visualization, supporting their drill-down queries and interactive exploration, thereby completing the closed loop from data to insight.

[0028] The above description is merely a preferred embodiment of the present invention and is used only to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for assessing enterprise-related transmission risks by incorporating graph neural networks, characterized in that, Includes the following steps: 1) Based on the target customer group enterprises and the multi-source data of enterprises, extract the enterprise relationship to establish a relationship graph. The nodes on the graph represent enterprises, and the edges represent the relationship between enterprises. The relationship includes two categories: direct relationship and suspected relationship. The weight of the edge is the weight of each relationship. 2) Organize the training samples of the supervised model, and based on the default and non-default samples in the target customer group, identify whether the enterprise nodes in the relationship graph are in default. Unidentified enterprise nodes are enterprises whose relationship transmission risk is to be predicted. 3) Establish an enterprise risk indicator assessment system. The indicator system covers the enterprise's own characteristics and graph characteristics. The enterprise's own characteristics are extracted from its own basic attributes and risk attributes, and the enterprise's relational characteristics are extracted from the enterprise's relational graph as the enterprise's graph structure characteristics. 4) The enterprise's own characteristics and relationship graph structure characteristics are fused to form enterprise association risk characteristics. Based on the association risk characteristics after fusion of training samples, a graph neural network algorithm is used to train a supervised machine learning model. Based on the trained supervised model, the probability of enterprise association transmission risk of unknown nodes is predicted. 5) Establish an enterprise-related risk assessment system, including updating enterprise relationship graph data, extracting enterprise-related risk features, training and updating supervised machine learning models, and visualizing enterprise relationships and transmission risks.

2. The method according to claim 1, characterized in that, In step 1), direct relationships between enterprises include branches, external investments, corporate shareholders, individual shareholders, directors, senior executives and supervisors, and historical directors, senior executives and supervisors; suspected relationships between enterprises include the same actual controller, the same phone number, the same email address, the same mailing address, the same registered address, the same domain name information, the same court documents, the same company name, the same patent information, and the same software copyright.

3. The method according to claim 1, characterized in that, The labeling of enterprise nodes in the enterprise relationship graph in step 2) includes: 1) Organize the training samples of the supervised model. Based on the default situation of enterprises that have been lent by financial institutions in the past and the default risk indicators in the public data, organize the training samples in the target customer group. The samples include three parts: default samples, non-default samples, and samples of default situation to be predicted. The sample data is stored in a relational database, and each sample is represented by a unique enterprise identifier. 2) Locate known defaulting enterprise samples in the relationship graph and clearly mark the risk status attributes of their nodes as "defaulted", "non-defaulted" or "unknown". Nodes marked as "defaulted" are the starting point of the entire risk transmission network. By directly marking them, all known risk outbreak points can be accurately anchored in the complex network. 3) Label the risk attributes of enterprise nodes in the relationship graph. Starting from each "defaulted" node, traverse all its first-degree related enterprises. For these first-degree related neighbors, assign the "related risk" attribute according to the type and strength of their association with the risk source.

4. The method according to claim 1, characterized in that, The enterprise risk indicator system comprises two parts: enterprise-specific risk characteristics and enterprise graph characteristics. Enterprise-specific characteristics are extracted from enterprise-owned data within a Spark environment through a typical ETL process of extraction, transformation, and loading. Spark SQL is used to perform efficient distributed joins and aggregation operations on the raw, multi-source, heterogeneous enterprise basic and risk data. Normalization is performed using StandardScaler or MinMaxScaler from the Spark MLlib library, handling categorical features, numerical indexes, textual features, and high-risk features respectively. Vectorization is achieved using TF-IDF or Word2Vec from Spark MLlib. The risk labels for each enterprise node—"whether it has defaulted" and "whether it has been executed"—serve as supervisory signals for model training.

5. The method according to claim 1, characterized in that, The GAT graph neural network algorithm is used to extract enterprise graph features and fuse enterprise features with enterprise graph features. Spark GraphFrames is used to manage and manipulate the enterprise relationship graph, storing node and edge vectors separately. Spark's distributed in-memory computing model is used to process large-scale graphs with billions of nodes and edges. The core mechanism of GAT is used for attention-driven feature propagation. The propagation process allows the model to assign different importance to different neighbors to achieve differentiated modeling of risk transmission. In order to capture the relationship transmission effect over longer distances, multiple layers of GAT are stacked to achieve multi-layer stacking and risk transmission perception.

6. The method according to claim 1, characterized in that, A supervised enterprise association transmission risk assessment model is trained based on the GAT algorithm. Through multiple iterations of the GAT layer, each enterprise node in the graph learns a new graph-enhanced feature vector that integrates its own attributes and surrounding association risks. This graph-enhanced feature vector is then used as the final feature input to a fully connected neural network for supervised learning. The output layer predicts a binary label and generates a transmission risk score based on the learned complex features. The training process is mainly carried out on Spark using a distributed gradient descent optimization algorithm, with cross-entropy loss function used for classification. Spark's parallel capabilities are utilized to distribute large-scale graph data and model parameters across several nodes in the cluster for computation. To prevent overfitting, Dropout, L2 regularization, and early stopping techniques are employed in the model.

7. The method according to claim 1, characterized in that, Real-time updates to the data pipeline and graph dynamics are achieved by building a real-time data communication pipeline based on Kafka components. Various data sources are continuously pushed to different Kafka topics. Spark Streaming is used as a reliable consumer to continuously pull these streaming data in a micro-batch processing manner. Within the Spark cluster, the data is first cleaned, transformed, and its relationships are parsed. The processed incremental graph data is then updated in real time to the graph database or distributed storage to ensure that the enterprise's relationship graph is always up-to-date.

8. The method according to claim 7, characterized in that, The feature extraction process adopts a stream-batch integrated design. For the enterprise's own static attribute system, Spark SQL is used to periodically extract and calculate from the data warehouse. For graph structure features with high real-time requirements, a real-time Spark job is launched when the graph is updated or a specific event is triggered. This job uses the GraphFrames API to calculate graph structure features for the enterprise nodes of interest in real time based on the latest graph snapshot. Finally, the static attributes and dynamic graph features are concatenated and merged in Spark memory to form a comprehensive, real-time enterprise association risk feature vector.

9. The method according to claim 8, characterized in that, The trained GNN model is encapsulated into an independent service through Spark MLlib's model export function. Once the real-time feature vector is ready, it is sent to the model service for online inference to obtain the enterprise's correlation transmission risk score at the current moment. This score, along with the underlying correlation path, is pushed to the front end for visualization through real-time communication technology. The front end uses visualization to intuitively present the complex correlation network and risk transmission path to risk control personnel, supporting their drill-down queries and interactive exploration, thereby completing the closed loop from data to insight.

Images