Credit extension risk assessment method and device, electronic equipment and storage medium

Abstract

The invention discloses a credit granting risk assessment method and device, electronic equipment and a storage medium, relates to the technical field of data analysis, can be applied to financial science and technology scenes, and comprises the following steps: obtaining multi-dimensional dynamic feature data, and generating a multi-scene sample data set containing the multi-dimensional dynamic feature data and corresponding sample tags; constructing a scene feature vector corresponding to the multi-dimensional dynamic feature data; a self-adaptive evaluation model is trained through the multiple scene feature vectors corresponding to the multi-scene sample data set and the sample labels, and in the training process of the self-adaptive evaluation model, model parameters and evaluation dimension weights are automatically adjusted according to the scene feature vectors; and inputting the real-time feature data of the to-be-evaluated object into the trained adaptive evaluation model to obtain a credit risk evaluation result. According to the invention, the accuracy, timeliness and expansibility of credit risk assessment can be improved.

Description

Credit risk assessment methods, devices, electronic equipment and storage media Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to a credit risk assessment method, apparatus, electronic device and storage medium. Background Technology

[0002] In the financial lending sector, credit assessment systems serve as the core technological support for risk control, and the rationality of their model architecture directly impacts the accuracy and applicability of the assessment results. With the increasing complexity of the market environment and the growing differentiation of industry characteristics, existing credit assessment systems are gradually revealing their technical limitations in model architecture design, making it difficult to meet dynamically changing assessment needs.

[0003] Existing credit assessment systems generally adopt a fixed-architecture model design. Their core technical characteristics include: model parameters and assessment dimension weights are manually preset and fixed in the long term, primarily relying on traditional algorithm frameworks such as linear regression, logistic regression, or static scoring cards. From a technical perspective, this fixed architecture cannot adapt to the business characteristics of different industries. For example, the cash flow fluctuations in manufacturing, the inventory turnover patterns in retail, and the R&D investment return models in the technology industry are fundamentally different. However, a fixed model architecture can only use uniform assessment dimensions and weights, leading to a significant increase in error rates when assessing across industries, and failing to guarantee the accuracy, timeliness, and scalability of credit risk assessment. Summary of the Invention

[0004] In view of this, this application provides a credit risk assessment method, apparatus, electronic device and storage medium, which can improve the accuracy, timeliness and scalability of credit risk assessment.

[0005] According to a first aspect of this application, a credit risk assessment method is provided, comprising: acquiring multi-dimensional dynamic feature data and generating a multi-scenario sample dataset containing the multi-dimensional dynamic feature data and corresponding sample labels, wherein the multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments, and different risk factors; constructing scenario feature vectors corresponding to the multi-dimensional dynamic feature data, wherein the scenario feature vectors are used to characterize the industry attributes, environmental states, and risk types corresponding to the multi-dimensional dynamic feature data; training an adaptive assessment model using multiple scenario feature vectors corresponding to the multi-scenario sample dataset and the sample labels, wherein the adaptive assessment model automatically adjusts model parameters and assessment dimension weights according to the scenario feature vectors during training; and inputting the real-time feature data of the object to be assessed into the trained adaptive assessment model to obtain a credit risk assessment result.

[0006] According to a second aspect of this application, a credit risk assessment device is provided, comprising: a generation module, configured to acquire multi-dimensional dynamic feature data and generate a multi-scenario sample dataset containing the multi-dimensional dynamic feature data and corresponding sample labels, wherein the multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments, and different risk factors; a construction module, configured to construct scenario feature vectors corresponding to the multi-dimensional dynamic feature data, wherein the scenario feature vectors are used to characterize the industry attributes, environmental states, and risk types corresponding to the multi-dimensional dynamic feature data; a training module, configured to train an adaptive assessment model using multiple scenario feature vectors corresponding to the multi-scenario sample dataset and the sample labels, wherein the adaptive assessment model automatically adjusts model parameters and assessment dimension weights according to the scenario feature vectors during training; and an input module, configured to input the real-time feature data of the object to be assessed into the trained adaptive assessment model to obtain a credit risk assessment result.

[0007] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the credit risk assessment method described above.

[0008] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the credit risk assessment method described above.

[0009] By employing the aforementioned technical solutions, the credit risk assessment method, apparatus, electronic device, and storage medium provided in this application effectively address the problem that existing fixed-architecture models cannot adapt to the business characteristics of different industries. This is achieved by acquiring multi-dimensional dynamic feature data reflecting the characteristics of different industries, constructing scenario feature vectors containing industry attributes, and training an adaptive assessment model that automatically adjusts model parameters and assessment dimension weights based on the scenario feature vectors. The multi-dimensional dynamic feature data covers the unique characteristics of each industry, providing targeted training basis for the adaptive assessment model. The industry attribute labels in the scenario feature vectors accurately identify the industry to which the data belongs, clarifying the adjustment direction of the model. During training, the adaptive assessment model learns the parameter and weight adjustment rules under different industry scenarios. When assessing an object to be assessed in a specific industry, it automatically calls the appropriate assessment dimension weights based on the industry scenario corresponding to its real-time feature data. This avoids the problem of increased cross-industry assessment error rates caused by using fixed models with uniform assessment dimensions and weights, thereby improving the accuracy of credit risk assessment, its adaptability to different industries, and its overall scalability.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 shows a flowchart of a credit risk assessment method provided by an embodiment of this application; Figure 2 shows a flowchart of a credit risk assessment method provided by another embodiment of this application; Figure 3 shows a structural schematic diagram of a credit risk assessment device provided by an embodiment of this application; Figure 4 shows a structural schematic diagram of a credit risk assessment device provided by another embodiment of this application. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] Existing credit assessment systems generally adopt a fixed-architecture model design. Their core technical characteristics include: model parameters and assessment dimension weights are manually preset and fixed in the long term, primarily relying on traditional algorithm frameworks such as linear regression, logistic regression, or static scoring cards. From a technical perspective, this fixed architecture cannot adapt to the business characteristics of different industries. For example, the cash flow fluctuations in manufacturing, the inventory turnover patterns in retail, and the R&D investment return models in the technology industry are fundamentally different. However, a fixed model architecture can only use uniform assessment dimensions and weights, leading to a significant increase in error rates when assessing across industries, and failing to guarantee the accuracy, timeliness, and scalability of credit risk assessment.

[0014] To address the aforementioned technical issues, this invention provides a credit risk assessment method, as shown in Figure 1. The method includes: step 110, acquiring multi-dimensional dynamic feature data and generating a multi-scenario sample dataset containing multi-dimensional dynamic feature data and corresponding sample labels. The multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments, and different risk factors.

[0015] Among them, multi-dimensional dynamic feature data refers to feature data that is collected from multiple dimensions (industry, market, risk, etc.) and changes dynamically over time. It covers the specific characteristics of different industries (such as R&D investment in the technology industry), the real-time status of the market environment (such as interest rate fluctuations), and the performance of various risk factors (such as logistics delays), and has real-time and multi-source characteristics. Sample labels are identifiers used to mark the actual credit risk assessment results corresponding to the multi-dimensional dynamic feature data (such as "high risk", "low risk" or specific risk level values), and are the reference standard for judging the correctness of prediction results during model training. Multi-scenario sample datasets refer to the collection composed of multi-dimensional dynamic feature data and corresponding sample labels.

[0016] In this embodiment of the disclosure, real-time data reflecting different industry characteristics, market environments, and risk factors is collected and combined with corresponding sample labels to construct a multi-scenario sample dataset for training an adaptive evaluation model. This overcomes the limitations of traditional fixed-architecture models that rely on single-dimensional static data. By incorporating dynamic features from multiple scenarios, it provides a comprehensive and real-time updated data foundation for subsequent model training. This enables the model to learn risk patterns in different scenarios, laying a data foundation for adaptively adjusting parameters and weights, thereby improving the model's adaptability to cross-industry, dynamic environments, and new risks.

[0017] Step 120: Construct scene feature vectors corresponding to multi-dimensional dynamic feature data. Scene feature vectors are used to characterize the industry attributes, environmental status, and risk types corresponding to multi-dimensional dynamic feature data.

[0018] Among them, the scenario feature vector is a structured vector data form that is formed by fusing multi-dimensional dynamic feature data with scenario information such as industry attributes, environmental status, and risk type. It is used to convey the scenario attributes corresponding to the data to the adaptive assessment model and is the core input element for the model to achieve scenario adaptation. The industry attribute is used to characterize the characteristics of the industry to which the data belongs (such as manufacturing, retail, technology, etc.). The environmental status refers to the state information of the market environment in which the data was generated (such as macro policies, economic cycles, market fluctuations, etc.) and is used to reflect the impact of the external environment on risk assessment. The risk type refers to the risk category contained in the data (such as credit risk, market risk, operational risk, etc.) and is used to clarify the risk direction that the model needs to focus on identifying.

[0019] In this embodiment of the disclosure, multi-dimensional dynamic feature data can be processed to transform the industry attributes, environmental conditions and risk types contained therein into a structured vector form, so that the originally scattered feature data can be accurately identified by the model to realize the scene attributes, and achieve deep binding between data and scene information.

[0020] Step 130: Train an adaptive evaluation model using multiple scene feature vectors and sample labels corresponding to the multi-scene sample dataset. During the training process, the adaptive evaluation model automatically adjusts the model parameters and evaluation dimension weights based on the scene feature vectors.

[0021] Among them, the adaptive evaluation model is a model that can autonomously adjust its internal parameters and evaluation logic according to the scene characteristics of the input data. Unlike the fixed architecture model, it has the ability to dynamically adapt to different scenarios. The model parameters refer to the core variables used for calculation and decision-making within the adaptive model (such as the node connection weights of the neural network, the filtering coefficients of the feature extraction layer, etc.), which directly affect the output results of the model. The evaluation dimension weight is the importance coefficient assigned by the adaptive model to different evaluation dimensions (such as enterprise operation, industry dynamics, etc.). The higher the weight, the greater the impact of the dimension on the final evaluation result.

[0022] In this embodiment of the disclosure, the adaptive evaluation model can be trained using multiple scene feature vectors and corresponding sample labels contained in the multi-scene sample dataset. During the training process, the model will autonomously adjust the internal model parameters (such as the connection weights of each layer of the network) and evaluation dimension weights (such as the degree of importance attached to cash flow features and inventory features) according to the specific scene represented by the scene feature vector, so that the model can form an appropriate evaluation logic in different scenarios.

[0023] By enabling adaptive evaluation models to autonomously learn the patterns of different parameters and weights corresponding to different scenarios during training, they can acquire scenario-adaptive capabilities. For example, when facing a manufacturing scenario, the adaptive evaluation model will automatically increase the evaluation weight of cash flow fluctuation characteristics; when facing a retail scenario, it will focus on inventory turnover characteristics, fundamentally solving the problem of high cross-industry evaluation error rates. At the same time, the adaptive evaluation model's dynamic response capability to scenarios can lay the foundation for subsequent responses to changes in the market environment and new risk factors, improving the accuracy and flexibility of evaluation.

[0024] Step 140: Input the real-time feature data of the object to be evaluated into the trained adaptive evaluation model to obtain the credit risk assessment result.

[0025] Among them, the real-time characteristic data of the object to be evaluated refers to the dynamic characteristic data collected in real time related to the object to be evaluated (such as enterprises or individuals), including its current operating data (such as real-time cash flow), the status of its industry, changes in the market environment, and potential risk signals, which are the direct inputs of the model evaluation; the credit risk assessment result is the quantitative output of the adaptive assessment model on credit risk, which usually includes a risk assessment score (such as a risk score of 0-100) and feature contribution weight (the degree of influence of each feature on the risk result), which is used to intuitively reflect the risk level of the object and key influencing factors.

[0026] In this embodiment of the disclosure, after the real-time feature data of the object to be evaluated is input into the trained adaptive evaluation model, the adaptive evaluation model will call the scenario adaptation logic learned during the training phase, automatically match the optimal parameters and evaluation dimension weights according to the scenario corresponding to the real-time feature data, and through internal hierarchical processing (such as feature extraction, scenario adaptation, and decision output), finally generate a credit risk assessment result containing risk assessment score and feature contribution weight, so as to achieve accurate judgment of the current risk status of a specific object.

[0027] Compared to traditional fixed-model evaluation methods that rely on static data and uniform parameters, this embodiment uses real-time feature data as input, enabling the model to evaluate based on the latest information and avoiding misjudgments caused by data lag. At the same time, the model leverages the scenario adaptability learned during the training phase to dynamically adjust the evaluation logic for the specific scenario of the object to be evaluated, ensuring that the results highly match the actual risk status. This significantly improves the timeliness and accuracy of credit risk assessment, providing a reliable basis for credit decisions.

[0028] In summary, the credit risk assessment method provided by this invention, by acquiring multi-dimensional dynamic feature data reflecting the characteristics of different industries, constructing scenario feature vectors containing industry attributes, and training an adaptive assessment model that can automatically adjust model parameters and assessment dimension weights according to the scenario feature vectors, can effectively solve the problem that existing fixed-architecture models cannot adapt to the business characteristics of different industries. The multi-dimensional dynamic feature data covers the unique characteristics of each industry, providing targeted training basis for the adaptive assessment model; the industry attribute labels in the scenario feature vectors can accurately identify the industry to which the data belongs, enabling the model to clearly define its adjustment direction; the adaptive assessment model learns the parameter and weight adjustment rules under different industry scenarios during training, and when assessing an object to be assessed in a specific industry, it automatically calls the assessment dimension weights adapted to that industry based on the industry scenario corresponding to its real-time feature data. This avoids the problem of increased cross-industry assessment error rates caused by fixed models using uniform assessment dimensions and weights, thereby improving the accuracy of credit risk assessment, its adaptability to different industries, and its overall scalability.

[0029] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the implementation of this embodiment, this embodiment also provides another credit risk assessment method, as shown in Figure 2. The method includes: step 210, acquiring multi-dimensional dynamic feature data and generating a multi-scenario sample dataset containing multi-dimensional dynamic feature data and corresponding sample labels. The multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments and different risk factors.

[0030] In the embodiments of this disclosure, step 210, when acquiring multi-dimensional dynamic feature data, may include the following steps: Step 210-1, collecting multi-modal data related to credit risk assessment from multi-source heterogeneous data sources.

[0031] Among them, multi-source heterogeneous data sources refer to data sources with different origins and structures, which may include internal enterprise systems, external open interfaces, policy platforms, public opinion monitoring tools, etc. The data collected may involve various types such as structured tables, text, images, and sensor signals; credit risk assessment refers to the process by which financial institutions analyze the credit status and repayment ability of borrowers or enterprises to determine their default risk and determine credit limits, interest rates, etc.

[0032] This disclosure embodiment can overcome the limitations of traditional credit assessment relying on single static data. It can collect multimodal data through multi-source heterogeneous data sources, such as internal enterprise operating data (e.g., ERP system transaction records, water and electricity consumption curves captured by IoT sensors), external market dynamic data (e.g., commodity futures price API interfaces, e-commerce platform sales rankings), and macroeconomic data (e.g., monetary policy announcements, natural disaster early warning GIS maps, etc.). These data cover multiple modalities such as structured data (transaction amount, inventory quantity) and unstructured data (policy texts, public opinion information). By integrating information scattered across different channels, it can provide comprehensive and real-time raw data support for credit risk assessment.

[0033] Step 210-2: Preprocess the multimodal data. Preprocessing includes at least data cleaning and standardization.

[0034] Data cleaning aims to remove duplicate values, outliers (such as transaction amounts that significantly exceed a reasonable range), and missing values ​​(such as inventory data that a company failed to report) to ensure data accuracy. Standardization, on the other hand, eliminates format differences and magnitude interference between different modalities by unifying data formats (such as converting transaction amounts in different currencies to the same currency unit) and scaling numerical ranges (such as mapping percentage data and absolute values ​​to the same interval), so that the data meets the requirements of subsequent feature engineering and model training.

[0035] Step 210-3: Perform feature engineering on the preprocessed multimodal data to generate multidimensional dynamic feature data that includes at least time-series features, correlation features, and mutation features.

[0036] In the embodiments of this disclosure, key features that reflect risk patterns can be extracted from the cleaned and standardized multimodal data through structured transformation and deep processing. Specifically, these features may include time-series features (such as the month-on-month growth rate of revenue and the weekly fluctuation trend of order volume), correlation features (such as the transaction dependence relationship between upstream and downstream enterprises in the industrial chain and the linkage effect between raw material prices and finished product inventory), and abrupt change features (such as a sudden drop in POS transaction frequency and a surge in negative public opinion information within a certain period of time). Ultimately, a feature data set covering multiple dimensions and dynamically updated over time is formed.

[0037] This processing method can overcome the limitations of traditional models that rely on static single features. By extracting three types of features—time series, correlation, and mutation—the data can more accurately map the nature of risk: time series features capture the dynamic changing trends of enterprise operations, which can solve the problem of insufficient timeliness of historical data in traditional systems; correlation features reveal the transmission relationship between the industrial chain, market environment, and enterprises, which can be used to make up for the problem of isolated risk perception; mutation features can identify abnormal signals in a timely manner, which can avoid the situation where traditional models are slow to respond to sudden risks.

[0038] Step 220: Construct scene feature vectors corresponding to multi-dimensional dynamic feature data. Scene feature vectors are used to characterize the industry attributes, environmental status, and risk types corresponding to multi-dimensional dynamic feature data.

[0039] In the embodiments of this disclosure, step 220, when constructing the scene feature vector corresponding to the multi-dimensional dynamic feature data, may include the following steps: Step 220-1, adding industry attribute tags to the multi-dimensional dynamic feature data based on a preset industry classification standard.

[0040] Among them, industry classification standards refer to pre-defined rules or systems for classifying industry categories, which are used to clarify the scope of different industries such as manufacturing, retail, and technology. Industry attribute tags are identifiers used to identify the industry to which multi-dimensional dynamic feature data belongs, such as "manufacturing", "retail" and "technology". They can intuitively reflect the industry characteristics of the data and are key information for the model to distinguish different industry scenarios.

[0041] For the embodiments of this disclosure, established industry classification rules (such as classification standards for manufacturing, retail, and technology industries) can be referenced to bind the multi-dimensional dynamic feature data obtained after feature engineering with the corresponding industry attributes. For example, data containing features such as ERP transaction flow and steel procurement costs can be labeled with the "manufacturing" label, and data containing features such as POS transaction frequency and e-commerce sales ranking can be labeled with the "retail" label, so that the data can clearly reflect the characteristics of its industry.

[0042] By using industry attribute tags, the adaptive assessment model can accurately identify the industry type corresponding to the data, providing a basis for adjusting the assessment logic for different industries. For example, after labeling multi-dimensional dynamic characteristic data of the manufacturing industry (such as raw material price correlation characteristics and cash flow time series characteristics) with "manufacturing," the adaptive assessment model can call on assessment dimensions suitable for manufacturing (such as focusing on supply chain stability); after labeling retail data with "retail," the adaptive assessment model can focus on dimensions such as inventory turnover and consumption trends. This avoids the errors caused by traditional fixed models using a uniform standard to assess different industries, improving the accuracy of cross-industry risk assessment.

[0043] Step 220-2: Based on macro-environmental parameters and domain fluctuation data, add environmental status labels to the multi-dimensional dynamic feature data.

[0044] Among them, macroeconomic environmental parameters refer to data reflecting the overall economic environment and policy conditions, such as monetary policy, regional GDP growth rate, and natural disaster warnings, which are used to reflect the impact of the external environment on enterprises; domain fluctuation data refers to dynamic change data of specific industries or markets, such as commodity futures price fluctuations, industry index changes, and competitor financing news, which are used to reflect real-time fluctuations within the industry; environmental status labels are used to identify the external environmental status of multi-dimensional dynamic feature data, such as "supply chain cost surge period" and "regional economic downturn period", which are used to help the model adapt to different environmental scenarios.

[0045] In this embodiment of the disclosure, information extracted from macroeconomic indicators and field dynamic data can be combined to mark the external environment status of multi-dimensional dynamic feature data. For example, when the regional GDP growth rate continues to decline, it can be marked as "regional economic downturn". This allows the data to not only contain its own characteristics, but also be associated with external environmental attributes.

[0046] By using environmental status labels, the micro-level characteristics of enterprises are correlated with the macro-environment and industry fluctuations, enabling adaptive assessment models to identify the impact of the external environment on enterprise risk. For example, if a manufacturing enterprise's multi-dimensional dynamic characteristic data (such as cash flow time series characteristics and procurement volume correlation characteristics) is labeled with "a period of soaring raw material prices," the adaptive assessment model will specifically strengthen the assessment weight of cost fluctuations on the enterprise's default risk, avoiding the problem of traditional models underestimating risk by ignoring such environmental factors.

[0047] Step 220-3: Integrate industry attribute tags, environmental status tags, and corresponding multi-dimensional dynamic feature data, and form a structured scene feature vector through feature splicing and dimension mapping.

[0048] In this embodiment of the disclosure, industry attribute labels, environmental status labels and multi-dimensional dynamic feature data can be combined by dimension through feature splicing, and then the integrated information can be transformed into a structured data form that can be recognized by the adaptive evaluation model by using dimension mapping processing (such as converting text labels into numerical vectors), and finally forming a scene feature vector containing industry attributes, environmental status and enterprise dynamic features.

[0049] Step 230: Train an adaptive evaluation model using multiple scene feature vectors and sample labels corresponding to the multi-scene sample dataset. During the training process, the adaptive evaluation model automatically adjusts the model parameters and evaluation dimension weights according to the scene feature vectors.

[0050] The hierarchical neural network architecture corresponding to the adaptive evaluation model includes a basic feature extraction layer, a scene adaptation layer, and a decision output layer. The basic feature extraction layer, the bottom layer of the hierarchical neural network architecture, is responsible for extracting general risk features (such as fused representations of temporal and correlation features) from the scene feature vector, providing basic feature input for upper-layer processing. The scene adaptation layer, the middle layer of the hierarchical neural network architecture, can dynamically adjust model parameters and evaluation dimension weights based on industry and environmental labels in the scene feature vector. The decision output layer, the top layer of the hierarchical neural network architecture, can generate evaluation results based on the scene-adapted features, optimize the model through sample labels, and finally output the risk assessment result.

[0051] In the embodiments of this disclosure, step 230, when training an adaptive evaluation model using multiple scene feature vectors and sample labels corresponding to the multi-scene sample dataset, may include the following steps: Step 230-1, using a basic feature extraction layer to perform general feature extraction on the scene feature vectors, the general features including the fusion representation of temporal features, correlation features and mutation features.

[0052] In this embodiment of the disclosure, the scene feature vector, which integrates industry attribute tags, environmental state tags and multi-dimensional dynamic features, can be processed by the basic feature extraction layer to extract universal features. These features are a fusion representation of time-series features (such as the time series trend of the month-on-month growth rate of enterprise revenue), correlation features (such as the dependence relationship of upstream and downstream transactions in the industrial chain), and abrupt change features (such as abnormal fluctuations such as a sudden drop in POS transaction frequency). This not only preserves the common laws of risk assessment under different scenarios, but also provides basic feature support for the targeted adjustment of the subsequent scenario adaptation layer.

[0053] The basic feature extraction layer's fusion representation of the three types of features can capture the dynamic changes (time-series features), related impacts (related features), and abnormal signals (mutation features) in enterprise operations. It can also extract common risk patterns in different scenarios, avoiding the difficulty in generalizing the model due to feature fragmentation. At the same time, the general features can provide a unified feature foundation for the scenario adaptation layer, making subsequent adjustments to parameters and weights based on industry and environmental labels more targeted.

[0054] Step 230-2: The scene adaptation layer calculates the feature dimension importance coefficient based on the industry attribute label and environmental state label in the scene feature vector, and dynamically adjusts the connection weight of the network node corresponding to the general feature based on the feature dimension importance coefficient to obtain the scene adaptation feature.

[0055] Among them, the feature dimension importance coefficient refers to a quantitative indicator that measures the degree of influence of different assessment dimensions on risk assessment in a specific scenario. The higher the coefficient, the greater the influence of the dimension on the assessment result. The scenario-adapted features are features obtained after the weights are adjusted by the scenario adaptation layer. They have been integrated into the assessment logic of a specific scenario and can accurately reflect the risk characteristics in that scenario, providing targeted input for the decision output layer.

[0056] The scenario adaptation layer, serving as an intermediate layer in the hierarchical neural network architecture of the adaptive assessment model, calculates the importance coefficients of different assessment dimensions (such as cash flow, inventory, and R&D investment) based on the industry attribute labels and environmental state labels carried in the scenario feature vector. It then dynamically adjusts the connection weights of the network nodes corresponding to the general features based on these coefficients. By driving the calculation of feature dimension importance coefficients and node weight adjustments through industry attribute labels and environmental state labels, the model can dynamically optimize its assessment logic for specific scenarios, significantly improving the scenario adaptability and accuracy of risk assessment.

[0057] For example, in manufacturing scenarios that are in a period of supply chain fluctuation, the importance coefficient of raw material procurement-related features and the weight of corresponding nodes will be increased. In retail scenarios that are in a period of low consumption, the weight of inventory turnover timing features will be increased, ultimately generating scenario-adaptive features that are suitable for specific "industry + environment" scenarios.

[0058] Step 230-3: Use the decision output layer to perform feature mapping and nonlinear transformation on the scene adaptation features, generate the evaluation result vector, and calculate the loss value between the evaluation result vector and the sample label.

[0059] As the top layer of the hierarchical neural network architecture of the adaptive evaluation model, the decision output layer receives the scene adaptation features processed by the scene adaptation layer, transforms them into the target evaluation space through feature mapping, and performs nonlinear transformation to capture complex risk relationships (such as the nonlinear correlation between supply chain fluctuations and corporate default probabilities), generating an evaluation result vector containing multi-dimensional indicators such as risk level and default probability. At the same time, this vector is compared with the sample labels (such as actual risk results), and the difference between the two (i.e., the loss value) is calculated through a loss function, providing a basis for optimizing model parameters.

[0060] Feature mapping and nonlinear transformation can effectively fit the nonlinear relationship between the macro environment, industry characteristics and enterprise risks, avoiding the assessment bias caused by the linear assumptions of traditional models. The calculation of loss values ​​can provide a quantitative basis for the iteration of adaptive assessment models. By optimizing the parameters of each layer through backpropagation, the assessment accuracy of the adaptive assessment model can be continuously improved in multiple scenarios, thereby enhancing the accuracy and adaptability of credit assessment.

[0061] Step 230-4: Based on the loss value generated by the decision output layer, optimize the feature extraction parameters of the basic feature extraction layer, the node connection weights of the scene adaptation layer, and the transformation coefficients of the decision output layer layer by layer through the backpropagation algorithm, so that the deviation between the evaluation result vector and the sample label of the adaptive evaluation model under different scene feature vector inputs is less than the first preset threshold, and the trained adaptive evaluation model is obtained.

[0062] The first preset threshold is the maximum allowable deviation between the pre-set evaluation result vector and the sample label. It is the standard for judging whether the model training is complete and ensures the evaluation accuracy of the adaptive evaluation model in multiple scenarios.

[0063] In this embodiment of the disclosure, the model parameters can be optimized layer by layer from the decision output layer using a backpropagation algorithm based on the loss value generated by the decision output layer. Specifically, for the basic feature extraction layer, its feature extraction parameters (such as the weight coefficients of general feature fusion) can be adjusted to more accurately capture temporal, correlation, and mutation features; for the scene adaptation layer, its node connection weights (such as the weights corresponding to the feature dimension importance coefficients under different industry / environment scenarios) can be optimized to enhance scene adaptation capability; for the decision output layer, its transformation coefficients (such as the activation function parameters of nonlinear transformation) can be corrected to improve the accuracy of the evaluation result vector. Through repeated iterative optimization, until the deviation between the output evaluation result vector and the sample label of the adaptive evaluation model is less than a first preset threshold when different scene feature vectors are input, the trained adaptive evaluation model is finally obtained.

[0064] By employing a backpropagation algorithm to collaboratively optimize parameters across layers, the basic feature extraction layer can more accurately extract general features, the scenario adaptation layer can more flexibly adjust scenario weights, and the decision output layer can more accurately generate evaluation results. This ensures the model maintains high-precision evaluation across different scenarios, such as manufacturing and retail, and economic boom and bust periods. Furthermore, the requirement that the deviation be less than a first preset threshold guarantees the stability of the adaptive evaluation model across multiple scenarios, providing a reliable model foundation for subsequent real-time credit assessment and mitigating the cross-scenario error problem caused by fixed parameters in traditional models.

[0065] Step 240: Convert the real-time business data of the object to be evaluated into a feature vector to be evaluated, and generate the corresponding feature vector of the scene to be evaluated. Use the feature vector to be evaluated and the feature vector of the scene to be evaluated as the real-time feature data of the object to be evaluated, and input them into the trained adaptive evaluation model.

[0066] In this embodiment of the disclosure, the real-time business data of the object to be evaluated (such as real-time transaction flow of the enterprise ERP system, POS transaction frequency, commodity futures price fluctuations, etc.) can be processed and first converted into an evaluation feature vector containing time-series features (such as real-time month-on-month revenue comparison), correlation features (such as real-time transaction correlation with suppliers), and mutation features (such as sudden changes in order volume). At the same time, referring to the processing logic of multi-scenario sample datasets, corresponding industry attribute labels and environmental status labels are added to the feature vector, and fused to form the evaluation scenario feature vector. Finally, these two types of vectors are used as the real-time feature data of the object to be evaluated and input into the trained adaptive evaluation model to provide input for subsequent risk assessment.

[0067] Step 250: Using the adaptive evaluation model, the model calls the corresponding model parameters and evaluation dimension weights of the feature vector of the scene to be evaluated, performs credit risk assessment on the feature vector to be evaluated, and outputs the credit risk assessment result including the risk assessment score and feature contribution weight.

[0068] In this embodiment of the disclosure, the trained adaptive evaluation model can be used to call the corresponding adaptive model parameters and adaptive evaluation dimension weights according to the specific scenario represented by the feature vector of the scenario to be evaluated, and perform credit risk assessment processing on the input feature vector to be evaluated. Finally, the credit risk assessment result covering the risk assessment score and feature contribution weight is output, so as to achieve accurate quantification of the risk of the object to be evaluated in a specific scenario.

[0069] By employing a scenario-driven parameter and weighting mechanism, the adaptive evaluation model can dynamically adjust its evaluation logic for specific scenarios. For example, for evaluation targets tagged as "retail industry + off-season for consumption," the contribution weight of inventory turnover characteristics is automatically increased; for targets in the "technology industry + peak period for R&D investment," the evaluation focuses on the correlation between R&D investment and revenue, thus avoiding cross-scenario errors caused by uniform standards. Simultaneously, the output feature contribution weights enhance the interpretability of the evaluation results, improving the accuracy of risk assessment, scenario adaptability, and decision-making transparency.

[0070] In specific application scenarios, after obtaining the credit risk assessment result, which includes the risk assessment score and the feature contribution weight, as a preferred approach, the implementation steps may further include: recording the deviation between the credit risk assessment result and the actual risk feature data subsequently generated by the assessed object to form a feedback dataset; based on the feedback dataset, using hypothesis testing methods to statistically analyze the assessment accuracy of the adaptive assessment model; when the analysis results show that the assessment accuracy is lower than a second preset threshold, triggering the automatic model iteration process, in which the model parameters of the adaptive assessment model are adjusted through a Bayesian optimization algorithm; the second preset threshold is a pre-set minimum standard for assessment accuracy, and when the actual assessment accuracy is lower than this value, it indicates that the model needs to be optimized to ensure that the model always maintains effective assessment capabilities.

[0071] By establishing a correlation between assessment results and actual risks through feedback datasets, the adaptive assessment model can promptly capture new risk characteristics (such as risk patterns in emerging industries). Statistical analysis of hypothesis testing ensures the scientific accuracy of assessment judgments and avoids the blindness of subjective adjustments. Parameter adjustment using the Bayesian optimization algorithm enables efficient model iteration, allowing the model to continuously adapt to changes in the market environment and industry characteristics (such as policy adjustments and the emergence of new risk factors). This significantly improves the long-term stability and assessment accuracy of the adaptive assessment model, forming a closed loop of "assessment-feedback-optimization," which can compensate for the shortcomings of traditional models that are rigid and difficult to evolve.

[0072] After obtaining the credit risk assessment results, including the risk assessment score and the feature contribution weight, as an optional approach, the implementation steps may further include: constructing a risk warning heatmap with the feature contribution weight as the vertical axis and the feature dimension as the horizontal axis; determining the target risk level range corresponding to the risk assessment score, and mapping the feature contribution weight of each feature dimension in the risk warning heatmap within the target risk level range to a color gradient value; and overlaying the corresponding warning indicator on the risk warning heatmap based on the risk level range where the risk assessment score is located.

[0073] The color gradient value is the color depth (e.g., light yellow to dark red) mapped to the feature contribution weight of each feature dimension. The darker the color, the greater the contribution of the feature to the current risk, so as to achieve the visual differentiation of risk factors. The warning mark is a warning mark superimposed according to the risk level range (e.g., yellow, orange, red).

[0074] The risk warning heatmap visualizes the risk contribution of each feature through color gradients, enabling decision-makers to quickly identify key risk points; the overlay of warning icons directly conveys the urgency of the risk. Compared to traditional systems that only output risk scores, this approach improves risk interpretability and response efficiency, shortens the time from risk identification to intervention, and reduces asset losses.

[0075] In summary, the technical solution in this application, through the acquisition and processing of multi-dimensional dynamic feature data, constructs scene feature vectors by combining industry attribute labels and environmental state labels, and trains an adaptive evaluation model based on a hierarchical neural network architecture. This enables precise adaptation to the business characteristics of different industries: the scene adaptation layer can calculate the importance coefficient of feature dimensions based on industry attribute labels and dynamically adjust the connection weights of network nodes, thereby solving the cross-industry error problem caused by fixed models and uniform evaluation dimensions and weights; by optimizing the parameters of each layer through the backpropagation algorithm, the evaluation bias of the model under different industry scenarios can be ensured to be controllable; and by combining hypothesis testing and Bayesian optimization iteration with feedback datasets, the model's dynamic adaptability to changes in industry characteristics can be further improved, significantly enhancing the accuracy, cross-industry adaptability, and system scalability of credit risk assessment.

[0076] Furthermore, as a specific implementation of the methods shown in Figures 1 and 2, this embodiment provides a credit risk assessment device, as shown in Figure 3. The device includes: a generation module 31, a construction module 32, a training module 33, and an input module 34.

[0077] The generation module 31 can be used to acquire multi-dimensional dynamic feature data and generate a multi-scenario sample dataset containing multi-dimensional dynamic feature data and corresponding sample labels. The multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments, and different risk factors. The construction module 32 can be used to construct scene feature vectors corresponding to the multi-dimensional dynamic feature data. The scene feature vectors are used to characterize the industry attributes, environmental states, and risk types corresponding to the multi-dimensional dynamic feature data. The training module 33 can be used to train an adaptive evaluation model using multiple scene feature vectors and sample labels corresponding to the multi-scenario sample dataset. During the training process, the adaptive evaluation model automatically adjusts the model parameters and evaluation dimension weights according to the scene feature vectors. The input module 34 can be used to input the real-time feature data of the object to be evaluated into the trained adaptive evaluation model to obtain the credit risk assessment result.

[0078] In some embodiments of this application, when acquiring multi-dimensional dynamic feature data, the generation module 31 can be used to collect multi-modal data related to credit risk assessment from multiple heterogeneous data sources; preprocess the multi-modal data, including at least data cleaning and standardization; and perform feature engineering on the preprocessed multi-modal data to generate multi-dimensional dynamic feature data that includes at least time-series features, correlation features, and mutation features.

[0079] In some embodiments of this application, the construction module 32 can be specifically used to add industry attribute labels to multi-dimensional dynamic feature data based on preset industry classification standards; add environmental status labels to multi-dimensional dynamic feature data according to macro environmental parameters and domain fluctuation data; and fuse the industry attribute labels, environmental status labels and corresponding multi-dimensional dynamic feature data to form a structured scene feature vector through feature splicing and dimension mapping processing.

[0080] In some embodiments of this application, the hierarchical neural network architecture corresponding to the adaptive evaluation model includes a basic feature extraction layer, a scene adaptation layer, and a decision output layer. The training module 33 can be specifically used to extract general features from the scene feature vector using the basic feature extraction layer. The general features include a fusion representation of temporal features, correlation features, and mutation features. The scene adaptation layer calculates the feature dimension importance coefficient based on the industry attribute label and environmental state label in the scene feature vector, and dynamically adjusts the connection weights of the network nodes corresponding to the general features according to the feature dimension importance coefficient to obtain scene adaptation features. The decision output layer performs feature mapping and nonlinear transformation on the scene adaptation features to generate an evaluation result vector, and calculates the loss value between the evaluation result vector and the sample label. Based on the loss value generated by the decision output layer, the feature extraction parameters of the basic feature extraction layer, the node connection weights of the scene adaptation layer, and the transformation coefficients of the decision output layer are optimized layer by layer through the backpropagation algorithm, so that the deviation between the evaluation result vector and the sample label of the adaptive evaluation model under different scene feature vector inputs is less than a first preset threshold, thus obtaining the trained adaptive evaluation model.

[0081] In some embodiments of this application, the input module 34 is specifically used to convert the real-time business data of the object to be evaluated into a feature vector to be evaluated, and generate a corresponding feature vector of the scenario to be evaluated; input the feature vector to be evaluated and the feature vector of the scenario to be evaluated as the real-time feature data of the object to be evaluated into the trained adaptive evaluation model; use the adaptive evaluation model to call the adaptive model parameters and adaptive evaluation dimension weights of the scenario corresponding to the feature vector to be evaluated, perform credit risk assessment processing on the feature vector to be evaluated, and output a credit risk assessment result including a risk assessment score and feature contribution weight.

[0082] In some embodiments of this application, as shown in FIG4, the device further includes: a recording module 35, an analysis module 36, and an adjustment module 37; the recording module 35 can be used to record the deviation value between the credit risk assessment result and the actual risk characteristic data subsequently generated by the object to be assessed, forming a feedback dataset; the analysis module 36 can be used to perform statistical analysis on the assessment accuracy of the adaptive assessment model based on the feedback dataset using hypothesis testing methods; the adjustment module 37 can be used to trigger the automatic model iteration process when the analysis result shows that the assessment accuracy is lower than a second preset threshold, and in the automatic model iteration process, adjust the model parameters of the adaptive assessment model through a Bayesian optimization algorithm.

[0083] In some embodiments of this application, the construction module 32 can also be used to construct a risk warning heatmap with feature contribution weight as the vertical axis and feature dimension as the horizontal axis; determine the target risk level range corresponding to the risk assessment score; map the feature contribution weight of each feature dimension in the risk warning heatmap within the target risk level range to a color gradient value; and overlay the corresponding warning mark on the risk warning heatmap based on the risk level range where the risk assessment score is located.

[0084] It should be noted that other corresponding descriptions of the functional units involved in the credit risk assessment device provided in this embodiment can be found in the corresponding descriptions in Figures 1 and 2, and will not be repeated here.

[0085] Based on the methods shown in Figures 1 and 2, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the credit risk assessment method shown in Figures 1 and 2.

[0086] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0087] Based on the methods shown in Figures 1 and 2, and the virtual device embodiments shown in Figures 3 and 4, in order to achieve the above objectives, this application also provides an electronic device, which may be a personal computer, tablet computer, server, or other network device, etc. The device includes a storage medium and a processor; the storage medium is used to store computer programs; the processor is used to execute the computer programs to implement the credit risk assessment methods shown in Figures 1 and 2.

[0088] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0089] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0090] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0091] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0092] This invention, through the acquisition and processing of multi-dimensional dynamic feature data, constructs scene feature vectors by combining industry attribute labels and environmental state labels. Relying on a hierarchical neural network architecture to train an adaptive evaluation model, it achieves precise adaptation to the business characteristics of different industries. The scene adaptation layer can calculate the importance coefficient of feature dimensions based on industry attribute labels and dynamically adjust the connection weights of network nodes, thereby solving the cross-industry error problem caused by fixed models and uniform evaluation dimensions and weights. Optimizing the parameters of each layer through the backpropagation algorithm ensures that the evaluation bias of the model under different industry scenarios is controllable. Combining hypothesis testing and Bayesian optimization iteration with the feedback dataset further enhances the model's dynamic adaptability to changes in industry characteristics, significantly improving the accuracy of credit risk assessment, cross-industry adaptability, and system scalability.

[0093] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0094] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A credit risk assessment method, characterized in that, include: Acquire multi-dimensional dynamic feature data and generate a multi-scenario sample dataset containing the multi-dimensional dynamic feature data and corresponding sample labels. The multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments, and different risk factors. Construct scene feature vectors corresponding to the multi-dimensional dynamic feature data. The scene feature vectors are used to characterize the industry attributes, environmental states, and risk types corresponding to the multi-dimensional dynamic feature data. Train an adaptive evaluation model using multiple scene feature vectors corresponding to the multi-scenario sample dataset and the sample labels. During training, the adaptive evaluation model automatically adjusts the model parameters and evaluation dimension weights according to the scene feature vectors. Input the real-time feature data of the object to be evaluated into the trained adaptive evaluation model to obtain the credit risk assessment result.

2. The method according to claim 1, characterized in that, The acquisition of multi-dimensional dynamic feature data includes: collecting multi-modal data related to credit risk assessment from multiple heterogeneous data sources; preprocessing the multi-modal data, the preprocessing including at least data cleaning and standardization; and performing feature engineering on the preprocessed multi-modal data to generate multi-dimensional dynamic feature data that includes at least time-series features, correlation features, and mutation features.

3. The method according to claim 1, characterized in that, Constructing the scene feature vector corresponding to the multi-dimensional dynamic feature data includes: adding industry attribute tags to the multi-dimensional dynamic feature data based on a preset industry classification standard; adding environmental state tags to the multi-dimensional dynamic feature data according to macro-environmental parameters and domain fluctuation data; and fusing the industry attribute tags, the environmental state tags, and the corresponding multi-dimensional dynamic feature data, and forming a structured scene feature vector through feature splicing and dimension mapping processing.

4. The method according to claim 1, characterized in that, The hierarchical neural network architecture corresponding to the adaptive evaluation model includes a basic feature extraction layer, a scene adaptation layer, and a decision output layer. Training the adaptive evaluation model using multiple scene feature vectors corresponding to the multi-scene sample dataset and the sample labels includes: using the basic feature extraction layer to extract general features from the scene feature vectors, where the general features include a fusion representation of temporal features, correlation features, and mutation features; the scene adaptation layer calculates the feature dimension importance coefficient based on the industry attribute labels and environmental state labels in the scene feature vectors, and dynamically adjusts the connection weights of the network nodes corresponding to the general features based on the feature dimension importance coefficients to obtain scene adaptation features; using the decision output layer to perform feature mapping and nonlinear transformation on the scene adaptation features to generate an evaluation result vector, and calculating the loss value between the evaluation result vector and the sample labels; based on the loss value generated by the decision output layer, optimizing the feature extraction parameters of the basic feature extraction layer, the node connection weights of the scene adaptation layer, and the transformation coefficients of the decision output layer layer by layer using a backpropagation algorithm, so that the deviation between the evaluation result vector and the sample labels under different scene feature vector inputs is less than a first preset threshold, thus obtaining the trained adaptive evaluation model.

5. The method according to claim 1, characterized in that, The step of inputting the real-time feature data of the object to be evaluated into the trained adaptive evaluation model to obtain the credit risk assessment result includes: converting the real-time business data of the object to be evaluated into a feature vector to be evaluated, and generating a corresponding feature vector of the scenario to be evaluated; inputting the feature vector to be evaluated and the feature vector of the scenario to be evaluated as the real-time feature data of the object to be evaluated into the trained adaptive evaluation model; using the adaptive evaluation model to call the adaptive model parameters and adaptive evaluation dimension weights of the scenario corresponding to the feature vector to be evaluated, performing credit risk assessment processing on the feature vector to be evaluated, and outputting a credit risk assessment result including a risk assessment score and feature contribution weight.

6. The method according to claim 5, characterized in that, After using the adaptive evaluation model to call the adaptive model parameters and adaptive evaluation dimension weights of the feature vector of the scene to be evaluated, and performing credit risk assessment on the feature vector to be evaluated, and outputting a credit risk assessment result including a risk assessment score and feature contribution weight, the method further includes: recording the deviation value between the credit risk assessment result and the actual risk feature data subsequently generated by the object to be evaluated, forming a feedback dataset; based on the feedback dataset, using a hypothesis testing method to statistically analyze the evaluation accuracy of the adaptive evaluation model; when the analysis result shows that the evaluation accuracy is lower than a second preset threshold, triggering an automatic model iteration process, in which the model parameters of the adaptive evaluation model are adjusted using a Bayesian optimization algorithm.

7. The method according to claim 5, characterized in that, After using the adaptive evaluation model to call the adaptive model parameters and adaptive evaluation dimension weights of the scene to be evaluated feature vector, and performing credit risk assessment processing on the feature vector to be evaluated, and outputting a credit risk assessment result including a risk assessment score and feature contribution weights, the method further includes: constructing a risk warning heatmap with the feature contribution weights as the vertical axis and the feature dimensions as the horizontal axis; determining the target risk level range corresponding to the risk assessment score, and mapping the feature contribution weights of each feature dimension in the risk warning heatmap within the target risk level range to color gradient values; and overlaying corresponding warning signs on the risk warning heatmap based on the risk level range where the risk assessment score is located.

8. A credit risk assessment device, characterized in that, include: The generation module is used to acquire multi-dimensional dynamic feature data and generate a multi-scenario sample dataset containing the multi-dimensional dynamic feature data and corresponding sample labels. The multi-dimensional dynamic feature data is real-time feature data reflecting different industry characteristics, different market environments and different risk factors. The construction module is used to construct the scene feature vector corresponding to the multi-dimensional dynamic feature data. The scene feature vector is used to characterize the industry attributes, environmental status and risk type corresponding to the multi-dimensional dynamic feature data. The training module is used to train an adaptive evaluation model using multiple scene feature vectors corresponding to the multi-scene sample dataset and the sample labels. During the training process, the adaptive evaluation model automatically adjusts the model parameters and evaluation dimension weights according to the scene feature vectors. The input module is used to input the real-time feature data of the object to be evaluated into the trained adaptive evaluation model to obtain the credit risk assessment result.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images