Financial Credit Risk Assessment and Dynamic Optimization Method Integrating Graph Neural Networks

Abstract

This invention discloses a method for financial credit risk assessment and dynamic optimization integrating graph neural networks, belonging to the field of financial management data processing technology. It calculates the risk correlation between enterprise nodes and funding nodes. If the risk correlation does not exceed a preset first threshold, the corresponding node is marked as a risk node, and multiple rounds of risk propagation simulation are performed based on the difference between the risk correlation and the preset first threshold. During the simulation, the risk correlation of the risk nodes is recalculated, and a set of risk-correlated nodes is obtained and divided into multiple data domains. Time-series graph analysis is performed on each data domain to extract the risk nodes in that domain, and differentiated risk optimization strategies are generated accordingly. Through graph neural networks and dynamic optimization mechanisms, the accuracy and real-time performance of financial credit risk assessment are significantly improved, enabling effective identification of potential risks and the development of targeted countermeasures.

Description

Technical Field

[0001] This invention relates to the field of financial management data processing technology, and more specifically, this application relates to a method for financial credit risk assessment and dynamic optimization that integrates graph neural networks. Background Technology

[0002] In the field of financial credit risk assessment, traditional methods typically rely on static corporate financial indicators and isolated transaction data analysis, making it difficult to comprehensively capture the dynamic propagation characteristics of risk within complex network structures. These methods often fail to effectively integrate corporate relationships and capital flow paths, resulting in a single-dimensional and unreal-time risk assessment that easily overlooks potential risk associations and evolutionary trends. Specifically, existing technologies are mostly based on simple statistical models or single machine learning algorithms, neglecting the interactive influence of inter-corporate transaction networks and capital flow networks, and failing to accurately quantify the strength of associations between risk nodes. Furthermore, due to fragmented data sources and rigid processing procedures, traditional methods struggle to achieve multi-round simulation and propagation analysis of risk, causing risk assessment results to lag behind actual market changes and failing to provide effective early warnings in the early stages of risk. In addition, existing technologies lack in-depth analysis of the temporal evolution of risk and cannot dynamically adjust assessment strategies based on historical data, resulting in insufficiently targeted risk management measures that are ill-suited to cope with complex and ever-changing financial environments.

[0003] To address the aforementioned issues, there is an urgent need in this field for a financial credit risk assessment method capable of integrating multi-source data, simulating risk propagation, and dynamically optimizing strategies to improve the comprehensiveness and timeliness of the assessment. The main problem with existing technologies is their inability to achieve dynamic, multi-dimensional risk assessment and optimization within complex corporate entities and cash flow networks, resulting in incomplete risk assessments and slow responses. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method for financial credit risk assessment and dynamic optimization that integrates graph neural networks. This technical solution solves the problems mentioned in the background section.

[0005] In a first aspect, embodiments of this application provide a method for financial credit risk assessment and dynamic optimization fused with graph neural networks, including the following steps: acquiring an enterprise entity graph constructed based on enterprise entity data and a capital flow graph constructed based on capital flow data; calculating and obtaining the risk correlation degree between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph; if the risk correlation degree does not exceed a preset first threshold, acquiring the corresponding enterprise node and capital node and recording them as risk nodes, recording the difference between the risk correlation degree and the preset first threshold as the first difference, and performing multiple rounds of risk propagation simulation accordingly; in the risk propagation simulation, recalculating the risk correlation degree of the risk nodes; if the recalculated risk correlation degree does not exceed the preset first threshold, summarizing the corresponding risk nodes to obtain a risk-related node set; dividing the risk-related node set into at least one data domain, performing time-series graph analysis on each data domain to obtain a risk evolution coefficient; if the risk evolution coefficient exceeds a preset second threshold, extracting the risk nodes corresponding to the corresponding data domain; generating differentiated risk optimization strategies based on the extracted risk nodes.

[0006] Secondly, embodiments of this application provide a financial credit risk assessment and dynamic optimization system integrating graph neural networks, including: a data acquisition module: used to acquire an enterprise entity graph constructed based on enterprise entity data and a capital flow graph constructed based on capital flow data, and calculate and obtain the risk correlation degree between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph; a data processing first module: used to acquire the corresponding enterprise node and capital node and record them as risk nodes if the risk correlation degree does not exceed a preset first threshold, record the difference between the risk correlation degree and the preset first threshold as the first difference, and perform multiple rounds of risk propagation simulation accordingly. The data processing module 2 is used to recalculate the risk correlation of risk nodes in the risk propagation simulation. If the recalculated risk correlation does not exceed a preset first threshold, the corresponding risk nodes are aggregated to obtain a risk-related node set. The data processing module 3 is used to divide the risk-related node set into at least one data domain, perform time-series graph analysis on each data domain, and obtain the risk evolution coefficient. The data processing module 4 is used to extract the risk nodes corresponding to the data domain if the risk evolution coefficient exceeds a preset second threshold. The data output module is used to generate differentiated risk optimization strategies based on the extracted risk nodes.

[0007] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0008] 1. By integrating enterprise entity graphs and capital flow graphs, and calculating the risk correlation between enterprise nodes and capital nodes, multi-dimensional features are integrated. The three-layer graph attention network layer and fully connected layer of the graph neural network model are used for feature aggregation, thereby capturing the complex relationships between risk factors more comprehensively and significantly improving the accuracy of risk assessment.

[0009] 2. By conducting multi-round risk propagation simulations based on the difference between the risk correlation degree and a preset first threshold, and by employing random walk and probabilistic selection strategies based on attention weights, this invention can simulate the diffusion path of risk in a two-layer network structure, promptly identify and summarize the set of risk-related nodes, thereby providing early warning in the early stages of risk and preventing risk spread and accumulation.

[0010] 3. By dividing the set of risk-related nodes into data domains and performing time-series graph analysis, the risk evolution coefficient is obtained. When the risk evolution coefficient exceeds a preset second threshold, the corresponding risk node is extracted to generate a differentiated strategy. This invention can dynamically track the risk evolution trend and adjust and optimize the strategy according to the actual risk event through a graph neural network model, thereby improving the pertinence and overall efficiency of risk management. Attached Figure Description

[0011] Figure 1 A schematic diagram of the structure of the financial credit risk assessment and dynamic optimization method fused graph neural network provided in the embodiments of this application;

[0012] Figure 2 This is a logical diagram illustrating the risk propagation simulation provided in an embodiment of this application.

[0013] Figure 3 This is a schematic diagram of the structure of the financial credit risk assessment and dynamic optimization system based on the fused graph neural network provided in this application embodiment. Detailed Implementation

[0014] This application's embodiments address the technical problem in the prior art where, in complex heterogeneous financial network environments, the accuracy of dynamically quantifying credit risk transmission and assessing implicit risks is insufficient, through a financial credit risk assessment and dynamic optimization method that integrates graph neural networks.

[0015] In existing technologies, financial credit risk assessment has long relied on static financial indicators and isolated transaction records for analysis, making it difficult to capture the transmission effects of risk within complex business networks. A commercial bank, when assessing creditworthiness of a core enterprise in its supply chain, found that traditional models could not effectively identify the implicit risk transmission from related guarantor companies. When the core enterprise experienced a liquidity crisis, the lack of dynamic tracking of fund flows led to delayed risk warnings, ultimately triggering a chain reaction of defaults. The essence of this assessment blind spot lies in the data silo effect and the lack of dynamic analysis capabilities.

[0016] To address these issues, the R&D team discovered that risk transmission exhibits network-like characteristics, necessitating the establishment of a cross-network risk quantification model. By constructing a two-layer graph of enterprise entities and capital flows, they attempted to capture cross-network risk correlations. However, initial experiments revealed that a fixed analysis scope led to the omission of key nodes, prompting a shift in research focus to dynamic adjustment mechanisms. To resolve the issue of unclear propagation paths, a random walk algorithm was introduced to simulate risk diffusion. Ultimately, a technical approach combining dynamic threshold adjustment and multi-dimensional clustering was developed, enabling precise tracking of risk evolution.

[0017] This application addresses the technical challenge of cross-network risk quantification and tracking, enabling dynamic optimization of the risk assessment scope. The method identifies key nodes along the risk transmission path, reducing the risk of misjudgment caused by a fixed analysis scope. Through a dual mechanism of joint clustering and time-series analysis, it significantly improves the timeliness of risk warnings and the targeting of response strategies, providing financial institutions with a dynamically adaptive risk assessment tool.

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] like Figure 1 The diagram shown is a structural schematic of the financial credit risk assessment and dynamic optimization method based on a fused graph neural network provided in this application embodiment. The method includes the following steps: obtaining an enterprise entity graph constructed from enterprise entity data and a capital flow graph constructed from capital flow data; calculating and obtaining the risk correlation degree between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph; if the risk correlation degree does not exceed a preset first threshold, obtaining the corresponding enterprise node and capital node and recording them as risk nodes; recording the difference between the risk correlation degree and the preset first threshold as the first difference and performing multiple rounds of risk propagation simulation accordingly; in the risk propagation simulation, recalculating the risk correlation degree of the risk nodes; if the recalculated risk correlation degree does not exceed the preset first threshold, summarizing the corresponding risk nodes to obtain a risk-related node set; dividing the risk-related node set into at least one data domain, performing time-series graph analysis on each data domain to obtain a risk evolution coefficient; if the risk evolution coefficient exceeds a preset second threshold, extracting the risk nodes corresponding to the corresponding data domain; and generating differentiated risk optimization strategies based on the extracted risk nodes.

[0020] Risk correlation refers to an indicator that quantifies the strength of risk transmission between corporate entities and funding nodes. Specifically, it can be achieved by aggregating multi-dimensional features through graph neural networks. This indicator can effectively identify key nodes in cross-network risk transmission.

[0021] Time series graph analysis refers to the technical means of dynamically tracking risk correlation sequences, such as calculating risk evolution trend coefficients through time slices to predict high-risk evolution areas.

[0022] The preset first threshold is obtained through training on historical risk data. Specifically, this includes: collecting risk correlation data of historical enterprise nodes and funding nodes and associating them with actual risk events; using ROC curve analysis to determine the optimal threshold, which maximizes the true positive rate and minimizes the false positive rate; and storing the threshold as the preset first threshold for subsequent risk correlation comparisons.

[0023] First, a graph neural network is used to fuse enterprise entity and capital flow data to generate cross-network risk correlation. When a high-risk correlation node is detected, the analysis scope is dynamically expanded based on the risk intensity difference, for example, increasing the preset number of nodes from 100 to 150. Through multiple rounds of risk propagation simulation, the distribution of risk nodes under different threshold conditions is recorded. Subsequently, joint clustering is performed by combining node similarity and risk difference, grouping nodes with similar risk characteristics into the same data domain. Time series analysis is conducted for each domain to identify key areas where risks are rapidly evolving. Finally, differentiated risk management solutions are generated for nodes in high-risk evolution domains, such as implementing liquidity monitoring for rapidly deteriorating supply chain nodes.

[0024] Furthermore, the risk correlation between enterprise nodes in the enterprise entity graph and fund nodes in the fund flow graph is calculated and obtained. Specifically, this includes: obtaining an enterprise entity graph constructed from enterprise entity data, which contains enterprise nodes and enterprise relationship edges. Enterprise nodes are enterprise feature vectors of a preset dimension, and enterprise relationship edges are transaction feature vectors containing transaction frequency and average transaction amount; and a fund flow graph constructed from fund flow data, which contains fund nodes and fund relationship edges. Fund nodes are institutional feature vectors of institution type and capital adequacy ratio, and fund relationship edges are... The system includes channel feature vectors containing channel fees, average clearing time, and historical default counts. By mapping enterprise nodes to fund nodes, cross-layer connection edge feature vectors are constructed, forming a two-layer network structure between the enterprise entity graph and the fund flow graph. The cross-layer connection edge feature vectors consist of account activity and fund aggregation ratio. The enterprise feature vectors, transaction feature vectors, institutional feature vectors, channel feature vectors, and cross-layer connection edge feature vectors are input into the graph neural network model. The graph neural network model aggregates the features of nodes and edges through a three-layer graph attention network and outputs the risk correlation through a fully connected layer.

[0025] In this embodiment, the enterprise feature vector refers to a vectorized representation constructed through multi-dimensional enterprise attributes. Specifically, it can be formed by standardizing indicators such as registered capital, debt-to-asset ratio, and tax level and then concatenating them to quantify the basic creditworthiness of the enterprise.

[0026] The funding node feature vector refers to the set of operational characteristics of financial institutions. Specifically, it can be vectorized by using institution type codes and capital adequacy ratio values ​​to characterize the stability of funding providers.

[0027] Cross-layer connection edge feature vectors refer to the interaction features between enterprise accounts and fund accounts. Specifically, the product of the average monthly number of transactions per account and the proportion of fund aggregation can be used as the calculation dimension to reflect the activity level and concentration risk of fund flows.

[0028] Graph attention network layer refers to a feature aggregation layer with an attention mechanism. Specifically, a multi-head attention mechanism can be used to dynamically weight the features of adjacent nodes to capture the differences in the impact of different connection edges on risk propagation.

[0029] Nodes in the enterprise entity graph are encoded using enterprise feature vectors of a predefined dimension, such as converting registered capital to a logarithmic scale and combining it with industry classification for encoding. Fund nodes in the capital flow graph are encoded using one-hot encoding based on institution type and concatenated with capital adequacy ratio values ​​to form institution feature vectors. When an enterprise node establishes a mapping relationship with a capital node through a public account, the feature vector of the cross-layer connection edge is generated by calculating account activity indicators and capital aggregation ratios. All node and edge feature vectors are input into a model containing a three-layer graph attention network. Each layer performs cross-graph aggregation of node features, and finally, a fully connected layer converts the aggregated features into risk correlation values. During feature aggregation, the graph attention mechanism dynamically adjusts weights based on the feature similarity of adjacent nodes, giving higher weights to high-risk correlation paths during calculation.

[0030] By employing a dynamically adjusted attention weighting mechanism, the differences in the contribution of different transaction channels to risk propagation can be accurately captured, overcoming the evaluation bias problem caused by fixed weight allocation in traditional methods. The introduction of cross-layer connecting edges enables the interaction between previously isolated enterprise characteristics and capital characteristics, providing a reliable data foundation for subsequent risk propagation simulations.

[0031] Furthermore, multiple rounds of risk propagation simulation are conducted, specifically including: in each round of risk propagation simulation, starting from the risk node, adjusting the preset node number range according to the first difference, and performing a random walk on the two-layer network structure according to the adjusted preset node number range; the preset node number range is used to determine the maximum number of neighboring nodes starting from the risk node; when selecting the next node, a probabilistic selection strategy based on attention weights is adopted, i.e., the next node... From the current node Neighbor set The probability of being selected Calculated using the following formula: ,in, and They are and eigenvectors, It is a shared weight matrix. It is an attention vector. This represents a vector concatenation operation. yes The new eigenvectors obtained after linear transformation of the shared weight matrix yes The new eigenvectors obtained after linear transformation of the shared weight matrix This represents the transpose of the attention vector. This represents the leakage linear rectifier function. Represents an exponential function. Indicates the current node The number of all neighboring nodes.

[0032] In this embodiment, Figure 2 This is a logical diagram illustrating the risk propagation simulation provided in an embodiment of this application.

[0033] Random walks are a method of traversing risk propagation paths on a graph structure. Specifically, they can be implemented using the transition probabilities of Markov chains. Their role is to discover potential risk transmission patterns through multiple path explorations.

[0034] The probabilistic selection strategy of attention weight refers to dynamically adjusting the path selection probability based on the correlation between features between nodes. Specifically, it can be implemented through the graph attention mechanism, which prioritizes the transmission path that is highly correlated with the current risk features.

[0035] When initiating risk propagation simulation on a two-layer network structure, a random walk process is performed starting from the node marked as the first risk association. Each time a next-hop node is selected, the transition probability distribution between nodes is dynamically adjusted by calculating the attention weights between the current node and its neighbors. This weight calculation process transforms the node feature vector through a shared weight matrix, concatenates it with the attention vector, and processes it using a nonlinear activation function to ultimately form a normalized probability distribution. After each round of simulation, the frequency of visits to each node is counted. When the visit frequency of a specific path exceeds a preset threshold, it is marked as a critical propagation path. This process iterates through multiple rounds to gradually construct the critical path topology for risk transmission.

[0036] Dynamic probability adjustment based on attention mechanism makes the simulation process more focused on high-risk transmission directions, and the access frequency statistics method can objectively reflect the importance level of the path. While ensuring computational efficiency, it significantly improves the accuracy of identifying critical paths of risk transmission, and provides a reliable basis for the formulation of subsequent risk management strategies.

[0037] Furthermore, the preset node number range is corrected based on the first difference, specifically including: extracting a preset scale expansion coefficient from the fused graph neural network database; and correcting the preset node number range according to the correction formula, the specific correction formula being as follows: ,in, This is the corrected range of the preset number of nodes. The preset range of the number of nodes, The first difference, The preset scaling factor, This is the floor function.

[0038] In this embodiment, the preset node number range is dynamically set based on the network size and computing resources. Specifically, it includes: calculating the average degree of nodes in the two-layer network structure, setting an initial node number range according to the simulation accuracy requirements; determining a value that balances computing efficiency and coverage through experiments, and storing it as the preset node number range.

[0039] The scale expansion coefficient is a parameter used to adjust the range of node quantity adjustments. This parameter determines the degree of influence of the risk correlation difference on the range of node quantity. The preset scale expansion coefficient is determined through risk propagation sensitivity analysis, specifically including: testing the impact of the scale expansion coefficient on risk path coverage under different risk scenarios; selecting a coefficient value that makes the risk path coverage rate reach more than 95%, and storing it as the preset scale expansion coefficient.

[0040] The first difference refers to the difference between the risk correlation degree and the preset first threshold, which can be obtained by arithmetic subtraction. This difference reflects the degree to which the risk correlation strength exceeds the threshold.

[0041] The correction formula is a mathematical expression used to dynamically adjust the range of the number of nodes. Specifically, it can be implemented through a linear combination. This formula combines the risk correlation difference with the scale expansion coefficient to achieve elastic expansion of the range of the number of nodes.

[0042] When the risk correlation exceeds a preset threshold, the system calculates a first difference and, combined with a preset expansion coefficient, linearly expands the initial node range. For example, when the initial node range N is set to 100, the first difference δ is 0.5, and the expansion coefficient k is 20, the corrected node range N′ will expand to 110. This dynamic adjustment mechanism allows the node sampling range of the risk propagation simulation to adaptively expand according to the actual risk spillover level, avoiding insufficient path coverage due to a fixed range and preventing the waste of computational resources caused by over-expansion. During implementation, the expansion coefficient can be preset according to different business scenarios; for example, a larger coefficient can be used during periods of tight liquidity to enhance the sensitivity of risk monitoring.

[0043] By dynamically adjusting the range of node numbers, the system ensures complete path coverage in high-risk scenarios while avoiding wasted computational resources in low-risk scenarios. For example, when a moderate risk overflow is detected in a node of an enterprise, the system can automatically expand the node sampling range by 20% to ensure that potential second-order risk propagation paths are captured, while keeping computational complexity within a reasonable range. This flexible adjustment mechanism significantly improves the balance between the accuracy and computational efficiency of risk propagation simulation.

[0044] Furthermore, the risk-related node set is divided into at least one data domain, specifically including: calculating the multidimensional similarity between risk nodes in the risk-related node set; using the difference between a preset first threshold and the recalculated risk correlation corresponding to the risk node as a second difference; using the second difference as the modulus in an independent dimension to obtain the first vector of the risk node; concatenating the corresponding multidimensional similarity as the modulus in another orthogonal independent dimension to obtain the second vector of the risk node; combining the first and second vectors to obtain the joint feature vector of the risk nodes; and clustering the joint feature vector using a predefined density clustering algorithm, wherein the neighborhood radius parameter of the predefined density clustering algorithm is adjusted according to the standard deviation of the joint feature vector, and the specific adjustment formula is as follows: ,in, The preset scale constant, The standard deviation of the joint eigenvectors, This represents the neighborhood radius parameter; it divides nodes within the same density connectivity domain into the same data domain.

[0045] In this embodiment, the second difference refers to the numerical difference between the risk correlation degree of each node in the risk-related node set and the preset first threshold. Specifically, it can be implemented by arithmetic subtraction. This difference reflects the degree to which a node exceeds the baseline risk level during the risk propagation process.

[0046] A multidimensional similarity vector is a comprehensive similarity index vector formed by weighted summation of attribute similarity and structural similarity. Specifically, it can be implemented by weighted calculation using cosine similarity and Jaccard similarity coefficient, and is used to comprehensively characterize the degree of similarity between nodes in feature attributes and network topology.

[0047] Predefined density clustering algorithms are machine learning algorithms that perform clustering analysis based on the principle of density reachability. Specifically, they can be implemented using the DBSCAN algorithm or its improved versions. They are suitable for discovering clusters of arbitrary shapes and effectively identifying noise points.

[0048] The neighborhood radius parameter refers to the distance threshold used in density clustering algorithms to determine whether data points belong to the same neighborhood. Specifically, it can be implemented using a standard deviation adjustment mechanism, which dynamically adjusts the search range according to the data distribution characteristics to adapt to data regions of different densities.

[0049] After the risk propagation simulation, the second difference of each node in the risk-associated node set is extracted as an independent dimension. This dimension is then concatenated with a multidimensional similarity vector obtained by weighted summation of attribute similarity and structural similarity to form a high-dimensional joint feature vector containing both risk intensity and similarity features. When using a predefined density clustering algorithm to cluster the joint feature vector, the neighborhood radius parameter is adaptively adjusted based on the standard deviation of the current data distribution. For example, the neighborhood search range is automatically expanded when the data distribution is relatively dispersed, and narrowed when the data distribution is dense. The minimum sample size ratio is dynamically determined based on the proportion of risk nodes in the global network. For example, when risk nodes account for 5% of the total number of nodes, this ratio can be set to 0.05. After density clustering, nodes located in the same density connectivity domain are grouped into the same data domain, and these domains exhibit high consistency in risk propagation characteristics and network topology.

[0050] By jointly analyzing risk intensity indicators with multidimensional similarity features, we can accurately identify clusters of nodes with similar risk propagation patterns and similar levels of risk spillover, providing a reliable data foundation for subsequent time-series analysis and the formulation of differentiated risk optimization strategies. The dynamic parameter adjustment mechanism ensures stable clustering performance across financial network scenarios of varying scales and complexities, significantly improving the environmental adaptability of the risk assessment model.

[0051] Furthermore, the multidimensional similarity between risk nodes in the risk-related node set is calculated, specifically including: obtaining attribute similarity by calculating the cosine similarity between the enterprise feature vectors or institutional feature vectors of two nodes; obtaining structural similarity by calculating the Jaccard similarity coefficient between the adjacency vectors of two nodes in the two-layer network structure; the adjacency vector is a Boolean vector that identifies whether a node has a connection with other nodes, taking all nodes in the two-layer network structure as the judgment object; the first weight and the second weight are extracted from the fused graph neural network database; and the multidimensional similarity is obtained by weighted summation of attribute similarity and structural similarity.

[0052] In this embodiment, the first weight and the second weight are determined through grid search and cross-validation, specifically including: testing the impact of different weight combinations on the clustering effect on historical data; selecting the weight combination that maximizes the clustering silhouette coefficient, and setting the attribute similarity weight as the first weight and the structural similarity weight as the second weight.

[0053] Attribute similarity refers to the degree of matching between nodes in terms of business attributes. Specifically, it can be achieved by using the cosine similarity algorithm to quantitatively compare the business characteristics of enterprises or the capital characteristics of financial institutions, in order to identify risk-related nodes with similar business attributes.

[0054] Structural similarity refers to the similarity of the positions of nodes in the network topology. Specifically, it can be achieved by comparing the connection patterns of nodes using the Jaccard similarity coefficient, which is used to discover risky nodes with similar propagation paths.

[0055] Adjacency vectors are Boolean representations of node connection relationships. Specifically, they can be implemented using binary vectors to record the connection states between a node and all nodes in the graph, and are used to quantify the network structure characteristics of nodes.

[0056] After the risk-related node set is formed, the attribute similarity and structural similarity between nodes are calculated first. Attribute similarity is processed using the cosine similarity algorithm to analyze enterprise or institutional feature vectors, such as measuring the similarity of features like registered capital and revenue scale. Structural similarity is calculated using the Jaccard coefficient to determine the overlap ratio of adjacency vectors, such as comparing the degree of overlap of connected objects between two nodes in a capital flow network. The two similarities are then weighted and summed according to preset weights, for example, with attribute similarity weight set to 0.6 and structural similarity weight set to 0.4, forming a multi-dimensional similarity index. This index reflects both the similarity of nodes' business characteristics and captures their structural equivalence in the network, providing a multi-dimensional basis for subsequent clustering.

[0057] The joint calculation mechanism of attribute and structural similarity can accurately identify node clusters with similar risk transmission characteristics, providing a reliable data foundation for the subsequent generation of differentiated risk strategies. This multidimensional similarity calculation method effectively improves the discriminative power of risk clustering, resulting in a substantial improvement in the accuracy of identifying high-risk nodes.

[0058] Furthermore, the risk evolution coefficient is obtained, specifically by generating time slices of a two-layer network structure through preset time intervals, and extracting the risk evolution coefficient for each data domain in continuous time. Arithmetic mean sequence of risk associations over a time slice Calculate the risk evolution coefficient using the following formula. : ,in, This represents the current time slice. , It is a natural constant. This represents the risk evolution coefficient of the current time slice.

[0059] In this embodiment, the preset time interval is set according to the business cycle and the frequency of risk changes. Specifically, it includes: analyzing the time distribution of historical risk events to determine the average time span of risk transmission; and combining the data update frequency to set the time interval to 30 days or quarters and storing it as the preset time interval.

[0060] Time slicing refers to a data processing method that divides a continuous time axis into equal-length intervals. Specifically, it can be implemented using a sliding window mechanism, such as generating a network snapshot every 30 days, and reflecting the dynamic evolution of risk by capturing changes in network structure at different times.

[0061] The arithmetic mean sequence is a quantitative indicator that reflects the fluctuation of the overall risk level in the data field over time. Specifically, it is achieved by calculating the arithmetic mean of the risk values ​​of all nodes in the field. This processing method can eliminate the interference of abnormal fluctuations of individual nodes on the judgment of the overall trend.

[0062] The risk evolution coefficient is a comprehensive parameter that measures the speed and intensity of risk propagation. Its calculation formula introduces an exponential time decay factor, for example, setting the weight of recent time slices higher than that of earlier slices. This design makes the impact of recent risk changes on the evolution coefficient more significant.

[0063] After continuously generating time slices, each data domain calculates the average risk correlation of its contained nodes within the corresponding time slice to form a sequence. The risk evolution coefficient is obtained by calculating the difference in risk correlation between adjacent time slices, multiplying it by an exponential decay factor that increases with the time slice number, and averaging these differences. During the coefficient calculation, for example, as the time slice number increases, the weighting factor assigned by the exponential function increases synchronously, making the contribution of recent risk changes to the evolution coefficient higher than that of earlier changes. This calculation mechanism effectively captures the acceleration characteristics of risk propagation and avoids the loss of time dimension information caused by simple arithmetic averaging in traditional methods.

[0064] This method accurately quantifies the dynamic evolution trend of risk propagation within the data domain, effectively identifying high-risk areas with accelerated propagation characteristics. For example, in monitoring the risk of broken capital chains, this method can promptly detect anomalies in the risk correlation of a supply chain cluster increasing over three consecutive time slices, thereby triggering an early warning mechanism. This time-sensitive risk evolution analysis provides precise decision-making basis for the formulation of dynamic optimization strategies, solving the lag problem of traditional methods in risk trend prediction.

[0065] Furthermore, the risk nodes corresponding to the data domains are extracted, including: comparing the risk evolution coefficient of the current time slice of each data domain with a preset second threshold; if the risk evolution coefficient of the current time slice is greater than the preset second threshold, the data domain is determined to be a high-risk evolution domain, and all risk nodes constituting the data domain are extracted, deduplicated, and uniformly marked as the extracted risk nodes.

[0066] In this embodiment, the preset second threshold is calculated using historical risk evolution data. Specifically, this includes: collecting risk evolution coefficients from historical data and associating them with actual risk outbreak events; determining the threshold through percentile analysis, such that the top 5% of the risk evolution coefficients are marked as high risk, and storing this threshold as the preset second threshold.

[0067] The risk evolution coefficient is a quantitative indicator of the dynamic change of risk calculated through time slice sequences. Specifically, it can be achieved by combining the mean difference of risk correlation in continuous time slices with an exponential decay weight, and is used to reflect the strength of the evolution trend of risk correlation within the data domain.

[0068] The data domain refers to the set of nodes formed through joint clustering. Specifically, density clustering is achieved by using the joint features of attribute similarity and risk correlation difference to divide groups of nodes with similar risk characteristics and evolution patterns.

[0069] Deduplication refers to the process of eliminating duplicate nodes across data domains. Specifically, it can be achieved through hash comparison of node unique identifiers to ensure the uniqueness of the marked nodes.

[0070] A two-layer network time-slice sequence is generated at preset time intervals, and the average risk correlation sequence is extracted from consecutive time slices within each data domain. The change in risk correlation between adjacent time slices is calculated based on exponential decay weights, and the weighted sum is used to obtain the risk evolution coefficient. When this coefficient exceeds a preset second threshold, an abnormal risk evolution trend is determined for that data domain. At this point, all nodes within the data domain are traversed, tracing their original enterprise and funding nodes. A hash table is used to store unique node identifiers to filter for duplicate nodes across domains, ultimately forming a set of risk nodes without redundancy. For example, if the average risk correlation of a data domain increases by 15%, 20%, and 25% in three consecutive time slices, respectively, its risk evolution coefficient may reach 1.8 times the preset threshold, triggering a high-risk evolution domain determination.

[0071] By dynamically calculating the risk evolution coefficient and setting adaptive thresholds, abnormal growth trends in risk correlation can be captured in a timely manner. For example, in scenarios involving the transmission of risks from broken capital chains, combinations of enterprise-capital nodes exhibiting accelerated deterioration characteristics can be identified 2-3 time slices in advance, securing a critical time window for risk management. Simultaneously, by tracing the original nodes and performing deduplication, it ensures that risk management strategies accurately cover all related entities, avoiding resource misallocation caused by duplicate node markings.

[0072] Furthermore, generating differentiated risk optimization strategies based on risk nodes according to preset rules also includes: associating and matching the generated risk handling strategies for risk nodes with subsequent risk event records from an external risk event database within an observation time window; if the associated risk node does not record any risk event within the observation time window, a positive sample label is generated for the risk handling strategy corresponding to the risk node; if a risk event is recorded, a negative sample label is generated; for each labeled sample, a training sample tuple is extracted and constructed; the training sample tuple contains: the feature vector of the risk node, the executed risk handling strategy, and the positive or negative sample label; using the newly added training set composed of the training sample tuple, the graph neural network model is adjusted; the adjustment process uses a weighted cross-entropy loss function, where the loss weight of negative samples is a preset multiple of that of positive samples, a mini-batch gradient descent algorithm is used, and the learning rate is set lower than the learning rate at the initial training of the model; when the performance evaluation index of the adjusted graph neural network model on the newly added training set reaches the preset model performance standard, the adjusted model parameters are updated to update the original graph neural network model.

[0073] In this embodiment, the extracted risk nodes generate differentiated risk optimization strategies based on preset rules. The preset rules are formulated based on risk type and node characteristics, specifically including: matching a predefined risk disposal strategy library according to the enterprise type, capital flow and risk evolution coefficient of the risk node; the strategy library is obtained by training a machine learning model and includes measures such as adjusting credit limits, increasing guarantee requirements and extending liquidation period.

[0074] Risk management strategies refer to fund supervision or transaction restriction measures formulated for high-risk nodes. Specifically, they can be implemented by adjusting credit limits, increasing guarantee requirements, or extending the settlement period. These strategies are automatically generated by a preset rule engine based on node characteristics.

[0075] The observation time window refers to the period for verifying the effect after the strategy is implemented. For example, it can be set to three months or half a year to observe whether the strategy has actually prevented the occurrence of risk events.

[0076] Positive sample labels are confirmation markers of the effectiveness of a strategy. They are generated when no risk events occur at a node during the observation period, indicating that the current strategy effectively suppresses risk transmission.

[0077] Negative sample labels are feedback indicators of policy failure. They are generated when risk events still occur at a node during the observation period, suggesting that the existing policy needs to be optimized.

[0078] The weighted cross-entropy loss function refers to applying a higher penalty weight to misclassified samples during model training. For example, the loss weight of negative samples is set to 3 times that of positive samples to enhance the model's ability to identify risky events.

[0079] Mini-batch gradient descent is an optimization method that uses only a portion of the training data to update parameters in each iteration, such as selecting 256 samples for training each time, thus ensuring training efficiency while avoiding memory overload.

[0080] Once a risk node is identified, a corresponding risk control strategy is automatically generated and implemented. During a preset observation period, new records in the external risk event database are continuously monitored. If a node remains risk-free during the observation period, its features, execution strategy, and positive label are used to form a training sample; if a risk event occurs, a negative-labeled sample is recorded. All labeled samples are integrated into a new training set for incremental model learning. During model tuning, the loss weight of negative samples is increased to make the model focus more on the feature patterns of strategy failure cases. A low learning rate is used for parameter updates, such as one-tenth of the initial learning rate, to avoid over-adjustment that could disrupt the existing knowledge structure. When the model's accuracy and recall on the validation set meet preset standards, such as accuracy reaching 85% and recall not less than 80%, the model version iteration is completed.

[0081] By constructing a closed-loop feedback mechanism for strategy effectiveness, the graph neural network model can continuously learn the latest risk transmission patterns and promptly identify scenarios where the original strategy fails. This dynamic optimization mechanism significantly improves the timeliness of risk warnings, avoids misjudgments or omissions caused by outdated strategies, and enhances the adaptability of credit assessment to complex financial environments.

[0082] Figure 3 This is a schematic diagram of the structure of the financial credit risk assessment and dynamic optimization system based on the fused graph neural network provided in this application embodiment. The system includes: a data acquisition module, used to acquire enterprise entity graphs constructed from enterprise entity data and capital flow graphs constructed from capital flow data, and calculate the risk correlation between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph; a first data processing module, used to acquire the corresponding enterprise nodes and capital nodes and record them as risk nodes if the risk correlation does not exceed a preset first threshold, and record the difference between the risk correlation and the preset first threshold as the first difference, and conduct multiple rounds of risk propagation simulation based on this; a second data processing module, used to recalculate the risk correlation of risk nodes in the risk propagation simulation, and if the recalculated risk correlation does not exceed the preset first threshold, summarize the corresponding risk nodes to obtain a risk-related node set; a third data processing module, used to divide the risk-related node set into at least one data domain, perform time-series graph analysis on each data domain, and obtain the risk evolution coefficient; a fourth data processing module, used to extract the risk nodes corresponding to the corresponding data domain if the risk evolution coefficient exceeds a preset second threshold; and a data output module, used to generate differentiated risk optimization strategies based on the extracted risk nodes.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0087] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A financial credit risk assessment and dynamic optimization method integrating graph neural networks, characterized in that, Includes the following steps: Obtain an enterprise entity graph constructed from enterprise entity data and a capital flow graph constructed from capital flow data, and calculate and obtain the risk correlation between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph. If the risk correlation does not exceed the preset first threshold, the corresponding enterprise node and fund node are obtained and recorded as risk nodes. The difference between the risk correlation and the preset first threshold is recorded as the first difference and multiple rounds of risk propagation simulation are carried out accordingly. In the risk propagation simulation, the risk correlation degree of risk nodes is recalculated. If the recalculated risk correlation degree does not exceed the preset first threshold, the corresponding risk nodes are aggregated to obtain a risk-related node set. The risk-related node set is divided into at least one data domain, and time series graph analysis is performed on each data domain to obtain the risk evolution coefficient; If the risk evolution coefficient exceeds the preset second threshold, then the risk node corresponding to the data domain is extracted. Generate differentiated risk optimization strategies based on the extracted risk nodes; Obtain an enterprise entity graph constructed based on enterprise entity data. The enterprise entity graph contains enterprise nodes and enterprise relationship edges. Enterprise nodes are enterprise feature vectors of a preset dimension, and enterprise relationship edges are transaction feature vectors containing transaction frequency and average transaction amount. The fund flow graph is constructed based on fund flow data. The fund flow graph includes fund nodes and fund relationship edges. Among them, the fund nodes are institutional feature vectors with institutional type and capital adequacy ratio, and the fund relationship edges are channel feature vectors with channel fee rate, average clearing time and historical default number. Based on this, multiple rounds of risk transmission simulations were conducted, specifically including: In each round of risk propagation simulation, starting from the risk node, the preset number of nodes is adjusted according to the first difference, and a random walk is performed on the two-layer network structure according to the adjusted preset number of nodes. The preset node quantity range is used to determine the maximum number of neighboring nodes starting from the risk node. When selecting the next node, a probabilistic selection strategy based on attention weights is adopted, i.e., the next node... From the current node Neighbor set The probability of being selected Calculated using the following formula: ,in, and They are and eigenvectors, It is a shared weight matrix. It is an attention vector. This represents a vector concatenation operation. yes The new eigenvectors obtained after linear transformation of the shared weight matrix yes The new eigenvectors obtained after linear transformation of the shared weight matrix This represents the transpose of the attention vector. This represents the leakage linear rectifier function. Represents an exponential function. Indicates the current node The number of all neighboring nodes; The risk evolution coefficient is obtained, specifically including: Time slices of a two-layer network structure are generated by pre-defined time intervals to extract the continuous data domain. Arithmetic mean sequence of risk associations over a time slice Calculate the risk evolution coefficient using the following formula. : ,in, This represents the current time slice. , It is a natural constant. This represents the risk evolution coefficient of the current time slice.

2. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 1, characterized in that, Calculate and obtain the risk correlation between enterprise nodes in the enterprise entity graph and fund nodes in the fund flow graph, specifically including: By mapping enterprise nodes to fund nodes, a cross-layer connection edge feature vector is constructed, so that the enterprise entity graph and the fund flow graph form a two-layer network structure. The cross-layer connection edge feature vector is composed of account activity and fund aggregation ratio. The enterprise feature vector, transaction feature vector, institution feature vector, channel feature vector, and cross-layer connection edge feature vector are all input into the graph neural network model; The graph neural network model aggregates the features of nodes and edges through a three-layer graph attention network and outputs the risk correlation degree through a fully connected layer.

3. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 2, characterized in that, The preset node quantity range is adjusted based on the first difference, specifically including: The preset scaling factor is extracted from the fused graph neural network database; The preset node quantity range is adjusted according to the correction formula, which is as follows: ,in, This is the corrected range of the preset number of nodes. The preset range of the number of nodes, The first difference, The preset scaling factor, This is the floor function.

4. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 1, characterized in that, The set of risk-related nodes is divided into at least one data domain, specifically including: Calculate the multidimensional similarity between risk nodes in the risk-related node set; The difference between the preset first threshold and the recalculated risk correlation degree corresponding to the risk node is used as the second difference; The second difference is used as the magnitude in the independent dimension direction to obtain the first vector of the risk node. The corresponding multidimensional similarity is used as the magnitude in another orthogonal independent dimension direction and concatenated to obtain the second vector of the risk node. The first vector and the second vector are combined to obtain the joint feature vector of the risk node. A predefined density clustering algorithm is used to cluster the joint feature vectors. The neighborhood radius parameter of the predefined density clustering algorithm is adjusted based on the standard deviation of the joint feature vectors, using the following specific adjustment formula: ,in, The preset scale constant, The standard deviation of the joint eigenvectors, Represents the neighborhood radius parameter; Nodes within the same density connectivity domain are grouped into the same data domain.

5. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 4, characterized in that, Calculate the multidimensional similarity between risk nodes in the risk-related node set, specifically including: Attribute similarity is obtained by calculating the cosine similarity between the enterprise feature vectors or institution feature vectors of two risk nodes. Structural similarity is obtained by calculating the Jaccard similarity coefficient between the adjacency vectors of two risk nodes in a two-layer network structure. The adjacency vector is a Boolean vector that identifies whether a risk node is connected to other risk nodes, taking all nodes in the two-layer network structure as the judgment object. The first and second weights are obtained by extracting them from a fused graph neural network database; Multidimensional similarity is obtained by weighted summation of attribute similarity and structural similarity.

6. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 5, characterized in that, Extract the risk nodes corresponding to the relevant data domain, including: The risk evolution coefficient of the current time slice in each data domain is compared with a preset second threshold; If the risk evolution coefficient of the current time slice is greater than the preset second threshold, the data domain is determined to be a high-risk evolution domain, and all risk nodes constituting the data domain are extracted, deduplicated, and uniformly marked as the extracted risk nodes.

7. The financial credit risk assessment and dynamic optimization method based on fused graph neural networks according to claim 2, characterized in that, The method of generating differentiated risk optimization strategies based on the extracted risk nodes also includes: The generated risk management strategies for risk nodes are correlated and matched with subsequent risk event records from an external risk event database within the observation time window. If no risk event is recorded for the associated risk node within the observation time window, a positive sample label is generated for the risk management strategy corresponding to the risk node; if a risk event is recorded, a negative sample label is generated. For each labeled sample, extract and construct a training sample tuple; The training sample tuple contains: the feature vector of the risk node, the risk management strategy implemented, and the positive or negative sample label; The graph neural network model is tuned using a new training set consisting of training sample tuples. The adjustment process uses a weighted cross-entropy loss function, where the loss weight of negative samples is a preset multiple of that of positive samples. It employs a mini-batch gradient descent algorithm and sets the learning rate to be lower than the learning rate used during the initial training of the model. Once the performance evaluation metrics of the adjusted graph neural network model on the newly added training set reach the preset model performance standard, the adjusted model parameters will be used to update the original graph neural network model.

8. A financial credit risk assessment and dynamic optimization system integrating graph neural networks, characterized in that, include: Data acquisition module: used to acquire enterprise entity graphs constructed based on enterprise entity data and capital flow graphs constructed based on capital flow data, and to calculate and obtain the risk correlation between enterprise nodes in the enterprise entity graph and capital nodes in the capital flow graph; The first data processing module is used to obtain the corresponding enterprise node and fund node and record them as risk nodes if the risk correlation does not exceed the preset first threshold. The difference between the risk correlation and the preset first threshold is recorded as the first difference and multiple rounds of risk propagation simulation are performed accordingly. The second data processing module is used to recalculate the risk correlation degree of risk nodes in the risk propagation simulation. If the recalculated risk correlation degree does not exceed the preset first threshold, the corresponding risk nodes are summarized to obtain a risk-related node set. The third data processing module is used to divide the set of risk-related nodes into at least one data domain, perform time series graph analysis on each data domain, and obtain the risk evolution coefficient. The fourth data processing module is used to extract the risk node corresponding to the data domain if the risk evolution coefficient exceeds the preset second threshold. Data output module: used to generate differentiated risk optimization strategies based on the extracted risk nodes; Obtain an enterprise entity graph constructed based on enterprise entity data. The enterprise entity graph contains enterprise nodes and enterprise relationship edges. Enterprise nodes are enterprise feature vectors of a preset dimension, and enterprise relationship edges are transaction feature vectors containing transaction frequency and average transaction amount. The fund flow graph is constructed based on fund flow data. The fund flow graph includes fund nodes and fund relationship edges. Among them, the fund nodes are institutional feature vectors with institutional type and capital adequacy ratio, and the fund relationship edges are channel feature vectors with channel fee rate, average clearing time and historical default number. Based on this, multiple rounds of risk transmission simulations were conducted, specifically including: In each round of risk propagation simulation, starting from the risk node, the preset number of nodes is adjusted according to the first difference, and a random walk is performed on the two-layer network structure according to the adjusted preset number of nodes. The preset node quantity range is used to determine the maximum number of neighboring nodes starting from the risk node. When selecting the next node, a probabilistic selection strategy based on attention weights is adopted, i.e., the next node... From the current node Neighbor set The probability of being selected Calculated using the following formula: ,in, and They are and eigenvectors, It is a shared weight matrix. It is an attention vector. This represents a vector concatenation operation. yes The new eigenvectors obtained after linear transformation of the shared weight matrix yes The new eigenvectors obtained after linear transformation of the shared weight matrix This represents the transpose of the attention vector. This represents the leakage linear rectifier function. Represents an exponential function. Indicates the current node The number of all neighboring nodes; The risk evolution coefficient is obtained, specifically including: Time slices of a two-layer network structure are generated by pre-defined time intervals to extract the continuous data domain. Arithmetic mean sequence of risk associations over a time slice Calculate the risk evolution coefficient using the following formula. : ,in, This represents the current time slice. , It is a natural constant. This represents the risk evolution coefficient of the current time slice.

Images