Risk control management system and method based on enterprise multi-dimensional data

Abstract

The application provides a risk control management system and method based on enterprise multidimensional data, and belongs to the technical field of enterprise risk management. The system collects multidimensional data such as static attributes, dynamic behaviors and external public opinions of enterprises, generates a standardized feature vector through feature engineering processing, constructs an enterprise heterogeneous graph network, extracts network topology structure features, fuses the network topology structure features, generates dynamic risk portraits such as credit rating, risk transmission sensitivity and fund flow health degree by using a multi-task risk portrait model, constructs a reinforcement learning environment based on the risk portraits and network features, simulates the evolution of the state of enterprises under different risk control strategies, outputs optimal risk control decisions, and optimizes the model through incremental learning based on decision feedback data, so that the system evolves itself. The application realizes closed-loop management from data collection, risk analysis, strategy formulation to effect feedback, and improves the comprehensiveness of risk identification, the real-time performance of strategy response and the continuity of system evolution.

Description

Technical Field

[0001] This invention relates to the field of enterprise risk management technology, and in particular to a risk control management system and method based on enterprise multidimensional data. Background Technology

[0002] Early corporate risk assessments primarily relied on financial indicator analysis and expert judgment, employing statistical methods such as linear regression and discriminant analysis to build risk assessment models. While these methods depend on historical financial data and offer good interpretability, they struggle to capture complex non-linear risk characteristics and require a high degree of data completeness and accuracy.

[0003] With the development of big data technology, machine learning methods are increasingly being applied to enterprise risk control. Credit scoring models built based on algorithms such as decision trees and support vector machines can handle more multi-dimensional enterprise data, improving the accuracy of risk assessment. However, these methods still primarily focus on individual enterprise characteristics, neglecting the impact of complex inter-enterprise networks on risk transmission.

[0004] In recent years, the introduction of graph computing technology has provided new insights into analyzing corporate relationships. By constructing corporate relationship networks, potential risks such as guarantee circles and supply chain risks can be identified. However, existing methods are mostly limited to the analysis of single types of relationships, lacking the ability to uniformly model multi-dimensional relationship networks, and the integration of network characteristics with traditional risk assessment models is insufficient.

[0005] In terms of dynamic risk assessment, existing technologies mostly employ time series analysis methods to predict future risk trends using historical enterprise data. However, these methods often struggle to respond in real-time to changes in the external environment and lack the ability to dynamically simulate risk transmission paths. Furthermore, existing risk control strategies largely rely on pre-set rules, making it difficult to adaptively adjust to changes in the enterprise's risk status.

[0006] Currently, enterprise risk control faces multiple technical challenges: how to effectively integrate multi-source heterogeneous data within an enterprise, how to accurately quantify the risk transmission effect in interconnected networks, how to achieve dynamic risk assessment and prediction, and how to generate personalized risk control strategies. Solving these problems requires building a complete technical system capable of integrating multi-dimensional data, network analysis, and dynamic modeling.

[0007] Therefore, there is an urgent need in this field for an innovative method that can overcome the limitations of existing technologies, make full use of multi-dimensional enterprise data and related network information, and achieve accurate, dynamic, and intelligent risk management. Summary of the Invention

[0008] The purpose of this invention is to provide a risk control management system and method based on enterprise multidimensional data, which integrates enterprise multi-source data and related networks, constructs dynamic risk profiles, uses reinforcement learning to generate optimal risk control strategies in real time, and continuously optimizes them through incremental learning to achieve accurate, efficient, and adaptive intelligent risk control management.

[0009] To achieve the above objectives, this invention provides a risk control management method based on enterprise multidimensional data, comprising the following steps:

[0010] Step S1: Obtain multi-dimensional data of the target company, and perform feature engineering on the multi-dimensional data to obtain standardized feature vectors;

[0011] Step S2: Construct a heterogeneous enterprise graph network based on the acquired multi-dimensional data, and extract the network topology features, where nodes represent enterprises and edges represent multi-dimensional relationships between enterprises.

[0012] Step S3: Fuse the standardized feature vector with the network topology features and input it into the multi-task risk profiling model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity and cash flow health.

[0013] Step S4: Using dynamic enterprise risk profiles and network topology features as state inputs, construct a reinforcement learning environment to simulate the enterprise state evolution path under different risk control strategies and output the optimal risk control decision.

[0014] Step S5: Execute the optimal risk control decision, and based on the enterprise status change data after the decision is executed, perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model.

[0015] Preferably, in step S1, the multi-dimensional data specifically includes:

[0016] Static attribute data: registered capital of the enterprise, industry code of the main business, and business registration address;

[0017] Dynamic behavioral data: inter-company transaction records, recent The quarterly tax payment sequence and average monthly bank statement, among which ≥8;

[0018] External public opinion data: Text data extracted from news media, social media, and regulatory announcements, which includes information on corporate co-occurrence relationships;

[0019] Feature engineering processing for multi-dimensional data includes:

[0020] One-hot encoding and normalization are performed on static attribute data to generate static feature vectors. ;

[0021] Extract transaction frequency and transaction amount features from transaction records in dynamic behavioral data, calculate the coefficient of variation for tax amount sequences, calculate volatility indicators for bank transaction records, and generate dynamic feature vectors. ;

[0022] External public opinion data is analyzed by extracting textual features using the BERT model and combining them with sentiment analysis. Enterprise associations are constructed based on co-occurrence relationships to generate public opinion feature vectors. ;

[0023] Will , , By concatenating and reducing the dimensionality, a standardized feature vector is obtained. .

[0024] Preferably, step S2 specifically includes the following steps:

[0025] Step S21: Define five types of enterprise relationships based on the acquired multi-dimensional data:

[0026] Address association: Extracting geographical cluster relationships based on business registration addresses;

[0027] Transaction Linkage: Building a supply chain relationship network based on transaction records;

[0028] Public opinion correlation: Establishing a public opinion dissemination network based on the co-occurrence relationships of enterprises in text data;

[0029] Equity Connections: Constructing a capital connection network based on equity shareholding ratios;

[0030] Guarantee Association: Constructing a credit association network based on guarantee relationship data;

[0031] Step S22: For the five types of relationships, construct a heterogeneous graph neural network based on meta-paths, where the node feature matrix... Using standardized feature vectors Perform initialization;

[0032] Step S23: For each type of association... Calculate the corresponding normalized adjacency matrix for each. ,in Representing relation type The adjacency matrix, This represents the corresponding degree matrix;

[0033] Step S24, in the In a multi-layer network, multi-relation feature fusion is performed using the following formula:

[0034] ;

[0035] in, Indicates the first The node feature matrix of the layer Indicates the first Layers based on relation types The trainable weight matrix, This represents a vector concatenation operation;

[0036] Step S25, after After layer propagation, the final output node representation Chinese corresponding enterprises The vector is used as the topological feature vector. .

[0037] Preferably, step S3 specifically includes the following steps:

[0038] Step S31: Standardize the feature vector With topological feature vectors The features are concatenated to form an enhanced feature vector. ;

[0039] Step S32: The enhanced feature vector The shared encoder of the input multi-task risk profiling model is used to extract shared feature representations through a multilayer perceptron. :

[0040] ;

[0041] in, This represents the weight matrix of the shared encoder. Indicates the bias term;

[0042] Step S33: Represent the shared features Input three task-specific decoders respectively:

[0043] Credit Rating Decoder:

[0044] ;

[0045] in, Indicates credit rating, This represents the weight matrix of the credit rating decoder. Indicates the bias term;

[0046] Risk Transmission Sensitivity Decoder:

[0047] Using a graph attention mechanism, the attention weights of neighboring nodes are first calculated. :

[0048] ;

[0049] in, and These represent the feature representations of the current node and its neighboring nodes, respectively. Represents the attention vector. Indicates transpose. This represents the weight matrix of the risk transmission sensitivity decoder. This represents a vector concatenation operation;

[0050] Then calculate the risk transmission sensitivity. :

[0051] ;

[0052] in, Representing neighboring nodes Risk score, Indicates the bias term. Indicates enterprise The set of neighboring nodes in a graph network;

[0053] Cash Flow Health Decoder:

[0054] ;

[0055] in, Represents a temporal convolutional network. This represents the weight matrix of the cash flow health decoder. Indicates the bias term;

[0056] Step S34: The total loss function of the multi-task risk profiling model is:

[0057] ;

[0058] in, This represents the mean squared error loss in credit rating. Cross-entropy loss, representing the sensitivity to risk transmission. Huber's losses indicate the health of its cash flow. , , This represents the task weight coefficient.

[0059] Preferably, the risk transmission sensitivity calculation incorporates a multi-hop propagation mechanism:

[0060] definition Jump risk transmission sensitivity :

[0061] ;

[0062] in, Represents a node of Jump to neighbor set, Indicates based on transaction association and public opinion association Jump propagation weights;

[0063] The final risk transmission sensitivity is the weighted sum of the sensitivities of each jump:

[0064] ;

[0065] in, Indicates the sensitivity to the final risk transmission. Indicates the first The weighting coefficient of the jump. This represents the maximum number of hops and the weights of each layer. Learned through attention mechanisms, and .

[0066] Preferably, the calculation of cash flow health specifically uses the dynamic behavioral data processed in step S1:

[0067] ;

[0068] in, This represents the baseline confirmation degree of fund flows calculated based on transaction amounts and bank statement data in transaction records. Indicators representing volatility in cash flows and Represents the normalization function. and Indicates the weighting coefficient;

[0069] Fund Flow Volatility Indicator Calculation using an improved coefficient of variation:

[0070] ;

[0071] in, Indicates the first Monthly transaction amount and bank statement summary data, express The average over several months Indicates the size of the time window. This represents the smoothing constant.

[0072] Preferably, step S4 specifically includes:

[0073] state space In Based on dynamic enterprise risk profiles This represents a weighted summary vector of the risk profile of the first-order neighbor nodes of the target enterprise extracted from the heterogeneous enterprise graph network;

[0074] Action space This includes adjustments to credit limits and risk warning levels, with the scope of adjustments determined based on the company's registered capital and operating scale.

[0075] reward function :

[0076] ;

[0077] in, This indicates the credit score at the current time step. This indicates the credit score at the previous time step. Indicates the execution of an action The resulting business costs, This indicates the sensitivity to risk transmission at the current time step. and The hyperparameters representing the balancing weights.

[0078] Preferably, business costs adopt a dynamic weighting mechanism:

[0079] ;

[0080] in, This represents the cost coefficient for adjusting the credit limit. This indicates the cost factor for raising the warning level; This indicates that increasing the monitoring frequency increases the cost factor. This indicates the amount of credit limit adjustment. This indicates an indicator function; its value is 1 if the condition within the parentheses is met, and 0 otherwise.

[0081] Dynamic cost coefficient:

[0082] ;

[0083] in, Indicates the basic cost coefficient. This represents a market environment indicator calculated based on sentiment analysis results from external public opinion feature vectors. This represents the sensitivity parameter.

[0084] Preferably, step S5 specifically includes:

[0085] Define model performance degradation index :

[0086] ;

[0087] in, Indicates the first The actual risk label value at each point in time. Indicates the first Model predictions at each time point This represents the standard deviation of the true risk label value;

[0088] when At that time, incremental learning based on importance weighting is triggered:

[0089] ;

[0090] in, This indicates the preset performance degradation threshold. Indicates the number of newly added samples. Indicates the sample index. Indicates the first The importance weight of each sample This represents the total loss function of the multi-task risk profiling model. Indicates the first The true risk label value of each sample Indicates the first The model prediction value for each sample;

[0091] The calculation formula is as follows:

[0092] ;

[0093] in, Indicates the first The prediction error for each sample. Indicates the first Data freshness of each sample Indicates the first Data integrity index for each sample , , Indicates the weighting coefficient;

[0094] Update the multi-task risk profiling model and reinforcement learning model using the optimized model parameters.

[0095] This invention also provides a risk control management system based on enterprise multidimensional data, including:

[0096] The data acquisition and preprocessing module is used to collect multi-dimensional data from the target enterprise and perform feature engineering on the multi-dimensional data to generate standardized feature vectors.

[0097] The heterogeneous graph network construction and topology feature extraction module is used to construct enterprise heterogeneous graph networks based on multi-dimensional data and extract network topology features.

[0098] The multi-task risk profile generation module is used to fuse standardized feature vectors with network topology features and input them into the multi-task risk profile model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity, and cash flow health.

[0099] The reinforcement learning decision module is used to construct a reinforcement learning environment with dynamic enterprise risk profiles and network topology features as state inputs, simulate the enterprise state evolution path under different risk control strategies, and output the optimal risk control decision.

[0100] The incremental learning optimization module is used to execute optimal risk control decisions and, based on the enterprise status change data after the decision is executed, to perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model.

[0101] Therefore, the present invention employs the above-mentioned risk control management system and method based on enterprise multidimensional data, and the beneficial technical effects are as follows:

[0102] (1) By integrating multi-dimensional data of enterprises with the topological characteristics of enterprise-related networks, a dynamic enterprise risk profile including credit rating, risk transmission sensitivity and cash flow health is generated. This overcomes the limitations of traditional methods that rely on only single-dimensional data and can more comprehensively reflect the overall risk status of enterprises.

[0103] (2) Based on dynamic enterprise risk profiles and network topology features, a reinforcement learning environment is constructed. By simulating the enterprise state evolution path under different risk control strategies, the optimal risk control decision that adapts to the current risk state of the enterprise can be output, which solves the problem that traditional static risk control strategies are difficult to adapt to the dynamic risk changes of enterprises.

[0104] (3) By executing risk control decisions and continuously optimizing the risk profiling model and reinforcement learning model based on feedback data, a complete technical closed loop from data collection, risk analysis, decision making to effect feedback is formed, enabling the system to have the ability to improve itself and continuously optimize. Attached Figure Description

[0105] Figure 1 This is a flowchart of the risk control management method based on enterprise multidimensional data according to the present invention;

[0106] Figure 2 This is an architecture diagram of the risk control management system based on enterprise multidimensional data of the present invention. Detailed Implementation

[0107] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0108] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0109] Example 1

[0110] like Figure 1 As shown, the risk control management method based on enterprise multidimensional data includes the following steps:

[0111] Step S1: Obtain multi-dimensional data of the target company and perform feature engineering on the multi-dimensional data to obtain standardized feature vectors.

[0112] Multidimensional data specifically includes:

[0113] Static attribute data: registered capital of the enterprise, industry code of the main business, and business registration address;

[0114] Dynamic behavioral data: inter-company transaction records, recent The quarterly tax payment sequence and average monthly bank statement, among which ≥8;

[0115] External public opinion data: Text data extracted from news media, social media and regulatory announcements, which includes information on corporate co-occurrence relationships.

[0116] Feature engineering processing for multi-dimensional data includes:

[0117] One-hot encoding and normalization are performed on static attribute data to generate static feature vectors. ;

[0118] Extract transaction frequency and transaction amount features from transaction records in dynamic behavioral data, and calculate the coefficient of variation for the tax amount sequence:

[0119] ;

[0120] in, Indicates standard deviation, This represents the mean.

[0121] Calculate volatility indicators from bank transaction records and generate dynamic feature vectors. ;

[0122] External public opinion data is analyzed by extracting textual features using the BERT model and combining them with sentiment analysis. Enterprise associations are constructed based on co-occurrence relationships to generate public opinion feature vectors. ;

[0123] eigenvectors , , By concatenating and reducing the dimensionality, a standardized feature vector is obtained. .

[0124] Step S2: Construct a heterogeneous enterprise graph network based on the acquired multi-dimensional data, and extract the network topology features, where nodes represent enterprises and edges represent multi-dimensional relationships between enterprises.

[0125] Step S2 specifically includes the following steps:

[0126] Step S21: Define five types of enterprise relationships based on the acquired multi-dimensional data:

[0127] Address association: Extracting geographical cluster relationships based on business registration addresses;

[0128] Transaction Linkage: Building a supply chain relationship network based on transaction records;

[0129] Public opinion correlation: Establishing a public opinion dissemination network based on the co-occurrence relationships of enterprises in text data;

[0130] Equity Connections: Constructing a capital connection network based on equity shareholding ratios;

[0131] Guarantee Association: Constructing a credit association network based on guarantee relationship data;

[0132] Step S22: For the five types of relationships, construct a heterogeneous graph neural network based on meta-paths, where the node feature matrix... Using standardized feature vectors Perform initialization;

[0133] Step S23: For each type of association... Calculate the corresponding normalized adjacency matrix for each. ,in Representing relation type The adjacency matrix, This represents the corresponding degree matrix;

[0134] Step S24, in the In a multi-layer network, multi-relation feature fusion is performed using the following formula:

[0135] ;

[0136] in, Indicates the first The node feature matrix of the layer Indicates the first Layers based on relation types The trainable weight matrix, This represents a vector concatenation operation;

[0137] Step S25, after After propagation through 3 layers, the final output node representation will be... Chinese corresponding enterprises The vector is used as the topological feature vector. The feature vector The structural information of enterprises in five types of relationship networks is comprehensively encoded and used for multi-source feature fusion in step S3.

[0138] Step S3: Fuse the standardized feature vector with the network topology features and input it into the multi-task risk profiling model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity, and cash flow health.

[0139] Step S3 specifically includes the following steps:

[0140] Step S31: Standardize the feature vector With topological feature vectors The features are concatenated to form an enhanced feature vector. ;

[0141] Step S32: The enhanced feature vector The shared encoder of the multi-task risk profiling model (3-layer MLP, with hidden layer dimensions of [64, 32, 16], using Layer Normalization) is input, and the shared feature representation is extracted through a multilayer perceptron. :

[0142] ;

[0143] in, This represents the weight matrix of the shared encoder. Indicates the bias term;

[0144] Step S33: Represent the shared features Input three task-specific decoders respectively:

[0145] Credit Rating Decoder:

[0146] ;

[0147] in, Indicates credit rating, This represents the weight matrix of the credit rating decoder. Indicates the bias term;

[0148] Risk Transmission Sensitivity Decoder:

[0149] Using a graph attention mechanism (4 attention heads, 16 attention dimensions), the attention weights of neighboring nodes are first calculated. :

[0150] ;

[0151] in, and These represent the feature representations of the current node and its neighboring nodes, respectively. Represents the attention vector. Indicates transpose. This represents the weight matrix of the risk transmission sensitivity decoder. This represents a vector concatenation operation;

[0152] Then calculate the risk transmission sensitivity. :

[0153] ;

[0154] in, Representing neighboring nodes Risk score, Indicates the bias term. Indicates enterprise The set of neighboring nodes in a graph network;

[0155] Cash Flow Health Decoder:

[0156] ;

[0157] in, This represents a temporal convolutional network (TCN) with a kernel size of 3, 4 layers, dilation coefficients [1, 2, 4, 8], and an output range of [0, 1]. This represents the weight matrix of the cash flow health decoder. Indicates the bias term;

[0158] Step S34: The total loss function of the multi-task risk profiling model is:

[0159] ;

[0160] in, This represents the mean squared error loss in credit rating. Cross-entropy loss, representing the sensitivity to risk transmission. Huber's losses indicate the health of its cash flow. , , This represents the task weight coefficient, in this embodiment... , , Set them to 0.4, 0.3, and 0.3 respectively.

[0161] Risk transmission sensitivity calculation introduces a multi-hop propagation mechanism:

[0162] definition Jump risk transmission sensitivity :

[0163] ;

[0164] in, Represents a node of Jump to neighbor set, Indicates based on transaction association and public opinion association Jump propagation weights;

[0165] The final risk transmission sensitivity is the weighted sum of the sensitivities of each jump:

[0166] ;

[0167] in, Indicates the sensitivity to the final risk transmission. Indicates the first The weighting coefficient of the jump. This represents the maximum number of hops and the weights of each layer. Learned through attention mechanisms, and .

[0168] The calculation of cash flow health specifically uses the dynamic behavioral data processed in step S1:

[0169] ;

[0170] in, This represents the baseline confirmation degree of fund flows calculated based on transaction amounts and bank statement data in transaction records. Indicators representing volatility in cash flows and Represents the normalization function. and Indicates the weighting coefficient;

[0171] Fund Flow Volatility Indicator Calculation using an improved coefficient of variation:

[0172] ;

[0173] in, Indicates the first Monthly transaction amount and bank statement summary data, express The average over several months Indicates the size of the time window. This represents the smoothing constant.

[0174] Step S4: Using dynamic enterprise risk profiles and network topology features as state inputs, construct a reinforcement learning environment to simulate the enterprise state evolution path under different risk control strategies and output the optimal risk control decision.

[0175] Step S4 specifically includes:

[0176] state space In Based on dynamic enterprise risk profiles This represents a weighted summary vector of the risk profile of the first-order neighbor nodes of the target enterprise extracted from the heterogeneous enterprise graph network;

[0177] Action space This includes adjustments to credit limits [-30%, -20%, -10%, 0, +10%, +20%] and risk warning level settings [normal attention, enhanced attention, key monitoring]. The scope of adjustment is determined based on the company's registered capital and operating scale.

[0178] reward function :

[0179] ;

[0180] in, This indicates the credit score at the current time step. This indicates the credit score at the previous time step. Indicates the execution of an action The resulting business costs, This indicates the sensitivity to risk transmission at the current time step. and The hyperparameter representing the balancing weights, in this embodiment , Set them to 0.6 and 0.4 respectively.

[0181] Operating costs are subject to a dynamic weighting mechanism:

[0182] ;

[0183] in, This represents the cost coefficient for adjusting the credit limit. This indicates the cost factor for raising the warning level; This indicates that increasing the monitoring frequency increases the cost factor. This indicates the amount of credit limit adjustment. This indicates an indicator function; its value is 1 if the condition within the parentheses is met, and 0 otherwise. This indicates the function that signals an escalation of the warning level. This function indicates an increase in monitoring frequency.

[0184] Dynamic cost coefficient:

[0185] ;

[0186] in, Indicates the basic cost coefficient. This represents a market environment indicator calculated based on sentiment analysis results from processed external public opinion feature vectors. This represents the sensitivity parameter.

[0187] Step S5: Execute the optimal risk control decision, and based on the enterprise status change data after the decision is executed, perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model.

[0188] Step S5 specifically includes:

[0189] Define model performance degradation index :

[0190] ;

[0191] in, Indicates the first The actual risk label value at each point in time. Indicates the first Model predictions at each time point This represents the standard deviation of the true risk label value;

[0192] when At that time, incremental learning based on importance weighting is triggered:

[0193] ;

[0194] in, This indicates the preset performance degradation threshold. Indicates the number of newly added samples. Indicates the sample index. Indicates the first The importance weight of each sample This represents the total loss function of the multi-task risk profiling model. Indicates the first The true risk label value of each sample Indicates the first The model prediction value for each sample;

[0195] The calculation formula is as follows:

[0196] ;

[0197] in, Indicates the first The prediction error for each sample. Indicates the first The data freshness of each sample is calculated based on the data timestamp. Indicates the first Data integrity index for each sample , , These represent weighting coefficients, set to 0.4, 0.3, and 0.3 respectively.

[0198] Update the multi-task risk profiling model and reinforcement learning model using the optimized model parameters.

[0199] Example 2

[0200] like Figure 2 As shown, the risk control management system based on enterprise multidimensional data includes:

[0201] The data acquisition and preprocessing module is used to collect multi-dimensional data from the target enterprise and perform feature engineering on the multi-dimensional data to generate standardized feature vectors.

[0202] The heterogeneous graph network construction and topology feature extraction module is used to construct enterprise heterogeneous graph networks based on multi-dimensional data and extract network topology features.

[0203] The multi-task risk profile generation module is used to fuse standardized feature vectors with network topology features and input them into the multi-task risk profile model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity, and cash flow health.

[0204] The reinforcement learning decision module is used to construct a reinforcement learning environment with dynamic enterprise risk profiles and network topology features as state inputs, simulate the enterprise state evolution path under different risk control strategies, and output the optimal risk control decision.

[0205] The incremental learning optimization module is used to execute optimal risk control decisions and perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model based on the enterprise status change data after the decision is executed.

[0206] Therefore, the present invention adopts the above-mentioned risk control management system and method based on enterprise multidimensional data, integrates enterprise multi-source data and related networks, constructs dynamic risk profiles, uses reinforcement learning to generate optimal risk control strategies in real time, and continuously optimizes them through incremental learning, so as to achieve accurate, efficient and adaptive intelligent risk control management.

[0207] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A risk control management method based on enterprise multidimensional data, characterized in that, Includes the following steps: Step S1: Obtain multi-dimensional data of the target company, and perform feature engineering on the multi-dimensional data to obtain standardized feature vectors; Step S2: Construct a heterogeneous enterprise graph network based on the acquired multi-dimensional data, and extract the network topology features, where nodes represent enterprises and edges represent multi-dimensional relationships between enterprises. Step S3: Fuse the standardized feature vector with the network topology features and input it into the multi-task risk profiling model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity and cash flow health. Step S4: Using dynamic enterprise risk profiles and network topology features as state inputs, construct a reinforcement learning environment to simulate the enterprise state evolution path under different risk control strategies and output the optimal risk control decision. Step S5: Execute the optimal risk control decision, and based on the enterprise status change data after the decision is executed, perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model; Step S2 specifically includes the following steps: Step S21: Define five types of enterprise relationships based on the acquired multi-dimensional data: Address association: Extracting geographical cluster relationships based on business registration addresses; Transaction Linkage: Building a supply chain relationship network based on transaction records; Public opinion correlation: Establishing a public opinion dissemination network based on the co-occurrence relationships of enterprises in text data; Equity Connections: Constructing a capital connection network based on equity shareholding ratios; Guarantee Association: Constructing a credit association network based on guarantee relationship data; Step S22: For the five types of relationships, construct a heterogeneous graph neural network based on meta-paths, where the node feature matrix... Using standardized feature vectors Perform initialization; Step S23: For each type of association... Calculate the corresponding normalized adjacency matrix for each. ,in Representing relation type The adjacency matrix, This represents the corresponding degree matrix; Step S24, in the In a multi-layer network, multi-relation feature fusion is performed using the following formula: ； in, Indicates the first The node feature matrix of the layer Indicates the first Layers based on relation types The trainable weight matrix, This represents a vector concatenation operation; Step S25, after After layer propagation, the final output node representation Chinese corresponding enterprises The vector is used as the topological feature vector. .

2. The risk control management method based on enterprise multidimensional data according to claim 1, characterized in that, In step S1, the multi-dimensional data specifically includes: Static attribute data: company registered capital, main business industry code, and business registration address; Dynamic behavioral data: inter-company transaction records, recent The quarterly tax payment sequence and average monthly bank statement, among which ≥8; External public opinion data: Text data extracted from news media, social media, and regulatory announcements, which includes information on corporate co-occurrence relationships; Feature engineering processing for multi-dimensional data includes: One-hot encoding and normalization are performed on static attribute data to generate static feature vectors. ; Extract transaction frequency and transaction amount features from transaction records in dynamic behavioral data, calculate the coefficient of variation for tax amount sequences, calculate volatility indicators for bank transaction records, and generate dynamic feature vectors. ; External public opinion data is analyzed by extracting textual features using the BERT model and combining them with sentiment analysis. Enterprise associations are constructed based on co-occurrence relationships to generate public opinion feature vectors. ; Will , , By concatenating and reducing the dimensionality, a standardized feature vector is obtained. .

3. The risk control management method based on enterprise multidimensional data according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Standardize the feature vector With topological feature vectors The features are concatenated to form an enhanced feature vector. ; Step S32: The enhanced feature vector The shared encoder of the input multi-task risk profiling model is used to extract shared feature representations through a multilayer perceptron. : ； in, This represents the weight matrix of the shared encoder. Indicates the bias term; Step S33: Represent the shared features Input three task-specific decoders respectively: Credit Rating Decoder: ； in, Indicates credit rating, This represents the weight matrix of the credit rating decoder. Indicates the bias term; Risk Transmission Sensitivity Decoder: Using a graph attention mechanism, the attention weights of neighboring nodes are first calculated. : ； in, and These represent the feature representations of the current node and its neighboring nodes, respectively. Represents the attention vector. Indicates transpose. This represents the weight matrix of the risk transmission sensitivity decoder. This represents a vector concatenation operation; Then calculate the risk transmission sensitivity. : ； in, Representing neighboring nodes Risk score, Indicates the bias term. Indicates enterprise The set of neighboring nodes in a graph network; Cash Flow Health Decoder: ； in, Represents a temporal convolutional network. This represents the weight matrix of the cash flow health decoder. Indicates the bias term; Step S34: The total loss function of the multi-task risk profiling model is: ； in, This represents the mean squared error loss in credit rating. Cross-entropy loss, representing the sensitivity to risk transmission. Huber's losses indicate the health of its cash flow. , , This represents the task weight coefficient.

4. The risk control management method based on enterprise multidimensional data according to claim 3, characterized in that, Risk transmission sensitivity calculation introduces a multi-hop propagation mechanism: definition Jump risk transmission sensitivity : ； in, Represents a node of Jump to neighbor set, Indicates based on transaction association and public opinion association Jump propagation weights; The final risk transmission sensitivity is the weighted sum of the sensitivities of each jump: ； in, Indicates the sensitivity to the final risk transmission. Indicates the first The weighting coefficient of the jump. This represents the maximum number of hops and the weights of each layer. Learned through attention mechanisms, and .

5. The risk control management method based on enterprise multidimensional data according to claim 4, characterized in that, The calculation of cash flow health specifically uses the dynamic behavioral data processed in step S1: ； in, This represents the baseline confirmation degree of fund flows calculated based on transaction amounts and bank statement data in transaction records. Indicators representing volatility in cash flows and Represents the normalization function. and Indicates the weighting coefficient; Fund Flow Volatility Indicator Calculation using an improved coefficient of variation: ； in, Indicates the first Monthly transaction amount and bank statement summary data, express The average over several months Indicates the size of the time window. This represents the smoothing constant.

6. The risk control management method based on enterprise multidimensional data according to claim 5, characterized in that, Step S4 specifically includes: state space In Based on dynamic enterprise risk profiles This represents a weighted summary vector of the risk profile of the first-order neighbor nodes of the target enterprise extracted from the heterogeneous enterprise graph network; Action space This includes adjustments to credit limits and risk warning levels, with the scope of adjustments determined based on the company's registered capital and operating scale. reward function : ； in, This indicates the credit score at the current time step. This indicates the credit score at the previous time step. Indicates the execution of an action The resulting business costs, This indicates the sensitivity to risk transmission at the current time step. and The hyperparameters representing the balancing weights.

7. The risk control management method based on enterprise multidimensional data according to claim 6, characterized in that, Operating costs are subject to a dynamic weighting mechanism: ； in, This represents the cost coefficient for adjusting the credit limit. This indicates the cost factor for raising the warning level; This indicates that increasing the monitoring frequency increases the cost factor. This indicates the amount of credit limit adjustment. This indicates an indicator function; its value is 1 if the condition within the parentheses is met, and 0 otherwise. Dynamic cost coefficient: ； in, Indicates the basic cost coefficient. This represents a market environment indicator calculated based on sentiment analysis results from external public opinion feature vectors. This represents the sensitivity parameter.

8. The risk control management method based on enterprise multidimensional data according to claim 7, characterized in that, Step S5 specifically includes: Define model performance degradation index : ； in, Indicates the first The actual risk label value at each point in time. Indicates the first Model predictions at each time point This represents the standard deviation of the true risk label value; when At that time, incremental learning based on importance weighting is triggered: ； in, This indicates the preset performance degradation threshold. Indicates the number of newly added samples. Indicates the sample index. Indicates the first The importance weight of each sample This represents the total loss function of the multi-task risk profiling model. Indicates the first The true risk label value of each sample Indicates the first The model prediction value for each sample; The calculation formula is as follows: ； in, Indicates the first The prediction error for each sample. Indicates the first Data freshness of each sample Indicates the first Data integrity index for each sample , , Indicates the weighting coefficient; Update the multi-task risk profiling model and reinforcement learning model using the optimized model parameters.

9. A risk control management system based on enterprise multidimensional data, characterized in that, The method for implementing risk control management based on enterprise multidimensional data as described in any one of claims 1-8 includes: The data acquisition and preprocessing module is used to collect multi-dimensional data from the target enterprise and perform feature engineering on the multi-dimensional data to generate standardized feature vectors. The heterogeneous graph network construction and topology feature extraction module is used to construct enterprise heterogeneous graph networks based on multi-dimensional data and extract network topology features. The multi-task risk profile generation module is used to fuse standardized feature vectors with network topology features and input them into the multi-task risk profile model to generate a dynamic enterprise risk profile that includes credit rating, risk transmission sensitivity, and cash flow health. The reinforcement learning decision module is used to construct a reinforcement learning environment with dynamic enterprise risk profiles and network topology features as state inputs, simulate the enterprise state evolution path under different risk control strategies, and output the optimal risk control decision. The incremental learning optimization module is used to execute optimal risk control decisions and perform incremental learning optimization on the multi-task risk profiling model and reinforcement learning model based on the enterprise status change data after the decision is executed.

Images