Relevance credit default early warning method based on graph neural network

Abstract

The invention discloses a graph neural network-based relevance credit default early warning method. The method comprises the following steps of S1, performing multi-source cross-mechanism data fusion and dynamic graph construction; s2, designing a space-time diagram neural network model to output a final risk score; s3, a federal learning framework: local training: locally training a sub-graph model by each participation mechanism, and retaining sensitive data; parameter aggregation: aggregating gradient information by the central server, and adding Gaussian noise by using differential privacy; model updating: the space-time diagram neural network model supports a heterogeneous graph structure through weighted average updating parameters; and S4, carrying out risk early warning and interpretability output. The method at least has the following beneficial effects: comprehensive risk coverage: dynamic heterogeneous graph construction: integrating multi-source heterogeneous data and dominant / implicit relationships, constructing a dynamic graph comprising enterprises, individuals and geographic nodes, dynamically adjusting edge weights through a time decay function, and quantifying timeliness of association strength;

Description

Technical Field

[0001] The present application belongs to the field of data processing technology, and specifically relates to a correlation credit default warning method based on graph neural network. Background Art

[0002] In the credit business, the qualifications and repayment ability of financial users directly determine the size of the financial risks that banks and other financial institutions have to bear.

[0003] Traditional credit risk assessment methods have the following significant flaws:

[0004] Omission of correlation risk: Traditional models (such as logistic regression and random forest) rely only on individual static characteristics (income, credit records), ignoring explicit and implicit connections between entities (such as supply chain transactions, equity control, and social networks), resulting in insufficient predictive ability for risk transmission (such as broken guarantee chains and group defaults).

[0005] Lack of dynamic modeling: Existing methods do not integrate time series characteristics (such as fluctuations in trading frequency and changes in economic cycles), making it difficult to capture the evolution of medium- and long-term risks and resulting in poor early warning timeliness.

[0006] Weak privacy protection: When collaborating on data across institutions, traditional centralized training requires sharing of original data, which violates privacy regulations such as the GDPR. Existing graph neural networks do not integrate federated learning and differential privacy technologies, posing a high risk of data leakage. Summary of the Invention

[0007] This application provides a correlation credit default warning method based on graph neural network, which aims to solve the problems of omission of correlation risk, lack of dynamic modeling, and weak privacy protection in the existing technology.

[0008] A correlation credit default warning method based on a graph neural network, the method comprising the following steps:

[0009] S1: Fusion of multi-source and cross-institutional data and construction of dynamic graphs;

[0010] S2: Combine cross-institutional data and dynamic graphs to design a spatiotemporal graph neural network model to output the final risk score;

[0011] S3: Deploy a federated learning framework to support joint modeling of cross-institutional data:

[0012] Local training: Each participating institution trains the subgraph model locally, retaining sensitive data;

[0013] Parameter aggregation: The central server aggregates gradient information and adds Gaussian noise using differential privacy;

[0014] Model update: The spatiotemporal graph neural network model updates parameters through weighted averaging;

[0015] S4: Risk warning and explainable output, combining risk scoring and joint modeling to output warning content.

[0016] Optionally, S1 includes:

[0017] S1.1: Data collection and preprocessing: Collect structured data, unstructured data, social data, and perform time series data processing;

[0018] S1.2: Dynamic heterogeneous graph construction.

[0019] Optionally, S1.2 includes:

[0020] S1.2.1: Determine the node type:

[0021] Enterprise node: Attributes include industry classification, debt-to-asset ratio, and equity structure;

[0022] Personal nodes: Attributes include career stability, long-term lending index, and social network centrality;

[0023] Geographic nodes: Attributes include regional GDP growth rate and judicial enforcement case density;

[0024] S1.2.2: Edge relationship definition:

[0025] Explicit relationships: equity holdings, supply chain transactions;

[0026] Implicit relationships: social closeness, address overlap;

[0027] S1.2.3: Dynamic edge weight update:

[0028] The time decay function is used to dynamically adjust the edge weights. The formula is:

[0029]

[0030] Among them, w ij (t) represents the correlation strength between node i and node j at time t, β is the transaction frequency weight coefficient, and transaction frequency (t) is the transaction frequency index, which refers to the normalized value of the number or amount of transactions between nodes i and j within the time window t. t is the current time, t 最新 is the time of the most recent transaction, t - t 最新 is the time interval, and λ is the time decay factor, which can be adjusted according to industry characteristics.

[0031] Optionally, S2 includes:

[0032] S2.1: Model Architecture:

[0033] Input layer: node feature vector: structured data + unstructured data embedding;

[0034] Edge feature matrix: contains relationship type, temporal weight, and historical risk labels (such as whether the associated company has ever defaulted);

[0035] S2.2: Dynamic Graph Attention Layer:

[0036] A multi-head attention mechanism is used to assign dynamic weights to nodes in different time slices;

[0037] Attention coefficient calculation:

[0038]

[0039] Among them, a T is the transpose of the learnable parameters, used to calculate the attention score, W is the weight matrix: for node feature h i Perform linear transformation to extract high-order features, h i is the node embedding, e ij is the edge feature, k is the attention head index, indicating the kth head in the multi-head attention mechanism, N i is the neighbor set of node i, the set of nodes directly connected to i in the graph structure, LeakyReLU is the activation function, h m Feature representation of neighbor node m, e im is the edge feature between node i and neighbor m;

[0040] S2.3: Spatiotemporal fusion module:

[0041] The nodes output by DyGAT are embedded into the input LSTM to capture the time series dependency. The formula is:

[0042]

[0043] Output layer: Combines the node risk probability and the community risk index to generate the final risk score.

[0044] Optionally, LeakyReLU is the activation function, defined as:

[0045] Usually γ=0.2

[0046] Where x is the input value;

[0047] Introduces nonlinearity, allowing negative values ​​to pass through with a smaller slope.

[0048] Optionally, S4 includes:

[0049] S4.1: Risk score generation:

[0050] Individual risk: Output probability value P through supervised learning 违约 ∈[0,1];

[0051] Group risk: Use the Louvain algorithm to detect high-risk communities and calculate the community risk index R c =∑ i∈c Pi·PageRank(i), PageRank in community risk index R c It plays the role of a dynamic risk amplifier, and measures the network influence of the node to convert the static risk probability P i Transformed into dynamic transmission risk;

[0052] Adjust the systemic risk threshold based on macroeconomic indicators.

[0053] Optionally, S4 further includes:

[0054] S4.2: Enhanced interpretability:

[0055] Risk transmission path visualization: Generate heat maps based on subgraph attention weights to highlight core risk nodes.

[0056] Optionally, interpretability enhancements also include natural language reporting: generating reports based on templates.

[0057] Compared with the prior art, this application has at least the following beneficial effects:

[0058] 1. Comprehensive risk coverage: Dynamic heterogeneous graph construction: Integrate multi-source heterogeneous data (corporate equity, supply chain transactions, social relationships, geographic overlap), give equal weight to explicit and implicit relationships, and build a dynamic graph containing corporate, individual, and geographic nodes. Dynamically adjust edge weights through a time decay function to quantify the timeliness of association strength.

[0059] Spatiotemporal graph neural network: Combining the dynamic graph attention mechanism (DyGAT) with LSTM, it simultaneously models spatial correlation and temporal dependency, improving the ability to capture risk transmission paths.

[0060] 2. Privacy compliance assurance: Federated learning framework: supports cross-institutional (bank, e-commerce) data collaboration, local training of subgraph models, aggregating only encrypted gradient parameters, and preventing the leakage of original data;

[0061] Differential privacy technology: Gaussian noise is added during gradient aggregation to meet regulatory requirements such as GDPR, with a measured privacy compliance score of 9.5 / 10.

[0062] 3. Dynamic early warning and high-precision prediction: Supports risk warning 3-6 months in advance (traditional models only 1-2 months), AUC improved to 0.92 (traditional GCN is 0.81), and the non-performing loan ratio has been measured to decrease by 33% (from 1.8% to 1.2%).

[0063] High-risk communities are detected through the Louvain algorithm, systemic risks are quantified in combination with PageRank, and thresholds (such as industry prosperity index) are dynamically adjusted to enhance macro-risk prevention and control capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A flowchart of a graph neural network-based correlation credit default warning method provided in accordance with one embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0066] The present application provides a graph neural network-based correlation credit default warning method, comprising the following steps:

[0067] S1: Multi-source cross-institutional data fusion and dynamic graph construction

[0068] S1.1: Data collection and preprocessing:

[0069] Structured data: business information, financial statements, personal credit records, transaction flow (frequency, amount, timestamp).

[0070] Unstructured data: Text data: Judicial decisions (using NLP (Natural Language Processing) to extract the subject, amount, and type of litigation).

[0071] Social data: Contact list relationship chains (mapped to closeness weights through graph embedding).

[0072] Time series data processing: Perform sliding window normalization on node attributes (such as corporate quarterly revenue) and edge weights (such as transaction frequency) to eliminate seasonal fluctuations.

[0073] S1.2: Dynamic heterogeneous graph construction:

[0074] S1.2.1: Determine the node type:

[0075] Enterprise node: Attributes include industry classification, debt-to-asset ratio, and equity structure.

[0076] Personal nodes: Attributes include career stability, long-borrowing index, and social network centrality.

[0077] Geographic nodes: Attributes include regional GDP growth rate and judicial enforcement case density.

[0078] S1.2.2: Edge relationship definition:

[0079] Explicit relationships: equity holding (weight = shareholding ratio), supply chain transactions (weight = transaction volume ratio).

[0080] Implicit relationships: social closeness (based on call frequency in the address book) and address overlap (groups of companies with the same registered place).

[0081] S1.2.3: Dynamic edge weight update:

[0082] The time decay function is used to dynamically adjust the edge weights. The formula is:

[0083]

[0084] Among them, w ij (t) represents the association strength between node i and node j at time t, β is the transaction frequency weight coefficient, which balances the impact of transaction frequency and time decay on edge weight and reflects the importance attached to recent transactions. Transaction frequency (t) is the transaction frequency indicator, which refers to the normalized value of the number or amount of transactions between nodes i and j within the time window t. t is the current time, t 最新 is the time of the most recent transaction, t - t 最新 is the time interval, and λ is the time decay factor, which can be adjusted according to industry characteristics (e.g., λ = 0.2λ = 0.2 for the fast-moving consumer goods industry, and λ = 0.1λ = 0.1 for the heavy industry);

[0085] S2: Spatiotemporal Graph Neural Network Model Design

[0086] S2.1: Model Architecture:

[0087] Input layer: node feature vector: structured data (numerical standardization) + unstructured data embedding (text features extracted through BERT);

[0088] Edge feature matrix: contains relationship type, time series weight, and historical risk label (such as whether the related company has ever defaulted).

[0089] S2.2: Dynamic Graph Attention Layer (DyGAT):

[0090] A multi-head attention mechanism (4 heads) is used to assign dynamic weights to nodes in different time slices.

[0091] Attention coefficient calculation:

[0092]

[0093] Among them, a T is the transpose of the learnable parameters, used to calculate the attention score, W is the weight matrix: for node feature h i Perform linear transformation to extract high-order features, h i is the node embedding, e ij is the edge feature, k is the attention head index, indicating the kth head in the multi-head attention mechanism, N i is the neighbor set of node i, the set of nodes directly connected to i in the graph structure, LeakyReLU is the activation function, h m Feature representation of neighbor node m, e im is the edge feature between node i and neighbor m;

[0094] LeakyReLU is an activation function: it introduces nonlinearity and allows negative values ​​to pass with a smaller slope. It is defined as:

[0095] Usually γ=0.2

[0096] Where x is the input value;

[0097] S2.3: Spatiotemporal fusion module:

[0098] The nodes output by DyGAT are embedded into the input LSTM (Long Short-Term Memory Network) to capture the time series dependency. The formula is:

[0099]

[0100] Output layer: Combines the node risk probability and the community risk index to generate the final risk score.

[0101] S3: Deploy a federated learning framework to support cross-institutional data joint modeling (such as collaboration between banks and e-commerce platforms):

[0102] Local training: Each participating institution (such as banks and e-commerce platforms) trains the subgraph model locally and retains sensitive data.

[0103] Parameter aggregation: The central server aggregates gradient information (not raw data) and adds Gaussian noise (∈=1.0,δ=10-5∈=1.0,δ=10-5) using differential privacy.

[0104] Model update: The spatiotemporal graph neural network model updates parameters through weighted averaging and supports heterogeneous graph structures (for example, institution A provides corporate data and institution B provides personal consumption data).

[0105] S4: Risk Warning and Explainable Output

[0106] S4.1: Risk score generation:

[0107] Individual risk: Output probability value P through supervised learning (labeled as historical default records) 违约 ∈[0,1].

[0108] Group risk: Use the Louvain algorithm to detect high-risk communities and calculate the community risk index R c =∑ i∈c Pi·PageRank(i), PageRank in community risk index R c It plays the role of a dynamic risk amplifier, and measures the network influence of the node to convert the static risk probability P i It is transformed into a dynamic transmission risk. Combined with the Louvain algorithm, it achieves a leap from "local risk identification" to "global risk quantification," providing mathematical support for group risk management.

[0109] Adjust the systemic risk threshold in combination with macroeconomic indicators (such as industry prosperity index).

[0110] S4.2: Enhanced interpretability:

[0111] Risk transmission path visualization: Generate heat maps based on subgraph attention weights to highlight core risk nodes (such as core enterprises in the guarantee circle);

[0112] Example: After a company defaults, the risk scores of its upstream and downstream suppliers show a "ripple diffusion" effect in the model.

[0113] Natural language reporting: Generates reports based on templates. For example, "Company A holds a 30% stake in Company B (weight = 0.3), and Company B has been involved in three lawsuits (type: contract disputes) in the past six months, causing Company A's risk score to increase by 15%. It is recommended that the supply chain finance contingency plan be triggered: shorten the payment period to 30 days."

[0114] In a specific embodiment, the above method is applied to a supply chain finance scenario:

[0115] Data input:

[0116] Enterprise data: 1,000 manufacturing companies, including three years of transaction flows, equity structure, and litigation records.

[0117] External data: industry prosperity index (from the National Bureau of Statistics), regional economic policy text (keyword extraction through NLP).

[0118] Graph construction:

[0119] Dynamic edge weight update: set β = 0.7, λ = 0.15β = 0.7, λ = 0.15, and update once every quarter.

[0120] Model training:

[0121] Hyperparameters: learning rate = 0.001, LSTM hidden layers = 256, federated learning epochs = 50.

[0122] Privacy protection: gradient clipping threshold = 1.0, Gaussian noise σ = 0.3σ = 0.3.

[0123] Verification results:

[0124] Prediction accuracy: AUC = 0.92 (traditional GCN AUC = 0.81).

[0125] Accuracy of risk transmission path: 93% (manual verification of the diffusion path of actual default cases).

[0126] In a specific embodiment, the above method is applied to a personal credit scenario:

[0127] Data input:

[0128] Personal nodes: 100,000 users, including address books, co-residents, and consumption installment records.

[0129] Side relationship: Call frequency > 5 times / week is defined as "strong social relationship".

[0130] Risk Warning:

[0131] Identifying high-risk communities: In a P2P lending group, after 20% of members defaulted, the model issued a three-month advance warning of increased risk among the remaining members.

[0132] Compliance Verification:

[0133] In federated learning, user data is retained locally, and the central server only receives encrypted gradient parameters, meeting GDPR requirements.

[0134] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A correlation-based credit default warning method based on graph neural network, characterized by: The method comprises the following steps: S1: Fusion of multi-source and cross-institutional data and construction of dynamic graphs; S2: Combine cross-institutional data and dynamic graphs to design a spatiotemporal graph neural network model to output the final risk score; S3: Deploy a federated learning framework to support joint modeling of cross-institutional data: Local training: Each participating institution trains the subgraph model locally, retaining sensitive data; Parameter aggregation: The central server aggregates gradient information and adds Gaussian noise using differential privacy; Model update: The spatiotemporal graph neural network model updates parameters through weighted averaging; S4: Risk warning and explainable output, combining risk scoring and joint modeling to output warning content.

2. The graph neural network-based correlation credit default warning method according to claim 1, characterized in that: S1 includes: S1.1: Data collection and preprocessing: Collect structured data, unstructured data, social data, and perform time series data processing; S1.2: Dynamic heterogeneous graph construction.

3. The method for credit default early warning based on graph neural network according to claim 2, characterized in that: S1.2 includes: S1.2.1: Determine the node type: Enterprise node: Attributes include industry classification, debt-to-asset ratio, and equity structure; Personal nodes: Attributes include career stability, long-term lending index, and social network centrality; Geographic nodes: Attributes include regional GDP growth rate and judicial enforcement case density; S1.2.2: Edge relationship definition: Explicit relationships: equity holdings, supply chain transactions; Implicit relationships: social closeness, address overlap; S1.2.3: Dynamic edge weight update: The time decay function is used to dynamically adjust the edge weights. The formula is: Among them, w ij (t) represents the correlation strength between node i and node j at time t, β is the transaction frequency weight coefficient, transaction frequency (t) is the transaction frequency index, which refers to the normalized value of the number or amount of transactions between nodes i and j within the time window t, t is the current time, t 最新 is the time of the last transaction, t - t 最新 is the time interval, and λ is the time decay factor, which can be adjusted according to industry characteristics.

4. The method for early warning of credit default based on graph neural network according to claim 1, characterized in that S2 include: S2.1: Model Architecture: Input layer: node feature vector: structured data + unstructured data embedding; Edge feature matrix: contains relationship type, temporal weight, and historical risk labels (such as whether the associated company has ever defaulted); S2.2: Dynamic Graph Attention Layer: A multi-head attention mechanism is used to assign dynamic weights to nodes in different time slices; Attention coefficient calculation: Among them, a T is the transpose of the learnable parameters, used to calculate the attention score, W is the weight matrix: for node feature h i Perform linear transformation to extract high-order features, h i is the node embedding, e ij is the edge feature, k is the attention head index, indicating the kth head in the multi-head attention mechanism, N i is the neighbor set of node i, the set of nodes directly connected to i in the graph structure, LeakyReLU is the activation function, h m Feature representation of neighbor node m, e im is the edge feature between node i and neighbor m; S2.3: Spatiotemporal fusion module: The nodes output by DyGAT are embedded into the input LSTM to capture the time series dependency. The formula is: Output layer: Combines the node risk probability and the community risk index to generate the final risk score.

5. The graph neural network-based correlation credit default warning method according to claim 4, characterized in that: Definition of the LeakyReLU activation function: Usually γ=0.2 Where x is the input value; Introduces nonlinearity, allowing negative values ​​to pass through with a smaller slope.

6. The method for early warning of credit default based on graph neural network according to claim 4, characterized in that S4 include: S4.1: Risk score generation: Individual risk: Output probability value P through supervised learning 违约 ∈[0,1]; Group risk: Use the Louvain algorithm to detect high-risk communities and calculate the community risk index R c =∑ i∈c Pi·PageRank(i), PageRank in community risk index R c It plays the role of a dynamic risk amplifier, and measures the network influence of the node to convert the static risk probability P i Transformed into dynamic transmission risk; Adjust the systemic risk threshold based on macroeconomic indicators.

7. The method for credit default early warning based on graph neural network according to claim 6, characterized in that: The S4 also includes: S4.2: Enhanced interpretability: Risk transmission path visualization: Generate heat maps based on subgraph attention weights to highlight core risk nodes.

8. The graph neural network-based correlation credit default warning method according to claim 7, characterized in that: Interpretability enhancements also include natural language reporting: generating reports based on templates.

Images