Credit evaluation method, system, equipment and medium for medical and creative enterprises in government affair field

Abstract

The invention discloses a method, a system, equipment and a medium for evaluating credits of medical and creative enterprises in the government affair field, and belongs to the technical field of credit risk evaluation, machine learning and transfer learning. The invention aims to solve the technical problems that in the existing government affair field, data and label accumulation is insufficient in enterprise credit evaluation, a traditional evaluation method is high in subjectivity, a machine learning model is poor in application effect, and high-quality data resources of financial institutions cannot be effectively utilized. According to the technical scheme, cross-domain data collaborative treatment is carried out, wherein cross-domain multi-source heterogeneous data of a financial source domain and a government affair target domain are collected, and standardization processing, enterprise subject identifier association and cross-domain wide table construction and synchronous updating are carried out on the cross-domain multi-source heterogeneous data; financial field feature engineering processing: generating a financial feature matrix through four-layer feature filtering and three-class derivative operator operation of financial source domain data; training a dynamic coupling architecture source domain model; carrying out three-layer mixed cross-domain migration; and training and optimizing the government affair target domain model.

Description

Technical Field

[0001] This invention relates to the fields of credit risk assessment, machine learning and transfer learning, specifically to a method, system, equipment and medium for credit evaluation of science and technology innovation enterprises in the government sector. Background Technology

[0002] In the context of the rapid development of the digital economy, science and technology innovation enterprises are a key force driving technological innovation and industrial upgrading. Their credit evaluation is of great significance for government departments to formulate precise financial and industrial support policies. A scientific and accurate evaluation system can effectively identify high-quality science and technology innovation enterprises, helping government departments to rationally allocate fiscal subsidies and implement tax incentives and other resources.

[0003] However, government departments currently face numerous challenges in conducting credit evaluations of science and technology innovation enterprises. Credit evaluation of these enterprises requires a comprehensive sample of diverse characteristics and labels specific to science and technology innovation. However, due to limitations in business scenarios and data collection methods, the data accumulated in the daily management process of government departments suffers from insufficient features and small sample sizes, making it impossible to construct a comprehensive and accurate credit evaluation model. Furthermore, the technical application of credit evaluation for science and technology innovation enterprises in the government sector mainly relies on traditional expert scoring methods and simple indicator weighting models. Expert scoring methods depend on the subjective judgment of a few experts in the field, resulting in inconsistent scoring standards and significant influence from personal experience and cognitive biases, making it difficult to guarantee the objectivity and consistency of evaluation results. While simple indicator weighting models quantify scores based on preset indicators, they lack in-depth analysis of the complex characteristics of science and technology innovation enterprises, and their weighting settings largely depend on business experience, failing to adapt to the dynamic development characteristics of this type of enterprise. This leads to significant deviations between the evaluation results and the actual credit status of science and technology innovation enterprises. In addition, some government departments have tried to introduce machine learning models to build a credit evaluation system for this type of enterprise. However, due to insufficient data accumulation and a small number of training samples, the model is prone to overfitting and has poor generalization ability, making it difficult to meet the accuracy and reliability requirements of credit evaluation for science and technology innovation enterprises in government scenarios. Summary of the Invention

[0004] The technical objective of this invention is to provide a method, system, device, and medium for credit evaluation of science and technology innovation enterprises in the government sector, in order to address the problems existing in the credit evaluation of science and technology innovation enterprises in the government sector, such as insufficient data and label accumulation, strong subjectivity of traditional evaluation methods, poor application effect of machine learning models, and inability to effectively utilize high-quality data resources from financial institutions.

[0005] The technical objective of this invention is achieved as follows: a method for credit evaluation of science and technology innovation enterprises in the government sector, the specific method of which is as follows:

[0006] Cross-domain data collaborative governance: Collect cross-domain multi-source heterogeneous data from the financial source domain and the government target domain, and perform standardized processing, enterprise entity identification association, and cross-domain wide table construction and synchronous update on the cross-domain multi-source heterogeneous data;

[0007] Feature engineering processing in the financial field: generating a financial feature matrix through four layers of feature filtering and three types of derivative operators on financial source domain data;

[0008] Source domain model training of dynamic coupling architecture: Adaptive regularization, phased learning, dynamic weight fusion and domain adaptive regularization constraints are achieved through the training of the "logistic regression-ensemble tree" dynamic coupling model.

[0009] Three-layer hybrid cross-domain transfer: Achieving three-layer hybrid cross-domain transfer through instance-level similarity weighting and pseudo-label generation, feature-level semantic mapping and filtering, and model-level parameter inheritance and local fine-tuning;

[0010] Training and optimization of the government target domain model: The training and optimization of the government target domain model is achieved through secondary screening of government domain features, reuse and fine-tuning of model parameters, dual-objective optimization mechanism and incremental learning sliding window triggering.

[0011] As a preferred approach, cross-domain data collaborative governance is implemented as follows:

[0012] Collect multi-source heterogeneous data from the financial source domain of the financial sector and the government target domain of the government sector, respectively.

[0013] By formulating governance standards for financial source domain data and government target domain data, and writing automated scripts to automatically execute development tasks, multi-source heterogeneous data is governed into standardized data and stored in the standard layer. Among them, the data jointly contained in the financial source domain and government target domain in the standard layer includes basic enterprise information, intellectual property information, and R&D investment information.

[0014] Based on the standard layer data, entity identifiers are defined for enterprises, and entity tables in various databases are merged based on these entity identifiers to form a wide table. A record is created for each enterprise entity, and one-to-many relationships are stored in a dictionary format.

[0015] Set up a scheduled task between the standard layer and the wide table to ensure that the wide table data is updated in a timely manner.

[0016] As a preferred approach, feature engineering in the financial field is as follows:

[0017] Based on the wide table data of the financial source domain, and considering the characteristics of financial data, a four-layer filtering mechanism is constructed in the feature selection stage. The first layer uses the Pearson correlation coefficient to remove features whose correlation with the default label is lower than a certain threshold. The second layer uses the variance inflation factor (VIF) to filter multicollinear features. The third layer uses the feature importance score pre-trained by the XGBOOST model to retain core features. At the same time, the fourth layer introduces mutual information entropy to evaluate the redundancy between features and removes features whose information gain is lower than a threshold.

[0018] In the feature derivation stage, three types of derivation operators are designed: time-series feature derivation operators, ratio feature derivation operators, and relationship network feature derivation operators. Among them, the time-series feature derivation operator is used to derive features from annual and monthly data across cycles; the relationship network feature derivation operator is used to identify beneficiaries and actual controllers of enterprises through equity penetration, and to perform graph calculation and identification on the characteristics of enterprise shareholding relationships and operating relationships, ultimately forming a feature matrix.

[0019] As a preferred option, the training of the source domain model of the dynamically coupled architecture is as follows:

[0020] Source domain model training phase: The "logistic regression-XGBOOST dynamic coupling architecture" is adopted to achieve efficient transfer of financial source domain knowledge to government target domain under the transfer learning framework;

[0021] Adaptive L1 regularization is introduced through the logistic regression model, with the regularization coefficient λ(t) = λ0·exp(-γ·t); where t represents the number of iterations and γ represents the decay rate; the regularization strength gradually increases with training iterations, enhancing the sparsity of the model while preserving key features of the source domain; the parameter C is determined to its optimal value through cross-validation grid search.

[0022] The XGBOOST model employs a phased learning strategy. The first 50 rounds use a learning rate of 0.2 for rapid convergence, while the rate is reduced to 0.05 for fine-tuning in the next 150 rounds. Pruning techniques and limiting tree depth are used to avoid overfitting. At the same time, the "feature importance alignment" technique from transfer learning is introduced, which dynamically calculates the importance weights of source domain features during training and transfers them to the target domain model.

[0023] In the model fusion stage, a dual-model dynamic weight algorithm W = σ(β1·AUC_LR + β2·KS_XGB) is designed, where β1 and β2 are optimized through a genetic algorithm. At the same time, to improve the model's cross-domain generalization ability, a "domain adaptive regularization" technique is introduced. A KL divergence constraint term L_DA = α·KL(P_S(f(X_S))||P_T(f(X_T))) is added to the loss function to make the feature distribution learned by the model as close as possible to the potential distribution of the government target domain. At the same time, an interpretable rule set is generated to provide a knowledge source for the subsequent transfer of the model to government scenarios.

[0024] As a preferred option, the three-layer hybrid cross-domain migration refers to using an instance-feature-model three-layer hybrid migration architecture to achieve efficient migration of source domain knowledge from the financial domain to the target domain in the government domain;

[0025] Among them, instance-level transfer calculates the feature similarity between financial and government samples through K-nearest neighbors, assigns dynamic weights of 1.4-1.7 to source domain samples with high similarity, and uses the logistic regression-XGBOOST ensemble model trained on financial source domain data to predict unlabeled data in the government domain and generate pseudo labels for unlabeled samples in the government domain (threshold is 0.8).

[0026] Feature-level transfer establishes a cross-domain feature mapping matrix to map financial domain features to semantically similar features in the government domain. Based on the feature importance obtained from the source domain model training, the government domain features are screened, and features above a threshold are retained for training the target model.

[0027] The model-level transfer adopts a strategy of "parameter inheritance + local fine-tuning". It retains the logistic regression layer and XGBOOST basic tree structure parameters in the source domain model, and fine-tunes the feature weights that are strongly correlated with government evaluation to optimize the model's adaptability to government scenarios, in response to the business needs of the government sector.

[0028] More ideally, the training and optimization of the government target domain model is based on transferred features and samples. A "logistic regression-XGBOOST architecture model" is constructed on government data, and the accuracy and discriminative power of credit scoring are improved through hierarchical training and dynamic optimization, as detailed below:

[0029] A second screening of the transferred features was performed using 5-fold cross-validation, retaining features that are highly correlated with government credit labels to form a feature matrix.

[0030] During the model training phase, the logistic regression layer parameters of the source domain model are reused (with the first 80% weights frozen), and only the weights of government-specific features are fine-tuned. The XGBOOST model is trained for 50 rounds with a learning rate of 0.05, and the tree depth is limited to 4 to reduce the risk of overfitting.

[0031] In the model optimization stage, a dual-objective dynamic adjustment mechanism is introduced: specific indicators for government affairs scenarios are established, and the science and technology innovation attribute (0.2) is combined with the traditional accuracy indicator (0.8) to form a composite loss function; incremental learning and updating are carried out, and a sliding window mechanism (window size 6 months) is designed. When new government affairs data accumulates to 500 records, the model parameters are fine-tuned through "hot start" to retain historical knowledge while adapting to the dynamic changes of enterprises.

[0032] The model output adopts a segmented mapping method, which maps the credit score of science and technology innovation enterprises to the government credit score range of 300-850 points based on the default probability after transfer fine-tuning, so as to meet the needs of the government sector for credit evaluation of science and technology innovation enterprises.

[0033] A credit evaluation system for science and technology innovation enterprises in the government sector, the system includes:

[0034] The data governance module is used to collect diverse and heterogeneous data from the financial source domain and the government target domain, execute cross-domain data governance standards through automated scripts, generate standardized data containing enterprise entity identifiers, and construct a cross-domain wide table containing a one-to-many relationship dictionary based on the entity identifiers and multi-source database entity tables, and maintain the synchronous update of the wide table data through scheduled tasks.

[0035] The financial feature engineering module is used to perform four layers of feature filtering on the wide table data of the financial source domain: removing low-correlation features through correlation analysis, filtering redundant features through collinearity test, selecting core features through model pre-training, and evaluating feature redundancy through mutual information entropy. It also designs three types of operators: time series derivation, ratio derivation, and relationship network derivation to generate a financial feature matrix.

[0036] The source domain model training module is used to train the source domain model using a dynamically coupled architecture of "logistic regression-XGBOOST model". In the logistic regression stage, adaptive regularization with decay over iteration is introduced, and cross-validation is used to optimize the sparsity and key feature retention of the model. The XGBOOST model adopts a phased learning strategy and overfitting avoidance techniques, and introduces "feature importance alignment" technology to dynamically transfer source domain feature weights. Through dynamic weight fusion algorithm and domain adaptive regularization constraints, a cross-domain generalizable interpretable rule set is generated.

[0037] The cross-domain migration module is used to achieve cross-domain migration through a three-tier hybrid migration strategy;

[0038] The target model optimization module is used to perform cross-validation secondary screening on the migration features to build a "logistic regression-XGBOOST model" adapted for government affairs; and reuse the core parameters of the source domain model to fine-tune the weights of government affairs-specific features; at the same time, it introduces a dual-objective optimization mechanism for government affairs scenarios to adapt to the dynamic changes of enterprises through incremental learning.

[0039] As a preferred approach, the three-layer hybrid transfer strategy includes instance-level transfer, feature-level transfer, and model-level transfer. Instance-level transfer refers to assigning weights to highly matched source domain samples based on similarity metrics and using the source domain model to generate high-confidence pseudo-labels for unlabeled government data. Feature-level transfer refers to constructing a semantic mapping relationship between financial and government features and selecting government target domain features based on the importance of source domain features. Model-level transfer refers to inheriting the basic parameters of the source domain model and performing local parameter fine-tuning on strongly correlated government features.

[0040] An electronic device includes: a memory and at least one processor;

[0041] The memory contains computer programs;

[0042] The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the credit evaluation method for science and technology innovation enterprises in the government sector as described above.

[0043] A computer-readable storage medium storing a computer program that can be executed by a processor to implement the credit evaluation method for science and technology innovation enterprises in the government sector as described above.

[0044] The credit evaluation method, system, equipment, and medium for science and technology innovation enterprises in the government sector of this invention have the following advantages:

[0045] (I) This invention utilizes the scoring model of science and technology innovation enterprises successfully trained by financial institutions with rich scenario data. It applies transfer learning technology to transfer the knowledge accumulated in the training of the bank model to the scenario of government departments. Through source dataset collection, target dataset collection, data preprocessing, feature engineering, source model training, transfer strategy selection, target model training and optimization, etc., the data from different sources are adapted. Appropriate transfer strategies are used to handle the problem of unlabeled samples. The aim is to build a scoring model that can effectively utilize government department data and accurately evaluate the status of science and technology innovation enterprises, so as to provide a scientific and reliable reference for government department decision-making.

[0046] (II) This invention utilizes transfer learning technology to transfer relevant knowledge accumulated in the financial sector during the training of scoring models for science and technology innovation enterprises, including feature selection and model parameter training, to the government sector, addressing the challenges of limited data dimensions and small sample sizes in government departments. Simultaneously, through a series of methods including source dataset collection, target dataset collection, data preprocessing, feature engineering, source model training, transfer strategy selection, target model training, and optimization, multiple source data are adapted to the evaluation scenario in the government sector. Appropriate transfer strategies are employed to handle unlabeled samples, thereby constructing a scoring model that can effectively utilize government data and accurately assess the creditworthiness of science and technology innovation enterprises. This provides a scientific, reliable, and objective decision-making reference for government departments to formulate support policies and optimize resource allocation, improving the accuracy and effectiveness of credit evaluation for science and technology innovation enterprises in the government sector. Furthermore, in the government sector, computing power is often a very valuable resource, especially in areas such as identifying equity penetration and associated risks, and the iterative training of models, which consume significant computing power and prolong project cycles. Based on this invention, while addressing the scientific nature of the credit evaluation method for science and technology innovation enterprises in the government sector, it significantly reduces computing power consumption on the government side, improving the efficiency of government services.

[0047] (III) This invention addresses the current situation where credit evaluation of science and technology innovation entities in various provinces and cities still relies primarily on rule-based strategies due to insufficient effective input features and the inability to obtain and extract good and bad labels when using government big data for credit evaluation. Furthermore, this invention effectively improves the discriminative power of the credit evaluation model for science and technology innovation enterprises. Testing shows that the KS of the scoring card built based on the expert model is only 0.08, lacking the ability to distinguish between good and bad credit. (See attached figure.) Figure 1 As shown in the attached figure; however, by applying this invention, after training based on samples and features from the government sector, the model's discrimination index is greatly improved, enabling it to be applied to credit assessment and lending for science and technology innovation enterprises by financial institutions, as well as credit evaluation of science and technology innovation enterprises in the government sector, as shown in the attached figure. Figure 2 As shown. Attached Figure Description

[0048] The invention will be further described below with reference to the accompanying drawings.

[0049] Appendix Figure 1 A flowchart illustrating the credit evaluation method for science and technology innovation enterprises in the government sector;

[0050] Appendix Figure 2 This is a schematic diagram of the distinguishing ability indicators based on the expert experience scoring card model;

[0051] Appendix Figure 3 This is a schematic diagram of the distinguishing ability index of the credit rating model based on this method. Detailed Implementation

[0052] The following detailed description of the credit evaluation method, system, equipment, and medium for science and technology innovation enterprises in the government sector of the present invention is provided with reference to the accompanying drawings and specific embodiments.

[0053] Example 1:

[0054] As attached Figure 1 As shown in the figure, this embodiment provides a method for credit evaluation of science and technology innovation enterprises in the government sector. The method is as follows:

[0055] S1. Cross-domain data collaborative governance: Collect cross-domain multi-source heterogeneous data from the financial source domain and the government target domain, and perform standardized processing, enterprise entity identification association, and cross-domain wide table construction and synchronous update on the cross-domain multi-source heterogeneous data.

[0056] S2. Feature Engineering Processing in the Financial Domain: Generate a financial feature matrix through four-layer feature filtering and three types of derivative operators on financial source domain data;

[0057] S3. Dynamically Coupled Architecture Source Domain Model Training: Through the training of the "logistic regression-ensemble tree" dynamic coupling model, adaptive regularization, phased learning, dynamic weight fusion, and domain adaptive regularization constraints are achieved.

[0058] S3, Three-layer hybrid cross-domain transfer: Achieve three-layer hybrid cross-domain transfer through instance-level similarity weighting and pseudo-label generation, feature-level semantic mapping and filtering, and model-level parameter inheritance and local fine-tuning;

[0059] S4. Training and optimization of the government target domain model: The training and optimization of the government target domain model is achieved through secondary screening of government domain features, reuse and fine-tuning of model parameters, dual-objective optimization mechanism and incremental learning sliding window triggering.

[0060] The cross-domain data collaborative governance in step S1 of this embodiment is as follows:

[0061] S101. Collect multi-source heterogeneous data from financial source domain data in the financial field and government target domain data in the government field respectively;

[0062] S102. By formulating governance standards for financial source domain data and government target domain data, and writing automated scripts to automatically execute development tasks, multi-source heterogeneous data is governed into standardized data and stored in the standard layer; among which, the data jointly contained in the standard layer by the financial source domain and the government target domain includes basic enterprise information, intellectual property information, and R&D investment information.

[0063] S103. Based on the standard layer data, formulate entity identification tags for enterprises, and merge the entity tables in each database based on the entity identification tags to form a wide table; and form a record for each enterprise entity. For one-to-many relationships, store them in a dictionary manner.

[0064] S104. Set up a scheduled task between the standard layer and the wide table to ensure that the wide table data is updated in a timely manner.

[0065] The financial domain feature engineering process in step S2 of this embodiment is as follows:

[0066] S201. Based on the wide table data of the financial source domain, and targeting the data characteristics of the financial field, a four-layer filtering mechanism is constructed in the feature selection stage. The first layer uses the Pearson correlation coefficient to remove features whose correlation with the default label is lower than a certain threshold. The second layer uses the variance inflation factor (VIF) to filter multicollinear features. The third layer uses the feature importance score pre-trained by the XGBOOST model to retain core features. At the same time, the fourth layer introduces mutual information entropy to evaluate the redundancy between features and removes features whose information gain is lower than the threshold.

[0067] S202. In the feature derivation stage, three types of derivation operators are designed: time-series feature derivation operators, ratio feature derivation operators, and relationship network feature derivation operators. Among them, the time-series feature derivation operator is used to derive features from annual and monthly data across cycles; the relationship network feature derivation operator is used to identify beneficiaries and actual controllers of enterprises through equity penetration, and to perform graph calculation and identification on the characteristics of enterprise shareholding relationships and operating relationships, ultimately forming a feature matrix.

[0068] The training of the dynamically coupled architecture source domain model in step S3 of this embodiment is as follows:

[0069] S301, Source Domain Model Training Stage: The "Logistic Regression-XGBOOST Dynamic Coupling Architecture" is adopted to achieve efficient transfer of financial source domain knowledge to government target domain under the transfer learning framework;

[0070] S302. Adaptive L1 regularization is introduced through the logistic regression model, with the regularization coefficient λ(t) = λ0·exp(-γ·t); where t represents the number of iterations and γ represents the decay rate; the regularization strength gradually increases with training iterations, enhancing the sparsity of the model while preserving key features of the source domain; the parameter C is determined by cross-validation grid search to determine the optimal value;

[0071] The S303 and XGBOOST models employ a phased learning strategy. The first 50 rounds use a learning rate of 0.2 for rapid convergence, while the rate is reduced to 0.05 for fine-tuning in the next 150 rounds. Pruning techniques and limiting tree depth are used to avoid overfitting. At the same time, the "feature importance alignment" technique from transfer learning is introduced, which dynamically calculates the importance weights of source domain features during training and transfers them to the target domain model.

[0072] S304. In the model fusion stage, a dual-model dynamic weight algorithm W = σ(β1·AUC_LR + β2·KS_XGB) is designed, where β1 and β2 are optimized through a genetic algorithm. At the same time, to improve the model's cross-domain generalization ability, a "domain adaptive regularization" technique is introduced. A KL divergence constraint term L_DA = α·KL(P_S(f(X_S))||P_T(f(X_T))) is added to the loss function to make the feature distribution learned by the model as close as possible to the potential distribution of the government target domain. At the same time, an interpretable rule set is generated to provide a knowledge source for the subsequent model transfer to government scenarios.

[0073] In step S4 of this embodiment, the three-layer hybrid cross-domain migration refers to using an instance-feature-model three-layer hybrid migration architecture to achieve efficient migration of source domain knowledge from the financial domain to the target domain of the government domain.

[0074] Among them, instance-level transfer calculates the feature similarity between financial and government samples through K-nearest neighbors, assigns dynamic weights of 1.4-1.7 to source domain samples with high similarity, and uses the logistic regression-XGBOOST ensemble model trained on financial source domain data to predict unlabeled data in the government domain and generate pseudo labels for unlabeled samples in the government domain (threshold is 0.8).

[0075] Feature-level transfer establishes a cross-domain feature mapping matrix to map financial domain features to semantically similar features in the government domain. Based on the feature importance obtained from the source domain model training, the government domain features are screened, and features above a threshold are retained for training the target model.

[0076] The model-level transfer adopts a strategy of "parameter inheritance + local fine-tuning". It retains the logistic regression layer and XGBOOST basic tree structure parameters in the source domain model, and fine-tunes the feature weights that are strongly correlated with government evaluation to optimize the model's adaptability to government scenarios, in response to the business needs of the government sector.

[0077] In this embodiment, step S5 involves training and optimizing the government target domain model based on transferred features and samples. A "logistic regression-XGBOOST architecture model" is constructed on government data, and hierarchical training and dynamic optimization are used to improve the accuracy and discriminative power of credit scoring, as detailed below:

[0078] S501. Five-fold cross-validation is used to perform secondary screening on the transferred features, retaining features that are highly correlated with government credit labels to form a feature matrix.

[0079] S502. During the model training phase, the logistic regression layer parameters of the source domain model are reused (80% of the weights are frozen). Only the weights of government-specific features are fine-tuned. The XGBOOST model is trained for 50 rounds with a learning rate of 0.05 and the tree depth is limited to 4 to reduce the risk of overfitting.

[0080] S503. In the model optimization stage, a dual-objective dynamic adjustment mechanism is introduced: establish specific indicators for government affairs scenarios, combine the science and technology innovation attribute (0.2) with the traditional accuracy indicator (0.8) to form a composite loss function; and carry out incremental learning and updating, design a sliding window mechanism (window size 6 months), and when new government affairs data accumulates to 500, fine-tune the model parameters through "hot start" to retain historical knowledge while adapting to the dynamic changes of enterprises;

[0081] S504. The model output adopts a segmented mapping method. Based on the default probability after transfer fine-tuning, the credit score of science and technology innovation enterprises is mapped to the government credit score range of 300-850 points to meet the needs of the government sector for credit evaluation of science and technology innovation enterprises.

[0082] Example 2:

[0083] This embodiment provides a credit evaluation system for science and technology innovation enterprises in the government sector, the system comprising:

[0084] The data governance module is used to collect diverse and heterogeneous data from the financial source domain and the government target domain, execute cross-domain data governance standards through automated scripts, generate standardized data containing enterprise entity identifiers, and construct a cross-domain wide table containing a one-to-many relationship dictionary based on the entity identifiers and multi-source database entity tables, and maintain the synchronous update of the wide table data through scheduled tasks.

[0085] The financial feature engineering module is used to perform four layers of feature filtering on the wide table data of the financial source domain: removing low-correlation features through correlation analysis, filtering redundant features through collinearity test, selecting core features through model pre-training, and evaluating feature redundancy through mutual information entropy. It also designs three types of operators: time series derivation, ratio derivation, and relationship network derivation to generate a financial feature matrix.

[0086] The source domain model training module is used to train the source domain model using a dynamically coupled architecture of "logistic regression-XGBOOST model". In the logistic regression stage, adaptive regularization with decay over iteration is introduced, and cross-validation is used to optimize the sparsity and key feature retention of the model. The XGBOOST model adopts a phased learning strategy and overfitting avoidance techniques, and introduces "feature importance alignment" technology to dynamically transfer source domain feature weights. Through dynamic weight fusion algorithm and domain adaptive regularization constraints, a cross-domain generalizable interpretable rule set is generated.

[0087] The cross-domain migration module is used to achieve cross-domain migration through a three-tier hybrid migration strategy;

[0088] The target model optimization module is used to perform cross-validation secondary screening on the migration features to build a "logistic regression-XGBOOST model" adapted for government affairs; and reuse the core parameters of the source domain model to fine-tune the weights of government affairs-specific features; at the same time, it introduces a dual-objective optimization mechanism for government affairs scenarios to adapt to the dynamic changes of enterprises through incremental learning.

[0089] The three-layer hybrid migration strategy in this embodiment includes instance-level migration, feature-level migration, and model-level migration. Instance-level migration refers to assigning weights to highly matched source domain samples based on similarity metrics and using the source domain model to generate high-confidence pseudo-labels for unlabeled government data. Feature-level migration refers to constructing a semantic mapping relationship between financial and government features and selecting government target domain features based on the importance of source domain features. Model-level migration refers to inheriting the basic parameters of the source domain model and performing local parameter fine-tuning on strongly related government features.

[0090] Example 3:

[0091] This embodiment also provides an electronic device, including: a memory and a processor;

[0092] The memory stores the instructions executed by the computer.

[0093] The processor executes computer execution instructions stored in the memory, causing the processor to execute the credit evaluation method for science and technology innovation enterprises in the government sector according to any embodiment of the present invention.

[0094] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0095] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0096] Example 4:

[0097] This invention also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the credit evaluation method for science and technology innovation enterprises in the government sector according to any embodiment of this invention. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or apparatus can read and execute the program code stored in the storage medium.

[0098] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0099] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0100] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0101] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the credit of science and technology innovation enterprises in the government sector, characterized in that, The method is as follows: Cross-domain data collaborative governance: Collect cross-domain multi-source heterogeneous data from the financial source domain and the government target domain, and perform standardized processing, enterprise entity identification association, and cross-domain wide table construction and synchronous update on the cross-domain multi-source heterogeneous data; Feature engineering processing in the financial field: generating a financial feature matrix through four layers of feature filtering and three types of derivative operators on financial source domain data; Dynamically Coupled Architecture Source Domain Model Training: Through the training of the "logistic regression-ensemble tree" dynamic coupling model, adaptive regularization, phased learning, dynamic weight fusion, and domain adaptive regularization constraints are achieved; Three-layer hybrid cross-domain transfer: Achieving three-layer hybrid cross-domain transfer through instance-level similarity weighting and pseudo-label generation, feature-level semantic mapping and filtering, and model-level parameter inheritance and local fine-tuning; Training and optimization of the government target domain model: The training and optimization of the government target domain model is achieved through secondary screening of government domain features, reuse and fine-tuning of model parameters, dual-objective optimization mechanism and incremental learning sliding window triggering.

2. The credit evaluation method for science and technology innovation enterprises in the government sector according to claim 1, characterized in that, The specifics of cross-domain data collaborative governance are as follows: Collect multi-source heterogeneous data from the financial source domain of the financial sector and the government target domain of the government sector, respectively. By formulating governance standards for financial source domain data and government target domain data, and writing automated scripts to automatically execute development tasks, multi-source heterogeneous data is governed into standardized data and stored in the standard layer. Among them, the data jointly contained in the financial source domain and government target domain in the standard layer includes basic enterprise information, intellectual property information, and R&D investment information. Based on the standard layer data, entity identifiers are defined for enterprises, and entity tables in various databases are merged based on these entity identifiers to form a wide table. A record is created for each enterprise entity, and one-to-many relationships are stored in a dictionary format. Set up a scheduled task between the standard layer and the wide table to ensure that the wide table data is updated in a timely manner.

3. The credit evaluation method for science and technology innovation enterprises in the government sector according to claim 1, characterized in that, The specific details of feature engineering in the financial field are as follows: Based on the wide table data of the financial source domain, and considering the characteristics of financial data, a four-layer filtering mechanism is constructed in the feature selection stage. The first layer uses the Pearson correlation coefficient to remove features whose correlation with the default label is lower than a certain threshold. The second layer uses the variance inflation factor to filter multicollinearity features. The third layer retains core features through the feature importance score pre-trained by the XGBOOST model. Meanwhile, the fourth layer introduces mutual information entropy to evaluate the redundancy between features and removes features whose information gain is lower than a threshold. In the feature derivation stage, three types of derivation operators are designed: time-series feature derivation operators, ratio feature derivation operators, and relational network feature derivation operators. Among them, the time-series feature derivation operators are used to derive features from annual and monthly data across periods. The relational network feature derivation operator is used to identify beneficiaries and actual controllers of enterprises through equity penetration, and to perform graph calculation and identification on the characteristics of enterprise shareholding and operation relationships, ultimately forming a feature matrix.

4. The credit evaluation method for science and technology innovation enterprises in the government sector according to claim 1, characterized in that, The training of the source domain model of the dynamically coupled architecture is as follows: Source domain model training phase: The "logistic regression-XGBOOST dynamic coupling architecture" is adopted to achieve efficient transfer of financial source domain knowledge to government target domain under the transfer learning framework; Adaptive L1 regularization is introduced through the logistic regression model, with the regularization coefficient λ(t) = λ0·exp(-γ·t); where t represents the number of iterations and γ represents the decay rate; the regularization strength gradually increases with training iterations, enhancing the sparsity of the model while preserving key features of the source domain; the parameter C is determined to its optimal value through cross-validation grid search. The XGBOOST model employs a phased learning strategy. The first 50 rounds use a learning rate of 0.2 for rapid convergence, while the rate is reduced to 0.05 for fine-tuning in the next 150 rounds. Pruning techniques and limiting tree depth are used to avoid overfitting. At the same time, the "feature importance alignment" technique from transfer learning is introduced, which dynamically calculates the importance weights of source domain features during training and transfers them to the target domain model. In the model fusion stage, a dual-model dynamic weight algorithm W = σ(β1·AUC_LR + β2·KS_XGB) is designed, where β1 and β2 are optimized through a genetic algorithm. At the same time, to improve the model's cross-domain generalization ability, a "domain adaptive regularization" technique is introduced. A KL divergence constraint term L_DA = α·KL(P_S(f(X_S))||P_T(f(X_T))) is added to the loss function to make the feature distribution learned by the model as close as possible to the potential distribution of the government target domain. At the same time, an interpretable rule set is generated to provide a knowledge source for the subsequent transfer of the model to government scenarios.

5. The credit evaluation method for science and technology innovation enterprises in the government sector according to claim 1, characterized in that, Three-layer hybrid cross-domain migration refers to the use of an instance-feature-model three-layer hybrid migration architecture to achieve efficient migration of source domain knowledge from the financial domain to the target domain in the government domain; Among them, instance-level transfer calculates the feature similarity between financial and government samples through K-nearest neighbors, assigns dynamic weights of 1.4-1.7 to source domain samples with high similarity, and uses the logistic regression-XGBOOST ensemble model trained on financial source domain data to predict unlabeled data in the government domain and generate pseudo-labels for unlabeled samples in the government domain. Feature-level transfer establishes a cross-domain feature mapping matrix to map financial domain features to semantically similar features in the government domain. Based on the feature importance obtained from the source domain model training, the government domain features are screened, and features above a threshold are retained for training the target model. The model-level transfer adopts a strategy of "parameter inheritance + local fine-tuning". It retains the logistic regression layer and XGBOOST basic tree structure parameters in the source domain model, and fine-tunes the feature weights that are strongly related to government evaluation to optimize the model's adaptability to government scenarios, in response to the business needs of the government sector.

6. The credit evaluation method for science and technology innovation enterprises in the government sector according to any one of claims 1-5, characterized in that, The training and optimization of the government target domain model is based on transferred features and samples. A "logistic regression-XGBOOST architecture model" is constructed on government data. Through hierarchical training and dynamic optimization, the accuracy and discriminative power of credit scoring are improved, as detailed below: A second screening of the transferred features was performed using 5-fold cross-validation, retaining features that are highly correlated with government credit labels to form a feature matrix. During the model training phase, the logistic regression layer parameters of the source domain model are reused, and only the weights of government-specific features are fine-tuned. The XGBOOST model is trained for 50 rounds with a learning rate of 0.05 and the tree depth is limited to 4 to reduce the risk of overfitting. In the model optimization stage, a dual-objective dynamic adjustment mechanism is introduced: specific indicators for government affairs scenarios are established, and the scientific and technological innovation attributes are combined with traditional accuracy indicators to form a composite loss function; incremental learning and updates are carried out, and a sliding window mechanism is designed. When new government affairs data accumulates to 500 records, the model parameters are fine-tuned through a "hot start" method, retaining historical knowledge while adapting to the dynamic changes of enterprises. The model output adopts a segmented mapping method, which maps the credit score of science and technology innovation enterprises to the government credit score range of 300-850 points based on the default probability after transfer fine-tuning, so as to meet the needs of the government sector for credit evaluation of science and technology innovation enterprises.

7. A credit evaluation system for science and technology innovation enterprises in the government sector, characterized in that, The system includes: The data governance module is used to collect diverse and heterogeneous data from the financial source domain and the government target domain, execute cross-domain data governance standards through automated scripts, generate standardized data containing enterprise entity identifiers, and construct a cross-domain wide table containing a one-to-many relationship dictionary based on the entity identifiers and multi-source database entity tables, and maintain the synchronous update of the wide table data through scheduled tasks. The financial feature engineering module is used to perform four layers of feature filtering on the wide table data of the financial source domain: removing low-correlation features through correlation analysis, filtering redundant features through collinearity test, selecting core features through model pre-training, and evaluating feature redundancy through mutual information entropy. It also designs three types of operators: time series derivation, ratio derivation, and relationship network derivation to generate a financial feature matrix. The source domain model training module is used to train the source domain model using a dynamically coupled architecture of "logistic regression-XGBOOST model". In the logistic regression stage, adaptive regularization with decay over iteration is introduced, and cross-validation is used to optimize the sparsity and key feature retention of the model. The XGBOOST model adopts a phased learning strategy and overfitting avoidance techniques, and introduces "feature importance alignment" technology to dynamically transfer source domain feature weights. Through dynamic weight fusion algorithm and domain adaptive regularization constraints, a cross-domain generalizable interpretable rule set is generated. The cross-domain migration module is used to achieve cross-domain migration through a three-tier hybrid migration strategy; The target model optimization module is used to perform cross-validation secondary screening on the migration features to build a "logistic regression-XGBOOST model" adapted for government affairs; and reuse the core parameters of the source domain model to fine-tune the weights of government affairs-specific features; at the same time, it introduces a dual-objective optimization mechanism for government affairs scenarios to adapt to the dynamic changes of enterprises through incremental learning.

8. The credit evaluation system for science and technology innovation enterprises in the government sector according to claim 7, characterized in that, The three-layer hybrid transfer strategy includes instance-level transfer, feature-level transfer, and model-level transfer. Instance-level transfer refers to assigning weights to highly matched source domain samples based on similarity metrics and using the source domain model to generate high-confidence pseudo-labels for unlabeled government data. Feature-level transfer refers to constructing a semantic mapping relationship between financial and government features and selecting government target domain features based on the importance of source domain features. Model-level transfer refers to inheriting the basic parameters of the source domain model and performing local parameter fine-tuning on strongly related government features.

9. An electronic device, characterized in that, include: Memory and at least one processor; The memory contains computer programs; The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the credit evaluation method for science and technology innovation enterprises in the government sector as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the credit evaluation method for science and technology innovation enterprises in the government sector as described in any one of claims 1 to 6.

Images