Comprehensive contribution evaluation method and system based on reinforcement learning

Abstract

The invention provides a comprehensive contribution assessment method and system based on reinforcement learning, and belongs to the technical field of financial science and technology. According to the method, multi-source heterogeneous data of enterprise customers are integrated, and financial index feature vectors and transaction behavior feature vectors are extracted to construct a fusion feature matrix. And a judgment mechanism of a fluctuation amplitude threshold is introduced, and when the fluctuation amplitude of the fusion feature matrix exceeds a limit, a transaction behavior trend is predicted to generate a dynamic risk vector. The contribution score is determined by fusing the feature matrix and the dynamic risk vector, and the industry enterprise feature data is grouped to calibrate the contribution score and correct the evaluation deviation by comparing the industry contribution benchmark and the condition exceeding the preset deviation threshold, so that the industry suitability and objectivity of contribution level judgment are ensured. And finally, the credit line is determined according to the contribution level, so that the bank credit resource optimization configuration is realized on the premise of dynamic risk controllability, the fund configuration efficiency and the risk control capability are improved, and the real operation demand of an enterprise can be accurately matched.

Description

Technical Field

[0001] This disclosure relates to the field of financial technology, specifically to a comprehensive contribution evaluation method and system based on reinforcement learning. Background Technology

[0002] Currently, corporate lending is a crucial pillar of the financial industry, directly impacting banks' capital allocation efficiency and risk control capabilities. As the primary recipients of bank credit, assessing the comprehensive contribution of corporate clients is not only a core element of credit decision-making but also key to optimizing resource allocation and enhancing market competitiveness. Accurately assessing the comprehensive contribution of corporate clients helps banks rationally allocate credit resources under controllable risk, thereby supporting the development of the real economy. However, traditional comprehensive contribution assessment methods typically rely on static indicator systems and fixed rule models, failing to adapt to the complex and ever-changing operating environments of corporate clients. Furthermore, when faced with multi-dimensional data from corporate clients, they cannot fully capture the true value of their contributions, making it difficult for traditional comprehensive contribution assessment methods to simultaneously meet the requirements of dynamism and accuracy.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] In view of this, this disclosure provides a comprehensive contribution evaluation method based on reinforcement learning, which can improve the accuracy of enterprise customer contribution evaluation.

[0005] In a first aspect, embodiments of this application provide a comprehensive contribution evaluation method based on reinforcement learning. The method includes: acquiring multi-source heterogeneous data from a company's financial statements and transaction records, the multi-source heterogeneous data including financial indicator values ​​and transaction behavior sequences; determining a customer status dataset based on the financial indicator values ​​and the transaction behavior sequences; extracting key feature vectors from the customer status dataset based on principal component analysis, the key feature vectors including financial indicator feature vectors and transaction behavior feature vectors; determining a fusion feature matrix based on the financial indicator feature vectors and the transaction behavior feature vectors; determining whether the fluctuation amplitude of the fusion feature matrix exceeds a preset fluctuation amplitude threshold; if it exceeds the preset fluctuation amplitude threshold, standardizing the transaction behavior feature vectors to determine standardized transaction behavior sequence features. The process involves: using a feature matrix; predicting transaction behavior trends based on a recurrent neural network and the standardized transaction behavior sequence feature matrix; determining a dynamic risk vector based on the transaction behavior trends; determining a contribution score based on the dynamic risk vector and the fused feature matrix; determining whether the absolute value of the deviation between the current contribution score and the industry's enterprise contribution benchmark exceeds a preset deviation threshold; if it exceeds the preset deviation threshold, acquiring feature data from multiple enterprise customers within the industry, including annual revenue, transaction frequency, and customer loyalty, and using this data as a dataset to be clustered; grouping the dataset to be clustered based on the K-means clustering algorithm to determine the grouping results for enterprise customer contribution levels; updating the contribution score based on the grouping results; determining the contribution level based on the updated contribution score; and determining the credit limit for enterprise customers based on the contribution level.

[0006] Secondly, embodiments of this application provide a comprehensive contribution evaluation method system based on reinforcement learning, the system comprising: a first acquisition module, a first determination module, an extraction module, a second determination module, a first judgment module, a processing module, a prediction module, a third determination module, a fourth determination module, a second judgment module, a second acquisition module, a fifth determination module, an update module, a sixth determination module, and a seventh determination module. The system comprises the following modules: a first acquisition module, used to acquire multi-source heterogeneous data from enterprise customers' financial statements and transaction records, including financial indicator values ​​and transaction behavior sequences; a first determination module, used to determine the customer status dataset based on the financial indicator values ​​and transaction behavior sequences; an extraction module, used to extract key feature vectors from the customer status dataset based on principal component analysis, including financial indicator feature vectors and transaction behavior feature vectors; a second determination module, used to determine a fused feature matrix based on the financial indicator feature vectors and transaction behavior feature vectors; a judgment module, used to determine whether the fluctuation range of the fused feature matrix exceeds a preset fluctuation range threshold; a processing module, used to standardize the transaction behavior feature vectors if the fluctuation range exceeds the preset threshold, to determine a standardized transaction behavior sequence feature matrix; and a prediction module, used to predict transaction behavior based on a recurrent neural network and the standardized transaction behavior sequence feature matrix. The system comprises the following modules: a trend determination module, a third determination module for determining a dynamic risk vector based on transaction behavior trends; a fourth determination module for determining a contribution score based on the dynamic risk vector and the fused feature matrix; a second judgment module for determining whether the absolute value of the deviation between the current contribution score and the industry's enterprise contribution benchmark exceeds a preset deviation threshold; a second acquisition module for acquiring feature data of multiple enterprise customers within the industry, including annual revenue, transaction frequency, and customer loyalty, if the deviation exceeds the preset deviation threshold, and using this data as a dataset to be clustered; a fifth determination module for grouping the dataset to be clustered based on the K-means clustering algorithm to determine the grouping results of enterprise customer contribution levels; an update module for updating the contribution score based on the grouping results; a sixth determination module for determining the contribution level based on the updated contribution score; and a seventh determination module for determining the credit limit of enterprise customers based on the contribution level.

[0007] This application provides a comprehensive contribution assessment method and system based on reinforcement learning. By integrating multi-source heterogeneous data such as financial statements and transaction records of enterprise customers, and using principal component analysis to extract feature vectors of financial indicators and transaction behavior, a fused feature matrix is ​​constructed. This overcomes the limitations of traditional methods in coordinating structured financial indicator data and unstructured transaction behavior data, achieving comprehensive aggregation of multi-dimensional information of enterprise customers. Simultaneously, a judgment mechanism with a preset fluctuation threshold is introduced. When the fluctuation of the fused feature matrix exceeds the threshold, the transaction behavior feature vector is standardized, and a recurrent neural network is used to predict transaction behavior trends to generate a dynamic risk vector, replacing the traditional fixed-rule model. This constructs a dynamic assessment system that responds in real-time to changes in the enterprise's operating status, promptly capturing the liquidity risk or long-term development potential behind short-term surges in transaction volume, solving the problem of traditional assessments lagging in responding to dynamic changes in enterprises. By determining the contribution score through the fused feature matrix and dynamic risk vector, the system balances the stability of the enterprise's historical comprehensive characteristics with the dynamism of real-time risk factors, avoiding the one-sidedness of assessments caused by traditional methods relying solely on financial data, and significantly improving the accuracy of enterprise customer contribution assessment. Meanwhile, by comparing with industry contribution benchmarks, K-means clustering is used to group industry enterprise characteristic data for cases exceeding preset deviation thresholds to calibrate contribution scores, correct assessment biases, and ensure the industry suitability and objectivity of contribution level determination. Ultimately, credit limits are determined based on accurate contribution levels, achieving optimized allocation of bank credit resources under the premise of dynamic risk control, improving capital allocation efficiency and risk control capabilities, and accurately matching the actual operational needs of enterprises, thus providing strong support for the development of the real economy. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments or conventional technologies of this disclosure, the accompanying drawings used in the description of the embodiments or conventional technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating an exemplary embodiment of the reinforcement learning-based comprehensive contribution evaluation method provided in this application.

[0010] Figure 2 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application.

[0011] Figure 3 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application.

[0012] Figure 4 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application.

[0013] Figure 5 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application.

[0014] Figure 6 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application.

[0015] Figure 7 This is a flowchart illustrating a reinforcement learning-based comprehensive contribution evaluation method provided in another exemplary embodiment of this application. Detailed Implementation

[0016] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are set forth to give a full understanding of embodiments of this disclosure.

[0017] The terms “a,” “one,” and “the” are used to indicate the existence of one or more elements / components / etc.; the terms “including” and “having” are used to indicate an open-ended inclusion and that other elements / components / etc. may exist in addition to those listed. The terms “first” and “second” are used only as markers and are not a limitation on the number of objects.

[0018] Currently, corporate lending is a crucial pillar of the financial industry, directly impacting banks' capital allocation efficiency and risk control capabilities. As the primary recipients of bank credit, assessing the comprehensive contribution of corporate clients is not only a core element of credit decision-making but also key to optimizing resource allocation and enhancing market competitiveness. Accurately assessing the comprehensive contribution of corporate clients helps banks rationally allocate credit resources under controllable risk, thereby supporting the development of the real economy. However, traditional comprehensive contribution assessment methods typically rely on static indicator systems and fixed rule models, failing to adapt to the complex and ever-changing operating environments of corporate clients. Furthermore, when faced with multi-dimensional data from corporate clients, they cannot comprehensively capture the true value of their contributions, making it difficult for traditional methods to simultaneously meet the requirements of dynamism and accuracy. For example, when a company's financial data and transaction behavior change rapidly, banks struggle to determine its comprehensive contribution level in real time. Similarly, a manufacturing company may increase its loan demand due to a surge in orders, but its industry risk may also rise. Traditional contribution assessment methods might approve or reject loans solely based on financial data, ignoring short-term solvency or long-term development potential revealed by transaction behavior, leading to inaccurate or delayed credit decisions.

[0019] One of the core challenges in assessing the comprehensive contribution of corporate clients to banks' corporate lending lies in constructing a dynamic assessment framework that integrates multi-dimensional information. For example, a corporate client's state space encompasses multiple aspects, including financial indicators, transaction behavior, and risk characteristics. The heterogeneity and high dynamism of this information make it difficult for traditional assessment models to effectively integrate it. Because the heterogeneity of corporate client data manifests in the differences in format, time span, and importance between different types of data (such as structured financial statements and unstructured transaction records), processing this data requires a system with strong information fusion capabilities. The complexity of information fusion further raises another technical challenge: how to update the contribution assessment strategy in real time within a dynamic environment to adapt to changes in the corporate client's state. For instance, a surge in a corporate client's transaction volume in a short period may reflect rapid growth in market demand, but it could also conceal potential liquidity risks. The system needs to quickly weigh these factors across multi-dimensional data and adjust the contribution assessment strategy accordingly.

[0020] Therefore, how to effectively integrate multi-dimensional and heterogeneous corporate customer data, and based on this, to accurately and dynamically adjust contribution assessment strategies, has become a technical problem that banks need to solve in the comprehensive contribution assessment of corporate customers with corporate credit lines.

[0021] This disclosure provides a comprehensive contribution evaluation method based on reinforcement learning, such as... Figure 1 The method shown is a reinforcement learning-based comprehensive contribution evaluation method. This method may include the following steps: Step S110: Obtain multi-source heterogeneous data from the enterprise customer's financial statements and transaction records. The multi-source heterogeneous data includes financial indicator values ​​and transaction behavior sequences. Step S120: Determine the customer status dataset based on financial indicator values ​​and transaction behavior sequences; Step S130: Based on the principal component analysis algorithm, extract the key feature vectors from the customer status dataset. The key feature vectors include financial indicator feature vectors and transaction behavior feature vectors. Step S140: Determine the fusion feature matrix based on the financial indicator feature vector and the transaction behavior feature vector; Step S150: Determine whether the fluctuation amplitude of the fused feature matrix exceeds the preset fluctuation amplitude threshold; Step S160: If it is determined that the fluctuation range exceeds the preset threshold, the trading behavior feature vector is standardized to determine the standardized trading behavior sequence feature matrix. Step S170: Based on the recurrent neural network, predict the trend of trading behavior according to the feature matrix of the standardized trading behavior sequence; Step S180: Determine the dynamic risk vector based on trading behavior trends; Step S190: Determine the contribution score based on the dynamic risk vector and the fusion feature matrix; Step S191: Determine whether the absolute value of the deviation between the current contribution score and the industry's enterprise contribution benchmark exceeds the preset deviation threshold; Step S192: If the deviation exceeds the preset threshold, acquire the feature data of multiple enterprise customers in the industry. The feature data includes annual income, transaction frequency and customer loyalty, and is used as the dataset to be clustered. Step S193: Based on the K-means clustering algorithm, group the dataset to be clustered to determine the grouping results of the enterprise customer contribution level; Step S194: Update the contribution score based on the grouping results; Step S195: Determine the contribution level based on the updated contribution score; Step S196: Determine the credit limit for corporate clients based on their contribution level.

[0022] According to the reinforcement learning-based comprehensive contribution evaluation method provided in this disclosure, the method can obtain multi-source heterogeneous data from enterprise customers' financial statements and transaction records, including financial indicator values ​​and transaction behavior sequences; determine a customer status dataset based on the financial indicator values ​​and transaction behavior sequences; extract key feature vectors from the customer status dataset based on principal component analysis, including financial indicator feature vectors and transaction behavior feature vectors; determine a fusion feature matrix based on the financial indicator feature vectors and transaction behavior feature vectors; determine whether the fluctuation range of the fusion feature matrix exceeds a preset fluctuation range threshold; if it exceeds the preset fluctuation range threshold, standardize the transaction behavior feature vectors to determine the standardized transaction behavior sequence feature matrix. The algorithm employs a recurrent neural network to predict transaction behavior trends based on a standardized transaction behavior sequence feature matrix. It then determines a dynamic risk vector based on these trends, and a contribution score based on the dynamic risk vector and the fused feature matrix. Next, it checks if the absolute value of the deviation between the current contribution score and the industry's benchmark contribution exceeds a preset deviation threshold. If the threshold is exceeded, it acquires feature data from multiple industry clients, including annual revenue, transaction frequency, and customer loyalty, and uses this data as a clustering dataset. Finally, it groups the dataset based on the K-means clustering algorithm to determine the grouping results for client contribution levels. The contribution score is updated based on the grouping results. The contribution level is determined based on the updated contribution score, and the credit limit for each client is determined based on the contribution level.

[0023] This method integrates multi-source heterogeneous data from corporate clients, including financial statements and transaction records. Principal component analysis (PCA) is used to extract feature vectors from financial indicators and transaction behavior, constructing a fused feature matrix. This overcomes the limitations of traditional methods in coordinating structured financial data and unstructured transaction behavior data, achieving comprehensive aggregation of multi-dimensional information about corporate clients. Simultaneously, a pre-set fluctuation threshold judgment mechanism is introduced. When the fluctuation of the fused feature matrix exceeds the threshold, the transaction behavior feature vector is standardized, and a recurrent neural network is used to predict transaction behavior trends to generate a dynamic risk vector. This replaces the traditional fixed-rule model, constructing a dynamic evaluation system that responds in real-time to changes in the company's operating status. It can promptly capture the liquidity risk or long-term development potential behind short-term surges in transaction volume, solving the problem of traditional evaluations lagging in responding to dynamic changes in the company. By determining the contribution score through the fused feature matrix and dynamic risk vector, the method balances the stability of the company's historical comprehensive characteristics with the dynamism of real-time risk factors, avoiding the one-sidedness of traditional methods that rely solely on financial data, and significantly improving the accuracy of corporate client contribution evaluation. Meanwhile, by comparing with industry contribution benchmarks, K-means clustering is used to group industry enterprise characteristic data for cases exceeding preset deviation thresholds to calibrate contribution scores, correct assessment biases, and ensure the industry suitability and objectivity of contribution level determination. Ultimately, credit limits are determined based on accurate contribution levels, achieving optimized allocation of bank credit resources under the premise of dynamic risk control, improving capital allocation efficiency and risk control capabilities, and accurately matching the actual operational needs of enterprises, thus providing strong support for the development of the real economy.

[0024] The steps of the reinforcement learning-based comprehensive contribution evaluation method provided in this disclosure are described in detail below: In one embodiment of this disclosure, step S110 involves acquiring multi-source heterogeneous data from the enterprise customer's financial statements and transaction records. This multi-source heterogeneous data includes financial indicator values ​​and transaction behavior sequences. Specifically, the financial indicator values ​​are structured data, stored in a fixed field + numerical format, reflecting the enterprise's solvency, profitability, and operational capabilities. Financial indicator values ​​may include total assets, net assets, and current liabilities from the balance sheet; annual revenue and net profit from the income statement; and operating cash flow from the cash flow statement. For example, the enterprise's SAP (System Applications and Products) system can be called via API (Application Programming Interface) to extract the balance sheet data for the first quarter of 2025, with the following financial indicator values: total assets of 50 million yuan, current liabilities of 20 million yuan, and net assets of 30 million yuan. The transaction behavior sequences are unstructured data, stored in a timeline + transaction attribute format, reflecting the real-time cooperation dynamics between the enterprise and the bank. The transaction sequence can include daily transaction amount (income / expenditure), transaction frequency (daily / monthly transaction count), transaction type (investment / consumption / transfer), and transaction interval, exhibiting dynamic and continuous characteristics. For example, by calling the bank's transaction system interface, the daily transaction records of the company in March 2025 can be extracted. The transaction sequence is as follows: March 1, income 500,000 yuan, expenditure 300,000 yuan; March 2, income 450,000 yuan, expenditure 350,000 yuan... March 31, income 520,000 yuan, expenditure 280,000 yuan.

[0025] In one embodiment of this disclosure, in step S120, a customer status dataset is determined based on financial indicator values ​​and a transaction behavior sequence. Specifically, for example, a company's financial indicator values ​​are: current assets of RMB 25 million, current liabilities of RMB 20 million, and net assets of RMB 30 million in March 2025. The transaction behavior sequence is: 60 valid transactions within 30 days in March 2025, with daily transaction amounts fluctuating between RMB 128,000 and RMB 3.2 million. Therefore, the company's customer status dataset is: current assets of RMB 25 million, current liabilities of RMB 20 million, and net assets of RMB 30 million in March 2025. Furthermore, there are 60 valid transactions within 30 days in March 2025, with daily transaction amounts fluctuating between RMB 128,000 and RMB 3.2 million.

[0026] In one embodiment of this disclosure, after determining the customer status dataset based on financial indicator values ​​and transaction behavior sequences in step S120, the following steps are further included: Figure 2 As shown, the specific content is as follows: Step S210: Determine customer status characteristics based on the customer status dataset. Customer status characteristics include liquidity ratio, transaction frequency, and net assets. Step S220: Based on the K-means clustering algorithm, use the current ratio, transaction frequency, and net assets as input features to cluster corporate customers; Step S230: Determine customers with different risk levels based on the clustering results. Risk level customers include low-risk customers, medium-risk customers, and high-risk customers.

[0027] Specifically, taking three companies—automobile manufacturer C01, machinery manufacturer C02, and textile company C03—as examples, customer status characteristics are extracted from various customer status datasets: Company C01 has current assets of 50 million yuan, current liabilities of 20 million yuan, a current ratio of 2.50, 120 transactions in 30 days (transaction frequency of 4.0 times / day), and net assets of 80 million yuan. Company C02 has current assets of 24 million yuan, current liabilities of 20 million yuan, a current ratio of 1.20, 60 transactions in 30 days (transaction frequency of 2.0 times / day), and net assets of 40 million yuan. Company C03 has current assets of 5 million yuan, current liabilities of 10 million yuan, a current ratio of 0.50, 15 transactions in 30 days (transaction frequency of 0.5 times / day), and net assets of 10 million yuan. The three customer status characteristics were standardized, and the following values ​​were calculated: mean current ratio 1.40, standard deviation 0.87; mean transaction frequency 2.17, standard deviation 1.51; and mean net assets 43.33 million yuan, standard deviation 30.55 million yuan. For example, the standardized current ratio of enterprise C02 is approximately (1.20-1.40) / 0.87≈-0.23, transaction frequency is approximately (2.0-2.17) / 1.51≈-0.11, and net assets are approximately (40 million-43.33) / 30.55≈-0.11. Subsequently, K=3 can be set, and based on the K-means clustering algorithm, using the three companies as initial centroids, similarity is calculated using Euclidean distance and the centroids are iteratively updated. After two iterations, convergence occurs. Company C01 is assigned to cluster 1 (standardized feature mean [1.26, 1.21, 1.20]), company C02 to cluster 2 (standardized feature mean [-0.23, -0.11, -0.11]), and company C03 to cluster 3 (standardized feature mean [-1.03, -1.10, -1.09]). Company C01 in cluster 1 is a low-risk customer due to its high current ratio (strong short-term solvency), high transaction frequency (active cooperation), and high net assets (strong risk resistance). Company C02 in cluster 2 has moderate characteristics (meeting solvency standards, stable cooperation, and average strength) and is a medium-risk customer. Company C03 in Cluster 3 is classified as a high-risk client due to its low current ratio (weak short-term solvency), low transaction frequency (inactive cooperation), and low net assets (weak risk resistance).

[0028] In the aforementioned method, on the one hand, by combining three characteristics strongly correlated with credit risk—current ratio, transaction frequency, and net assets—the specificity of customer status assessment is ensured, avoiding biases caused by feature redundancy or missing key information in traditional assessments. On the other hand, the K-means clustering algorithm breaks the subjectivity of traditional manual risk classification, making the risk classification results more objective. This assigns clear risk labels to corporate customers, helping banks accurately identify low-risk, high-quality customers and focus on managing high-risk customers in their credit business. This provides a data-driven decision-making basis for balancing resource optimization and risk control.

[0029] In one embodiment of this disclosure, step S130 involves extracting key feature vectors from the customer status dataset based on a principal component analysis algorithm. These key feature vectors include financial indicator feature vectors and transaction behavior feature vectors. The process also includes the following steps: Figure 3 As shown, the specific content is as follows: Step S310: Standardize the financial indicators to determine standardized financial indicators; Step S320: Based on the TF-IDF algorithm, determine the transaction behavior sequence vector according to the transaction behavior sequence; Step S330: Based on the principal component analysis algorithm, calculate the covariance matrix of standardized financial indicators and transaction behavior sequence vectors; Step S340: Perform eigenvalue decomposition on the covariance matrix of standardized financial indicators to determine the eigenvectors of the financial indicators; Step S350: Perform eigenvalue decomposition on the covariance matrix of the trading behavior vector to determine the trading behavior feature vector.

[0030] Specifically, taking three financial indicators of three companies (C01 Automobile Manufacturing, C02 Machinery Manufacturing, and C03 Textile Industry) as examples, the financial indicators are standardized to determine standardized financial indicators. The standardized financial indicators for company C01 are: standardized current assets of 1.17, standardized current liabilities of 0.58, and standardized net assets of 1.20. The standardized financial indicators for company C02 are: standardized current assets of 0.11, standardized current liabilities of 0.58, and standardized net assets of -0.11. The standardized financial indicators for company C03 are: standardized current assets of 1.06, standardized current liabilities of -1.15, and standardized net assets of -1.09. It should be noted that the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm can be used for text analysis. Here, "transaction type" is considered as "vocabulary," and "transaction behavior sequence" is considered as "document." The transaction behavior sequence can be vectorized using the TF-IDF algorithm to generate a transaction behavior sequence vector. For example, a company's average monthly transaction amount is [5, 6, 4, ... ], transaction frequency is [10, 12, 8, ... ], and transaction type percentage is [0.4, 0.3, 0.2, ... ]. The TF-IDF algorithm is used to convert the transaction behavior sequence into a transaction behavior sequence vector with a dimension of 12 × 3 = 36. Then, principal component analysis is applied to calculate the covariance matrix of the standardized financial indicators and the transaction behavior sequence vector. For example, the covariance matrix of the standardized financial indicators is a 3 × 3 matrix, and the covariance matrix of the transaction behavior sequence vector is a 36 × 36 matrix. Eigenvalue decomposition yields the eigenvectors of the financial indicators and the eigenvectors of the transaction behavior, retaining the first k principal components (assuming k = 2). This results in a cumulative explained variance of 85%. For example, the first principal component eigenvector of the standardized financial indicators is [0.6, 0.5, 0.3], explaining 50% of the variance, and the second principal component is [0.2, -0.7, 0.4], explaining 35% of the variance. The standardized financial indicators and TF-IDF vectors are projected onto the principal component space to obtain the dimensionality-reduced financial standard feature vectors. The transaction behavior vectors and TF-IDF vectors are then projected onto the principal component space to obtain the transaction behavior feature vectors. For example, a corporate client's financial indicator feature vector is [-1.2, 0.8], and its transaction behavior feature vector is [2.1, -0.5].

[0031] In the above method, standardization eliminates the dimensional interference of financial indicators, and the TF-IDF algorithm transforms unstructured transaction behavior sequences into quantitative transaction behavior sequence vectors, solving the "incomparability" problem of multi-source heterogeneous features. Furthermore, principal component analysis is used to perform eigenvalue decomposition on the covariance matrix, selecting the key feature vectors with the strongest explanatory power (such as the comprehensive principal components of financial indicators and the unique transaction types of transaction behavior). This effectively reduces data dimensionality and eliminates redundant information (such as the high correlation between financial indicators), laying a feature foundation for accurate comprehensive contribution assessment of corporate clients.

[0032] In one embodiment of this disclosure, in step S140, a fusion feature matrix is ​​determined based on the financial indicator feature vector and the transaction behavior feature vector. Specifically, for example, a corporate customer's financial indicator feature vector is [-1.2, 0.8], and its transaction behavior feature vector is [2.1, -0.5]. The two are then concatenated to obtain a fusion feature matrix of [-1.2, 0.8, 2.1, -0.5].

[0033] In one embodiment of this disclosure, after determining the fused feature matrix based on the financial indicator feature vector and the transaction behavior feature vector in step S140, the following steps are also included: Figure 4 As shown, the specific content is as follows: Step S410: Standardize the fused feature matrix to determine the standardized fused feature matrix; Step S420: Determine whether the feature dimension of the standardized fusion feature matrix exceeds a preset dimension threshold; Step S430: If the dimensionality threshold is exceeded, the dimensionality of the standardized fusion feature matrix is ​​reduced based on the principal component analysis algorithm to determine the dimensionality-reduced fusion feature matrix.

[0034] Specifically, taking machinery manufacturing company D01, e-commerce company D02, and textile company D03 as examples, the fusion feature matrices of the three companies are standardized, resulting in: the standardized fusion matrix of company D01 is [0.95, 0.72, 0.35, 0.86, 0.87]; the standardized fusion matrix of company D02 is [0.09, 0.56, 0.88, 0.28, 0.20]; and the standardized fusion matrix of company D03 is [-1.04, -1.28, -1.24, -1.14, -1.13]. Next, it is determined whether the feature dimension of the standardized fusion feature matrix exceeds a preset dimension threshold. It should be noted that the preset dimension threshold is set based on the processing power and evaluation accuracy requirements of the machine learning algorithm. When the dimension is too high, it will lead to a significant increase in training parameters, thereby increasing resource consumption. The number of columns in the standardized fusion feature matrix can be used as the feature dimension for this determination. For example, the standardized fusion feature matrices of companies D01, D02, and D03 all have a feature dimension of 5 (5 columns of features), while the preset dimension threshold is 4. Since 5 > 4, it is determined that the preset dimension threshold is exceeded. Therefore, based on the principal component analysis algorithm, a linear transformation is used to map the high-dimensional features to a low-dimensional space, while preserving the original data to the maximum extent. For example, calculate the covariance matrix of the standardized fusion feature matrix (e.g., the covariance between net assets and average monthly transaction amount is ≈0.98), and perform eigenvalue decomposition to obtain 5 eigenvalues ​​(e.g., principal component 1 λ=3.21, principal component 2 λ=1.04) and their corresponding explained variance proportions. Select the first two principal components (which explain a cumulative 85% of the variance), and their eigenvectors are [0.45, 0.42, 0.40, 0.48, 0.43] and [-0.12, -0.08, 0.65, -0.15, 0.73], respectively. Multiply the standardized matrix by these two eigenvectors to obtain the dimensionality-reduced 3×2 fusion feature matrix, where principal component 1 of D01 is approximately 2.15 and principal component 2 is approximately 0.62, principal component 1 of D02 is approximately 0.83 and principal component 2 is approximately 0.59, and principal component 1 of D03 is approximately -2.98 and principal component 2 is approximately -1.21. Therefore, the dimensionality-reduced fusion feature matrix is: Enterprise D01 [2.15, 0.62], Enterprise D02 [0.83, 0.59], Enterprise D03 [-2.98, -1.21].

[0035] In the aforementioned method, a pre-defined dimensionality threshold judgment mechanism accurately identifies the dimensionality overshoot problem of high-dimensional standardized fusion feature matrices. This avoids the potential for a surge in machine learning model training parameters, increased resource consumption, and overfitting risks caused by high-dimensional data, thus meeting the efficiency and stability requirements of banks' corporate credit assessment. Dimensionality reduction based on principal component analysis (PCA) preserves core assessment information from the original data to the greatest extent possible, such as the comprehensive strength of the enterprise (e.g., the scale correlation reflected by the covariance of net assets and average monthly transaction amount ≈ 0.98) and the uniqueness of transaction patterns (e.g., the weighting characteristics of transaction frequency and type proportions). Simultaneously, it significantly reduces the dimensionality of the feature matrix, providing low-dimensional, high-information-density input data for subsequent recurrent neural network risk prediction and linear regression weighted scoring. Ultimately, this improves the operational efficiency and accuracy of the bank's comprehensive contribution assessment process for corporate clients, providing reliable data support for credit decisions.

[0036] In one embodiment of this disclosure, in steps S150-S180, it is determined whether the fluctuation amplitude of the fused feature matrix exceeds a preset fluctuation amplitude threshold; if it exceeds the preset fluctuation amplitude threshold, the trading behavior feature vector is standardized to determine a standardized trading behavior sequence feature matrix; based on a recurrent neural network, the trading behavior trend is predicted according to the standardized trading behavior sequence feature matrix; and a dynamic risk vector is determined according to the trading behavior trend. Specifically, it should be noted that historical time-series data (such as the past 30 days) of the trading behavior features in the fused feature matrix are selected, the standard deviation of each trading behavior feature is calculated, and the mean of the standard deviations of each feature or the standard deviation of key features (such as daily trading amount) is used as the fluctuation amplitude of the fused feature matrix. For example, taking machinery manufacturing enterprise D01 (whose transaction behavior features in the fusion feature matrix include daily transaction amount (ten thousand yuan), transaction frequency (times / day), and transaction interval (day)) as an example, the transaction behavior data of enterprise D01 over the past 30 days are extracted as follows: Daily transaction amount: [115,120,98,150,102,85,130,...] (30 data points in total, mean ≈ 1.1 million yuan); Transaction frequency: [2,3,1,4,2,1,3,...] (mean ≈ 2.2 times / day); Transaction interval: [1,0.5,2,0.3,1.5,2.5,0.8,...] (mean ≈ 1. (2 days). The standard deviations of each transaction behavior characteristic are calculated as follows: daily transaction amount standard deviation σ1 = 225,000 yuan; transaction frequency standard deviation σ2 ≈ 1.1 times / day; transaction interval standard deviation σ3 ≈ 0.7 days. At this point, the daily transaction amount standard deviation can be taken as the core volatility range (since transaction amount is directly related to liquidity risk, priority is given to amount volatility), meaning the volatility range of the enterprise D01 fusion feature matrix is ​​225,000 yuan. The bank's preset daily transaction amount volatility threshold is 200,000 yuan (based on historical manufacturing data). Since 225,000 > 200,000, the volatility range of the fusion feature matrix is ​​determined to exceed the preset volatility threshold.

[0037] If the fluctuation exceeds a preset threshold, the transaction behavior feature vector is standardized. The standardized data for 30 days is arranged chronologically, resulting in a standardized transaction behavior sequence feature matrix of dimension (30,3). A Recurrent Neural Network (RNN) model is constructed using an LSTM (Long Short-Term Memory) architecture to predict the standardized transaction behavior sequence feature matrix of company D01. The model parameters are: input dimension (30,3), 2-layer LSTM, Adam optimizer, learning rate 0.001, MSE (Mean Square Error) loss function, and model convergence after 50 iterations (validation set loss < 0.05). The prediction result outputs the transaction behavior trend for the next 7 days. The transaction behavior trend can include predicted trends in daily transaction amount, transaction frequency, and transaction interval. For example, the following prediction results can be generated: Day 1, daily transaction amount = 1.35 million yuan, transaction frequency = 4.5 times / day, transaction interval = 0.2 days; Day 2, daily transaction amount = 1.42 million yuan, transaction frequency = 5.0 times / day, transaction interval = 0.1 days; Day 3, daily transaction amount = 1.5 million yuan, transaction frequency = 5.2 times / day, transaction interval = 0.1 days; Day 4, daily transaction amount = 900,000 yuan, transaction frequency = 1.0 time / day, transaction interval = 2.5 days; Day 5, daily transaction amount = 850,000 yuan, transaction frequency = 0.8 times / day, transaction interval = 3.0 days; Day 6, daily transaction amount = 820,000 yuan, transaction frequency = 0.5 times / day, transaction interval = 3.2 days; Day 7, daily transaction amount = 780,000 yuan, transaction frequency = 0.3 times / day, transaction interval = 3.5 days. It can be seen that the daily transaction amount, transaction frequency, and transaction intervals increased over the first three days, reflecting a sharp increase in short-term trading activity. Conversely, the daily transaction amount decreased over the following four days, the transaction frequency decreased, and the transaction intervals increased, reflecting a sharp decrease in trading activity.

[0038] Understandably, the predicted trading behavior trend needs to be transformed into a quantifiable risk indicator that can be directly used for subsequent contribution assessment, i.e., a dynamic risk vector. For example, arranged according to the risk score for the next 7 days, the dynamic risk vector is [0.42, 0.48, 0.50, 0.75, 0.82, 0.88, 0.92], where the risk score for the last 4 days is consistently higher than 0.75, clearly indicating a significant increase in short-term risk.

[0039] In the aforementioned method, on the one hand, the standard deviation method accurately captures abnormal fluctuations in transaction behavior within the fused feature matrix and promptly triggers risk prediction, thus addressing the pain point of traditional contribution assessment's inability to identify short-term transaction risks in real time and significantly improving resource utilization efficiency. On the other hand, the recurrent neural network based on the LSTM architecture effectively captures the long-term and short-term temporal dependencies of transaction behavior characteristics, such as accurately predicting the trend of a sudden increase in transaction activity in the first three days of a company's D01 and the subsequent trend of a decrease in transaction activity, overcoming the limitation of traditional linear models in handling transaction temporal characteristics. Simultaneously, the transaction behavior trend is transformed into a quantified dynamic risk vector, clearly indicating short-term risk changes. This not only solves the problems of lagging and difficult-to-quantify traditional risk assessment results but also provides real-time and actionable risk basis for subsequent weighted contribution score calculation and credit resource allocation. Ultimately, this helps banks dynamically avoid liquidity risks in corporate lending, significantly improving the timeliness, accuracy, and reliability of comprehensive contribution assessment.

[0040] In one embodiment of this disclosure, step S190, determining the contribution score, may include the following steps: obtaining the weight value corresponding to the dynamic risk vector; performing a weighted operation on the fused feature matrix and the dynamic risk vector based on the weight value corresponding to the dynamic risk vector, and determining the weighted operation result as the contribution score. Specifically, the contribution score can be determined by the following formula: Where S is the contribution score. For the weight vector, As a dynamic risk vector, This is the fused feature matrix.

[0041] For example, a dynamic risk vector can include credit score, transaction frequency, and the proportion of abnormal behavior. We obtain the weights for credit score (0.4), transaction frequency (0.35), and abnormal behavior proportion (0.25), respectively. Therefore, the weight vector corresponding to the dynamic risk vector is [0.4, 0.35, 0.25]. The dynamic risk vector is [0.85, 0.6, 0.1], and the fused feature matrix is ​​[0.9, 0.88, 0.7]. Substituting these into the above formula, we get: Contribution score S = (0.4 × 0.85 × 0.9) + (0.35 × 0.6 × 0.8) + (0.25 × 0.1 × 0.7) = 0.4925.

[0042] In one embodiment of this disclosure, in steps S191 to S194, it is determined whether the absolute value of the deviation between the current contribution score and the contribution benchmark of enterprises in the industry exceeds a preset deviation threshold; if it is determined that it exceeds the preset deviation threshold, feature data of multiple enterprise customers in the industry are obtained, including annual income, transaction frequency and customer loyalty, and used as a dataset to be clustered; the dataset to be clustered is grouped based on the K-means clustering algorithm to determine the grouping results of enterprise customer contribution levels; the contribution score is updated according to the grouping results.

[0043] Updating the contribution score based on the grouping results includes the following steps: Figure 5 As shown, the details are as follows: Step S510: Determine the average contribution score within the corresponding enterprise customer contribution level grouping results based on the enterprise customer contribution level grouping results; Step S520: Determine whether the current contribution score does not exceed the average contribution score within the corresponding enterprise customer contribution level grouping results; Step S530: If it is determined that the average contribution score has not been exceeded, the current contribution score is updated to the average contribution score of the corresponding enterprise customer contribution level grouping results; Step S540: Determine the contribution level based on the updated contribution score. The contribution levels include low-level contribution, medium-level contribution, and high-level contribution.

[0044] Specifically, it should be noted that when setting industry-specific benchmarks, it is necessary to select enterprise clients in the same industry, of similar size, and with the same number of years of cooperation as the current enterprise, and calculate the average weighted contribution score of these clients over the past 12 months. This ensures the industry relevance of the benchmark. The preset deviation threshold is determined based on the standard deviation of the contribution scores of enterprises within the industry (usually one standard deviation to avoid normal fluctuations triggering calibration). This threshold is used to determine whether the current enterprise's contribution score is abnormally deviating. If the absolute value of the deviation exceeds the preset deviation threshold, it indicates that the individual score differs too much from the level of similar enterprises in the industry, and calibration is required through subsequent steps. Taking machinery manufacturing enterprise D01 as an example, 30 "medium-sized machinery manufacturing enterprises (500-1000 employees, cooperating with banks for 3-5 years)" in China were selected, and the average weighted contribution score of these enterprises over the past 12 months was calculated. The result is 0.6, meaning that the industry-specific contribution benchmark is 0.6. The deviation between the current contribution score (0.4925) and the industry benchmark (0.6) is 0.4925 - 0.6 = -0.1075, and the absolute value of the deviation is |-0.1075| = 0.1075. Setting the preset deviation threshold to 0.1, since 0.1075 > 0.1, it is determined that the absolute value of the deviation between the current contribution score and the industry benchmark exceeds the preset deviation threshold. Then, feature data of 30 medium-sized machinery manufacturing enterprises (500-1000 employees, 3-5 years of cooperation) in China are extracted from the bank's corporate customer database. This feature data may include annual revenue, transaction frequency, and customer loyalty, and is used as the dataset to be clustered. For example, enterprise D06 has an annual revenue of 52 million yuan, a transaction frequency of 75 times / year, and a customer loyalty score of 88 (rating, 0-100). The dataset to be clustered may include the annual revenue, transaction frequency, and customer loyalty of 30 companies, and organize the above data into a structured dataset of 30 rows × 3 columns.

[0045] K-means clustering (K=3) was performed on the clustering dataset (feature data of 30 medium-sized machinery manufacturing enterprises). This resulted in three groups: a high-contribution group (containing 8 enterprises): standardized feature mean of "annual revenue 1.2, transaction frequency 1.5, customer loyalty 1.3" (corresponding to the original data: annual revenue ≥ 55 million yuan, transaction frequency ≥ 70 times / year, customer loyalty ≥ 85 points); a medium-contribution group (containing 15 enterprises): standardized feature mean of "annual revenue 0.1, transaction frequency 0.2, customer loyalty 0.1" (corresponding to the original data: annual revenue 35-55 million yuan, transaction frequency 50-70 times / year, customer loyalty 75-85 points); and a low-contribution group (containing 7 enterprises): standardized feature mean of "annual revenue -1.1, transaction frequency -1.0, customer loyalty -0.9" (corresponding to the original data: annual revenue ≤ 35 million yuan, transaction frequency ≤ 50 times / year, customer loyalty ≤ 75 points). Among them, the standardized characteristics of enterprise D01 are annual revenue of 0.32, transaction frequency of 0.17, and customer loyalty of 0.4. Its centroid distance is closest to that of the medium contribution level group (distance ≈ 0.21), therefore enterprise D01 is classified into the medium contribution level group. Based on the grouping results of step S630, the average contribution score of each group is calculated as follows: High contribution level grouping: The weighted contribution scores of the 8 companies are as follows: [0.72,0.68,0.75,0.81,0.69,0.73,0.77,0.66], average contribution score = 0.73.

[0046] Contribution Level Grouping: The weighted contribution scores of the 15 companies are as follows: [0.4925,0.55,0.52,0.48,0.58,0.51,0.47,0.53,0.56,0.46,0.54,0.50,0.45,0.57,0.49], average contribution score = 0.51.

[0047] Low contribution level grouping: The weighted contribution scores of the 7 companies are as follows: [0.28,0.30,0.25,0.22,0.29,0.26,0.24], average contribution score = 0.26.

[0048] Among them, the average contribution score of the group in which enterprise D01 belongs is 0.51.

[0049] The current contribution score of enterprise D01 is 0.4925, and the average contribution score within its medium contribution level group is 0.51. Since 0.4925 < 0.51, the current contribution score does not exceed the average contribution score within the corresponding enterprise customer contribution level group. Therefore, enterprise D01's current contribution score is updated from 0.4925 to 0.51, which is the average contribution score within the corresponding enterprise customer contribution level group. The updated contribution score 0.51 ∈ (0.3, 0.6], therefore, enterprise D01's final contribution level is determined to be medium contribution. If the updated score is 0.62 (e.g., if an enterprise's original score was 0.58, and the average score after being grouped into the high contribution group was 0.62, the updated score is 0.62), then 0.62 ∈ (0.6, 1], and the final contribution level is determined to be high contribution.

[0050] In the aforementioned method, on the one hand, by judging the contribution benchmark of enterprises within the industry and the preset deviation threshold, abnormal deviations between individual contribution scores and the level of similar enterprises in the industry are accurately identified, solving the assessment bias problem caused by only looking at individual data and ignoring industry characteristics in traditional assessments. On the other hand, by grouping similar enterprises within the industry based on the K-means clustering algorithm and calibrating individual scores by combining the average contribution scores of the groups, the subjectivity of manually setting industry benchmarks is avoided, and the score anomalies caused by short-term operational fluctuations of individuals are corrected by the mean substitution. This ensures that the contribution scores can reflect the long-term value of enterprises and the consistency with the level of similar enterprises in the industry. It is both in line with the actual operating capabilities of enterprises and in line with the overall level of the industry. This provides a fair and accurate rating basis for subsequent assessment strategy updates and credit resource allocation, effectively avoiding resource misallocation caused by abnormal individual data or industry perception biases, and significantly improving the objectivity, industry adaptability, and decision support capabilities of the bank's comprehensive contribution assessment for corporate clients.

[0051] In one embodiment of this disclosure, step S195, determining the contribution level based on the updated contribution score, whereby the contribution levels include low-level contribution, medium-level contribution, and high-level contribution, includes the following steps: determining whether the contribution score falls within a preset low-level contribution score range; if it falls within the preset low-level contribution score range, then determining the contribution level of the current contribution score as low-level contribution; if it does not fall within the preset low-level contribution score range, then determining whether the contribution score falls within a preset medium-level contribution score range; if it falls within the preset medium-level contribution score range, then determining the contribution level of the current contribution score as medium-level contribution; if it does not fall within the preset medium-level contribution score range, then determining whether the contribution score falls within a preset high-level contribution score range; if it falls within the preset high-level contribution score range, then determining the contribution level of the current contribution score as high-level contribution.

[0052] Specifically, taking three companies from different industries—textile company D03, machinery manufacturing company D01, and e-commerce company D04—as examples, we first determine whether each company's weighted contribution score falls within the preset low-level contribution score range [0, 0.3]. Company D03's weighted contribution score is 0.25, falling within this range, and its contribution level is determined to be low-level contribution. The weighted contribution score of company D01 is 0.4925, and the weighted contribution score of company D04 is 0.75. Neither falls within the preset low-level contribution score range. Therefore, it is further determined whether they belong to the preset medium-level contribution score range (0.3, 0.6). Since company D01's weighted contribution score of 0.4925 falls within the preset medium-level contribution score range, its contribution level is determined to be medium-level. Since company D04's weighted contribution score of 0.75 does not fall within the preset medium-level contribution score range, it is determined whether they belong to the preset high-level contribution score range (0.6, 1). Since company D04's weighted contribution score of 0.75 falls within the preset high-level contribution score range, its contribution level is determined to be high-level.

[0053] The above method transforms the contribution score into discrete contribution levels of "low / medium / high," which solves the pain points of traditional contribution assessment methods, such as the difficulty in implementing scores and the lack of basis for decision-making. This allows the contribution assessment results to be directly linked to the bank's corporate credit resource allocation strategy (such as limiting credit for low-level contributions and prioritizing allocation for high-level contributions), which can help banks achieve refined stratification of corporate customer contributions and accurate matching of credit resources.

[0054] In one embodiment of this disclosure, step S196, determining the credit limit of the enterprise customer based on the contribution level, includes the following steps: Figure 6 As shown, the specific content is as follows: Step S610: Generate an evaluation strategy signal based on the contribution level; Step S620: Calculate the signal strength value of the evaluation strategy signal based on the support vector machine algorithm; Step S630: Determine whether the signal strength value exceeds the preset strength threshold; Step S640: If it is determined that the preset strength threshold is exceeded, the credit resource allocation ratio is determined according to the contribution level; Step S650: Determine the credit limit for corporate clients based on the credit resource allocation ratio.

[0055] Specifically, taking e-commerce company D04, whose final contribution level is high-level contribution, as an example, an evaluation strategy signal is generated based on D04's high-level contribution. This signal is presented in a structured JSON format and includes "Company ID: D04, Contribution Level: High-level Contribution, Strategy Instruction List: ["Prioritize Credit Resources", "Open Fast Approval Channel", "Provide Customized Financial Services", "Assign a Dedicated Account Manager"], and Generation Time: 2025-XX-XX 10:00:00. Next, the signal strength value of the evaluation strategy signal is calculated based on the support vector machine algorithm. First, the evaluation strategy signal is converted into a feature vector (contribution level quantification value = 1, industry matching degree = 0.9 (highly matched with the average level of bank e-commerce clients), historical cooperation stability = 0.95 (no overdue records for 5 consecutive years), feature vector is [1, 0.9, 0.95]). Then, it is input into a support vector machine regression model trained with 1000 "evaluation strategy signal - business feedback data" from the bank over the past 3 years (using the RBF kernel function, model prediction error < 0.05), and the calculated signal strength value is 0.85. It is then determined whether this signal strength value exceeds the preset strength threshold, which is set at 0.7. Since 0.85 > 0.7, it is determined to exceed the preset strength threshold. Combining the characteristics of the e-commerce industry—"high capital turnover rate and strong credit demand"—the credit resource allocation ratio is determined to be 15%, and the credit limit for D04 is determined based on the credit resource allocation ratio. First, calculate the industry benchmark credit line (D04's net assets are 80 million yuan, and the credit coefficient for the electronic trade industry is set at 0.6 according to the disclosure document, so the industry benchmark credit line = 80 million yuan × 0.6 = 48 million yuan). Then, calculate the final credit line = 48 million yuan × 15% = 7.2 million yuan. This amount is higher than the minimum credit line of 5 million yuan for the electronic trade industry. Finally, the credit line of 7.2 million yuan will be recorded in D04's customer credit file.

[0056] In the aforementioned method, on the one hand, by transforming the high-level contribution of enterprise D04 into a structured evaluation strategy signal containing specific business instructions, the pain point of traditional contribution evaluation methods—where contribution levels are merely abstract classifications and cannot be directly translated into business operations—is addressed, ensuring that the evaluation results are actionable. On the other hand, based on the support vector machine algorithm, the strategy signal is transformed into a quantified signal strength value. The model trained on historical data ensures the objectivity of the strength value, avoiding bias from subjective human judgment. Simultaneously, a judgment mechanism with preset strength thresholds accurately identifies high-priority enterprises like D04, ensuring that banks prioritize credit resources for clients with signal strength exceeding the threshold and higher expected returns, avoiding resource misallocation. Furthermore, different credit resource allocation ratios are determined based on the characteristics of different industries, rather than using a uniform ratio, reflecting industry adaptability and avoiding extensive cross-industry resource allocation. Thus, it achieves precise matching of credit limits with enterprise financial strength and industry characteristics, while ensuring the basic operational needs of enterprises, effectively balancing optimal resource utilization and risk control, and significantly improving the accuracy, industry adaptability, and resource utilization efficiency of banks' corporate credit decisions.

[0057] In one embodiment of this disclosure, after determining the credit limit of the enterprise customer according to the credit resource allocation ratio in step S650, the following steps are further included: Figure 7 As shown, the specific content is as follows: Step S710: Based on the linear regression model, determine the predicted value of credit resource utilization rate according to the credit limit of the enterprise customer; Step S720: Determine whether the predicted value of the credit resource utilization rate exceeds the preset utilization rate threshold; Step S730: If it is determined that the preset utilization rate threshold has not been exceeded, the credit resource allocation ratio is adjusted; Step S740: Determine the final credit limit for the corporate customer based on the adjusted credit resource allocation ratio.

[0058] Specifically, taking e-commerce company D04 with a credit line of 7.2 million yuan as an example, based on the bank's historical data of 800 e-commerce clients over the past three years (including credit lines and corresponding actual credit resource utilization rates, excluding extreme samples), a linear regression model was trained. Substituting company D04's credit line of 7.2 million yuan into the linear regression model, the preliminary calculated predicted credit resource utilization rate was 78.5%. The preset utilization rate threshold is 80%. It can be seen that 78.5% does not exceed this preset threshold. Therefore, the credit resource allocation ratio will be adjusted, and the adjustment formula is as follows: in, The adjusted credit resource allocation ratio. The basic credit resource allocation ratio is s, where s is the contribution score and k is the incremental coefficient.

[0059] Among them, the basic credit resource allocation ratio was determined based on the high contribution level of enterprise D04 and the characteristics of the e-commerce industry, such as high capital turnover and strong credit demand. The percentage is 15%. Due to the controllable overall risk and high resource utilization efficiency requirements of the e-commerce industry, based on bank statistics of e-commerce industry resource allocation data over the past three years, the incremental coefficient k is set to 2%. This means that for every 0.1 increase in the contribution score, the corresponding credit resource allocation ratio increases by 0.2%, used to quantify the adjustment effect of the contribution score on the credit resource allocation ratio. Its contribution score is 0.75. Substituting the above parameters into the adjustment formula, the adjusted credit resource allocation ratio is: =15%+(0.75×2%)=15%+1.5%=16.5%. The adjusted credit resource allocation ratio of 16.5% is used to calculate the final credit limit, that is, the final credit limit of enterprise D04 = 48 million yuan × 16.5% = 7.92 million yuan.

[0060] In the aforementioned method, on the one hand, by predicting the utilization rate of credit resources based on a linear regression model, the pain point of traditional credit granting—which only determines the credit limit based on credit rating and cannot predict resource idleness—is addressed, providing a quantitative basis for optimizing credit limits. On the other hand, through a judgment mechanism that presets a utilization rate threshold, the idle risk caused by excessively high initial credit limits is accurately identified, preventing banks from allocating resources to low-utilization, low-efficiency credit limits. Simultaneously, if the preset utilization rate threshold is not exceeded, the credit resource allocation ratio is adjusted, achieving adaptive optimization of the credit resource allocation ratio. This avoids the subjectivity of manual adjustments and ensures that the resource utilization rate corresponding to the adjusted credit limit is closer to the target threshold, balancing the needs of enterprise operations with the efficiency of bank resources. Finally, the final credit limit obtained by recalculating through the adjusted credit resource allocation ratio improves the utilization efficiency of bank credit resources, ensures that basic enterprise needs are met, further reduces the risk of resource idleness, and enhances the overall efficiency of the bank's corporate lending business.

[0061] This disclosure also provides a comprehensive contribution evaluation method system based on reinforcement learning. The system may include a first acquisition module, a first determination module, an extraction module, a second determination module, a first judgment module, a processing module, a prediction module, a third determination module, a fourth determination module, a second judgment module, a second acquisition module, a fifth determination module, an update module, a sixth determination module, and a seventh determination module. Specifically, the first acquisition module acquires multi-source heterogeneous data from the financial statements and transaction records of enterprise customers, including financial indicator values ​​and transaction behavior sequences; the first determination module determines a customer status dataset based on the financial indicator values ​​and transaction behavior sequences; the extraction module extracts key feature vectors from the customer status dataset based on principal component analysis, including financial indicator feature vectors and transaction behavior feature vectors; the second determination module determines a fusion feature matrix based on the financial indicator feature vectors and transaction behavior feature vectors; the judgment module determines whether the fluctuation range of the fusion feature matrix exceeds a preset fluctuation range threshold; the processing module standardizes the transaction behavior feature vectors to determine a standardized transaction behavior sequence feature matrix if the fluctuation range exceeds the preset threshold; and the prediction module predicts the transaction behavior sequence feature matrix based on a recurrent neural network. The system comprises seven modules: a feature matrix predicting transaction behavior trends; a third determination module determining a dynamic risk vector based on these trends; a fourth determination module determining a contribution score based on the dynamic risk vector and the fused feature matrix; a second judgment module determining whether the absolute value of the deviation between the current contribution score and the industry's benchmark contribution exceeds a preset deviation threshold; a second acquisition module acquiring feature data from multiple enterprise customers within the industry, including annual revenue, transaction frequency, and customer loyalty, if the threshold is exceeded, and using this data as the dataset to be clustered; a fifth determination module grouping the dataset based on the K-means clustering algorithm to determine the grouping results for enterprise customer contribution levels; an update module updating the contribution score based on the grouping results; a sixth determination module determining the contribution level based on the updated contribution score; and a seventh determination module determining the credit limit for enterprise customers based on the contribution level.

[0062] It should be noted that the embodiments of the reinforcement learning-based comprehensive contribution evaluation method system provided in this application can be used to execute the processing flow of the reinforcement learning-based comprehensive contribution evaluation method embodiments described above. Its functions will not be repeated here, but can be referred to the detailed description of the above method embodiments.

[0063] As described above, the comprehensive contribution evaluation method system based on reinforcement learning provided in this disclosure integrates multi-source heterogeneous data such as financial statements and transaction records of enterprise customers. It uses principal component analysis to extract feature vectors of financial indicators and transaction behavior to construct a fused feature matrix. This overcomes the limitations of traditional methods in coordinating structured financial indicator data and unstructured transaction behavior data, achieving comprehensive aggregation of multi-dimensional information about enterprise customers. Simultaneously, a judgment mechanism with a preset fluctuation threshold is introduced. When the fluctuation of the fused feature matrix exceeds the limit, the transaction behavior feature vector is standardized, and a recurrent neural network is used to predict transaction behavior trends to generate a dynamic risk vector, replacing the traditional fixed-rule model. This constructs a dynamic evaluation system that responds in real-time to changes in the enterprise's operating status, promptly capturing the liquidity risk or long-term development potential behind short-term surges in transaction volume, solving the problem of traditional evaluations lagging in responding to dynamic changes in enterprises. By determining the contribution score through the fused feature matrix and dynamic risk vector, it takes into account both the stability of the enterprise's historical comprehensive characteristics and incorporates the dynamism of real-time risk factors, avoiding the one-sidedness of evaluation caused by traditional methods relying solely on financial data, and significantly improving the accuracy of enterprise customer contribution evaluation. Meanwhile, by comparing with industry contribution benchmarks, K-means clustering is used to group industry enterprise characteristic data for cases exceeding preset deviation thresholds to calibrate contribution scores, correct assessment biases, and ensure the industry suitability and objectivity of contribution level determination. Ultimately, credit limits are determined based on accurate contribution levels, achieving optimized allocation of bank credit resources under the premise of dynamic risk control, improving capital allocation efficiency and risk control capabilities, and accurately matching the actual operational needs of enterprises, thus providing strong support for the development of the real economy.

[0064] This disclosure also provides an electronic device including one or more processors and memory resources, represented by memory, for storing instructions executable by the processors, such as application programs. The application programs stored in the memory may include one or more modules, each corresponding to a set of instructions. Furthermore, the processor is configured to execute instructions to perform the aforementioned reinforcement learning-based comprehensive contribution evaluation method.

[0065] The electronic device may also include a power supply component configured to perform power management of the electronic device, a wired or wireless network interface configured to connect the electronic device to a network, and an input / output (I / O) interface. The electronic device can be operated based on operating devices stored in memory, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0066] In one embodiment, a computer device, which may be a server, is also provided. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computational and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device stores data. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a reinforcement learning-based comprehensive contribution evaluation method.

[0067] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a reinforcement learning-based comprehensive contribution evaluation method. The display unit of the computer device is used to form a visually visible image and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0068] This disclosure also provides a non-transitory computer-readable storage medium. When the instructions in the storage medium are executed by the processor of the aforementioned electronic device, the electronic device is able to execute a comprehensive contribution evaluation method based on reinforcement learning, including: acquiring multi-source heterogeneous data from the financial statements and transaction records of enterprise customers, the multi-source heterogeneous data including financial indicator values ​​and transaction behavior sequences; determining a customer status dataset based on the financial indicator values ​​and transaction behavior sequences; extracting key feature vectors from the customer status dataset based on a principal component analysis algorithm, the key feature vectors including financial indicator feature vectors and transaction behavior feature vectors; determining a fusion feature matrix based on the financial indicator feature vectors and transaction behavior feature vectors; determining whether the fluctuation amplitude of the fusion feature matrix exceeds a preset fluctuation amplitude threshold; if it is determined that it exceeds the preset fluctuation amplitude threshold, then... The eigenvectors are standardized to determine a standardized transaction behavior sequence feature matrix. Based on a recurrent neural network, transaction behavior trends are predicted using the standardized transaction behavior sequence feature matrix. A dynamic risk vector is determined based on the transaction behavior trend. A contribution score is determined based on the dynamic risk vector and the fused feature matrix. The absolute value of the deviation between the current contribution score and the industry's benchmark contribution exceeds a preset deviation threshold. If the threshold is exceeded, feature data from multiple enterprise customers within the industry, including annual revenue, transaction frequency, and customer loyalty, are acquired and used as the clustering dataset. The dataset is grouped using the K-means clustering algorithm to determine the grouping results for enterprise customer contribution levels. The contribution score is updated based on the grouping results. The contribution level is determined based on the updated contribution score. Finally, the credit limit for enterprise customers is determined based on the contribution level.

[0069] This disclosure can take the form of a computer program product implemented on one or more storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0070] It should be noted that although the steps of the reinforcement learning-based comprehensive contribution evaluation method of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps, such as omitting certain steps, combining multiple steps into one step, and / or decomposing a step into multiple steps, should all be considered part of this disclosure.

[0071] It should be understood that this disclosure is not limited to the detailed structure and arrangement of the modules in the reinforcement learning-based comprehensive contribution evaluation system proposed in this specification. This disclosure is capable of other implementations and can be implemented and performed in various ways. The foregoing variations and modifications fall within the scope of this disclosure. It should be understood that this disclosure, as disclosed and defined in this specification, extends to all alternative combinations of two or more individual features mentioned or apparent in the text and / or figures. All these different combinations constitute multiple alternative aspects of this disclosure. The embodiments described in this specification illustrate the best known mode for implementing this disclosure and will enable those skilled in the art to utilize this disclosure.

Claims

1. A comprehensive contribution evaluation method based on reinforcement learning, characterized in that, include: Obtain multi-source heterogeneous data from corporate clients' financial statements and transaction records, wherein the multi-source heterogeneous data includes financial indicator values ​​and transaction behavior sequences; A customer status dataset is determined based on the financial indicator values ​​and the transaction behavior sequence. Based on the principal component analysis algorithm, key feature vectors are extracted from the customer status dataset. The key feature vectors include financial indicator feature vectors and transaction behavior feature vectors. A fusion feature matrix is ​​determined based on the financial indicator feature vector and the transaction behavior feature vector; Determine whether the fluctuation amplitude of the fused feature matrix exceeds a preset fluctuation amplitude threshold; If the fluctuation exceeds the preset threshold, the trading behavior feature vector is standardized to determine the standardized trading behavior sequence feature matrix. Based on a recurrent neural network, the trend of trading behavior is predicted according to the feature matrix of the standardized trading behavior sequence. Determine the dynamic risk vector based on the aforementioned trading behavior trends; The contribution score is determined based on the dynamic risk vector and the fusion feature matrix; Determine whether the absolute value of the deviation between the current contribution score and the industry's enterprise contribution benchmark exceeds a preset deviation threshold; If the deviation exceeds the preset threshold, then the characteristic data of multiple enterprise customers in the industry are obtained. The characteristic data includes annual income, transaction frequency and customer loyalty, and is used as a dataset to be clustered. The dataset to be clustered is grouped based on the K-means clustering algorithm to determine the grouping results of enterprise customer contribution levels; Update the contribution score based on the grouping results; The contribution level is determined based on the updated contribution score. The credit limit for corporate clients is determined based on the contribution level.

2. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, After determining the customer status dataset based on the financial indicator values ​​and the transaction behavior sequence, the method further includes: Customer status characteristics are determined based on the customer status dataset, including liquidity ratio, transaction frequency, and net assets. Based on the K-means clustering algorithm, the current ratio, the transaction frequency, and the net assets are used as input features to cluster corporate customers; Customers with different risk levels are identified based on the clustering results. These risk levels include low-risk customers, medium-risk customers, and high-risk customers.

3. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, The extraction of key feature vectors from the customer status dataset based on principal component analysis includes: The financial indicators are standardized to determine standardized financial indicators. Based on the TF-IDF algorithm, a transaction behavior sequence vector is determined according to the transaction behavior sequence; Based on the principal component analysis algorithm, the covariance matrix of the standardized financial indicators and the transaction behavior sequence vector is calculated; The covariance matrix of the standardized financial indicators is subjected to eigenvalue decomposition to determine the feature vector of the financial indicators. The covariance matrix of the transaction behavior vector is subjected to eigenvalue decomposition to determine the transaction behavior feature vector.

4. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, After determining the fused feature matrix based on the financial indicator feature vector and the transaction behavior feature vector, the method further includes: The fused feature matrix is ​​standardized to determine the standardized fused feature matrix; Determine whether the feature dimension of the standardized fusion feature matrix exceeds a preset dimension threshold; If the dimensionality threshold is exceeded, the standardized fusion feature matrix is ​​dimensionality-reduced based on the principal component analysis algorithm to determine the dimensionality-reduced fusion feature matrix.

5. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, Indeed, the contribution score includes: Obtain the weight value corresponding to the dynamic risk vector; Based on the weight values ​​corresponding to the dynamic risk vector, a weighted operation is performed on the fused feature matrix and the dynamic risk vector, and the weighted operation result is determined as the contribution score.

6. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, Updating the contribution score based on the grouping results includes: Determine the average contribution score within the corresponding enterprise customer contribution level grouping results based on the enterprise customer contribution level grouping results; Determine whether the current contribution score does not exceed the average contribution score within the corresponding enterprise customer contribution level grouping results; If it is determined that the contribution score has not exceeded the average contribution score, then the current contribution score is updated to the average contribution score of the corresponding enterprise customer contribution level grouping results; The contribution level is determined based on the updated contribution score, and the contribution level includes low-level contribution, medium-level contribution, and high-level contribution.

7. The comprehensive contribution evaluation method based on reinforcement learning according to claim 1, characterized in that, The process of determining the credit limit for corporate clients based on the contribution level includes: An evaluation strategy signal is generated based on the contribution level; The signal strength value of the evaluation strategy signal is calculated based on the support vector machine algorithm; Determine whether the signal strength value exceeds a preset strength threshold; If the contribution level exceeds the preset strength threshold, the credit resource allocation ratio is determined based on the contribution level. The credit limit for corporate clients is determined based on the credit resource allocation ratio.

8. The comprehensive contribution evaluation method based on reinforcement learning according to claim 7, characterized in that, After determining the credit limit for corporate clients based on the credit resource allocation ratio, the method further includes: Based on the linear regression model, the predicted value of credit resource utilization rate is determined according to the credit limit of the enterprise customer; Determine whether the predicted value of the credit resource utilization rate exceeds a preset utilization rate threshold; If it is determined that the preset utilization rate threshold has not been exceeded, the credit resource allocation ratio is adjusted. The final credit limit for corporate clients is determined based on the adjusted credit resource allocation ratio.

9. A comprehensive contribution evaluation method system based on reinforcement learning, characterized in that, include: The first acquisition module is used to acquire multi-source heterogeneous data from the financial statements and transaction records of enterprise customers. The multi-source heterogeneous data includes financial indicator values ​​and transaction behavior sequences. The first determining module is used to determine the customer status dataset based on the financial indicator values ​​and the transaction behavior sequence. The extraction module is used to extract key feature vectors from the customer status dataset based on the principal component analysis algorithm. The key feature vectors include financial indicator feature vectors and transaction behavior feature vectors. The second determining module is used to determine a fusion feature matrix based on the financial indicator feature vector and the transaction behavior feature vector; The first judgment module is used to determine whether the fluctuation amplitude of the fused feature matrix exceeds a preset fluctuation amplitude threshold. The processing module is used to standardize the trading behavior feature vector if it is determined that the fluctuation range exceeds the preset threshold, so as to determine the standardized trading behavior sequence feature matrix. The prediction module is used to predict trading behavior trends based on the standardized trading behavior sequence feature matrix using a recurrent neural network. The third determining module is used to determine the dynamic risk vector based on the trend of the trading behavior. The fourth determining module is used to determine the contribution score based on the dynamic risk vector and the fusion feature matrix; The second judgment module is used to determine whether the absolute value of the deviation between the current contribution score and the contribution benchmark of enterprises in the industry exceeds a preset deviation threshold. The second acquisition module is used to acquire feature data of multiple enterprise customers in the industry if it is determined that the deviation exceeds the preset threshold. The feature data includes annual income, transaction frequency and customer loyalty, and is used as a dataset to be clustered. The fifth determining module is used to group the dataset to be clustered based on the K-means clustering algorithm to determine the grouping results of the enterprise customer contribution level; An update module is used to update the contribution score based on the grouping results; The sixth determining module is used to determine the contribution level based on the updated contribution score; The seventh determination module is used to determine the credit limit of the enterprise customer based on the contribution level.

Images