A Method and Apparatus for Multi-Source Heterogeneous Enterprise Credit Data Fusion Based on MCP

Abstract

This application relates to a method and apparatus for fusing multi-source heterogeneous corporate credit data based on MCP. The method includes: collecting multi-source heterogeneous corporate credit data to obtain multi-source data; dividing the multi-source data into multiple modules according to the data source; using the Lasso penalty algorithm to filter the original variables in each module, eliminating redundant variables by minimizing the censored quantile regression loss function, and outputting modular data; treating each modular data as an independent censored dataset, constructing a multi-source data integration model, the objective function of which includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term; using the CQRFabs iterative algorithm to solve the objective function, obtaining the quantile regression model, calculating the estimated default time at different quantile levels, integrating the results to obtain the corporate default time distribution, and correcting for censored data interference using the censored quantile loss function to output prediction results containing default time and distribution characteristics at different confidence levels.

Description

Technical Field

[0001] This application relates to the field of multi-source data fusion technology, and in particular to a method and apparatus for fusion of multi-source heterogeneous enterprise credit data based on MCP. Background Technology

[0002] Corporate credit assessment is a core component of financial risk control, supply chain management, and regulatory decision-making, and its accuracy directly impacts credit security and resource allocation efficiency. With the development of the digital economy, corporate credit data is becoming increasingly multi-sourced and heterogeneous: data sources encompass financial statements, government regulations, ESG ratings, supply chain transactions, and more; data structures include structured indicators, semi-structured text, and unstructured data. Efficiently integrating multi-source, heterogeneous data to accurately characterize corporate default risk has become a key challenge in the field of credit assessment.

[0003] In existing technologies, corporate credit scoring methods are mainly divided into two categories: one is classification models based on a single dataset, such as Logistic Regression and Support Vector Machine (SVM), which predict whether a default will occur through binary classification, but rely on balanced datasets and are difficult to handle high-dimensional and extremely imbalanced data; the other is survival analysis models, which can predict the timing of default, but are mostly based on a single data source and do not consider cross-domain data correlation.

[0004] The aforementioned technologies suffer from insufficient precision in high-dimensional variable screening, limited prediction of default timing, and the tendency of existing models to output single-point estimates, failing to characterize risk distribution at different confidence levels and thus hindering dynamic risk control decisions.

[0005] Based on this, this application provides a method and apparatus for fusing multi-source heterogeneous enterprise credit data based on MCP. Summary of the Invention

[0006] To address the issues of insufficient accuracy in high-dimensional variable screening, single-point prediction of default time, and the inability of existing models to output single-point estimates, thus failing to characterize risk distribution under different confidence levels and support dynamic risk control decisions, this application provides a method and apparatus for fusing multi-source heterogeneous enterprise credit data based on MCP.

[0007] Firstly, this application provides a method for fusing multi-source heterogeneous enterprise credit data based on MCP, which adopts the following technical solution, including:

[0008] Collect multi-source heterogeneous enterprise credit data to obtain multi-source data. Divide the multi-source data into multiple modules according to the data source. Each module contains original variables and corresponding enterprise identifiers, observation time, censoring status and standardized core variable matrix.

[0009] The Lasso penalty algorithm is used to filter the original variables in each module. By minimizing the censored quantile regression loss function with L1 norm penalty, the core variables that have a significant impact on default risk are retained, redundant variables are removed, and standardized modular data are output.

[0010] Each modular data is treated as an independent censored dataset, and a multi-source data integration model is constructed. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term.

[0011] The non-smoothness of the censored quantile regression loss function is handled by constructing a quadratic differentiable substitution function using the MM algorithm. Important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration. Standardized feature data of the enterprise to be evaluated are extracted from the standardized core variable matrix. The coefficient matrix is ​​combined with the standardized feature data to obtain the quantile regression model.

[0012] The estimated default time at different quantile levels is calculated using the quantile regression model, and the distribution results of corporate default time are integrated. The data interference is automatically corrected by censoring the quantile loss function in the objective function, and the final output is a prediction result containing the default time and distribution characteristics at different confidence levels.

[0013] Preferably, the Lasso penalty algorithm is used to filter the original variables in each module. By minimizing the censored quantile regression loss function with L1 norm penalty, core variables that have a significant impact on default risk are retained, redundant variables are eliminated, and standardized modular data is output, including:

[0014] Based on the original variables and their corresponding enterprise identifiers, observation times, and censoring states, an original variable matrix is ​​constructed, where the rows of the original variable matrix correspond to enterprise samples, the columns correspond to the original variables, and the elements are the observed values ​​of the original variables.

[0015] Define a censored quantile regression loss function with L1 norm penalty, and obtain the variable coefficient vector by minimizing the censored quantile regression loss function. Identify the original variables with non-zero coefficients in the variable coefficient vector as core variables and the original variables with zero coefficients as redundant variables to complete the variable screening.

[0016] The selected core variables are Z-score standardized to obtain standardized core variables. The standardized core variables, enterprise identifiers, time series observations, and censoring points are integrated to form the standardized modular data. The modular data includes enterprise identifiers, time series observations, censoring points, and a standardized core variable matrix. The rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to the standardized core variables.

[0017] Preferably, each modular data is treated as an independent censored dataset, and a multi-source data integration model is constructed. The objective function of the integration model includes: the censoring quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term, including:

[0018] Each of the modular data is labeled as an independent censored dataset. Each dataset includes: the standardized core variable matrix, the observation default time, and the censoring point, where the censoring point is the observation deadline under the right censoring state.

[0019] The quantile regression loss function for each module is as follows: For each independent censored dataset, the quantile loss of all enterprise samples is calculated and summed. The quantile loss function is as follows: When the error term is greater than or equal to 0, the loss is the product of the quantile level and the error term; when the error term is less than 0, the loss is the product of the difference between the quantile level and 1 and the error term. The quantile level is a confidence level parameter between 0 and 1.

[0020] Contrast penalty: For each core variable, pairwise comparisons are performed between different independent censored datasets. When the coefficients of the same core variable have the same sign in two independent censored datasets, a weighted penalty is applied to the squared difference between the two coefficients; when the signs are inconsistent, no penalty is applied to the difference in the coefficients of the core variable.

[0021] Global Lasso penalty term: Summing the absolute values ​​of the coefficients of all core variables in all the independent censored datasets and multiplying by the Lasso penalty parameter to achieve global variable selection across datasets and eliminate redundant core variables;

[0022] The censoring quantile regression loss function, the Contrast penalty term, and the global Lasso penalty term of all the independent censored datasets are added together to obtain the objective function that integrates the heterogeneity and correlation of multi-source data, wherein the variable coefficient vector set contains the core variable coefficients of all the independent censored datasets.

[0023] Preferably, the step of constructing a quadratic differentiable substitution function using the MM algorithm to handle the non-smoothness of the censored quantile regression loss function, selecting important variables and estimating the quantile regression coefficient matrix through forward-backward step-by-step iteration, extracting standardized feature data of the enterprise to be evaluated from the standardized core variable matrix, and combining the coefficient matrix with the standardized feature data to obtain the quantile regression model includes:

[0024] The MM algorithm is used to handle the non-smoothness of the censored quantile regression loss function: the non-smooth part of the objective function containing quantile loss is transformed into a quadratically differentiable substitution function. The substitution function is tangent to the censored quantile regression loss function at the current iteration point and its value is greater than or equal to that of the censored quantile regression loss function. The censored quantile regression loss function is approximately minimized by optimizing the substitution function.

[0025] Based on the substitution function, the important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration;

[0026] The rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to the standardized core variables. The standardized observations of the standardized core variables in the row of the enterprise to be evaluated are extracted to form the standardized feature data.

[0027] Using the coefficients of the core variables at different quantile levels in the quantile regression coefficient matrix as weights, the standardized feature data are weighted and summed to obtain the quantile regression model. The quantile regression model can output the estimated value of the default time point at different quantile levels based on the input standardized feature data.

[0028] Preferably, based on the substitution function, the quantile regression coefficient matrix is ​​obtained by selecting important variables and estimating them through forward-backward step-by-step iteration, including:

[0029] Based on the variable coefficient vector of the current substitution function iteration, the core variable that causes the objective function to decrease the most is selected and added to the model;

[0030] The variables selected for the integrated model are subjected to coefficient compression. Variables with coefficients close to zero are identified as redundant variables and removed. Through multiple iterations until the objective function converges, the quantile regression coefficient matrix is ​​obtained. The rows of the quantile regression coefficient matrix correspond to the quantile level, the columns correspond to the standardized core variables, and the elements are the coefficients of the standardized core variables at different quantile levels.

[0031] Preferably, the step of calculating the estimated default time at different quantile levels using the quantile regression model, integrating the results to obtain the corporate default time distribution, and automatically correcting for data interference by removing the quantile loss function in the objective function, finally outputs a prediction result containing the default time and distribution characteristics at different confidence levels, including:

[0032] Based on the coefficients of the core variables corresponding to different quantile levels in the quantile regression model as weights, the standardized feature data are weighted and summed to calculate the estimated value of the default time point at each quantile level.

[0033] By integrating the estimated default time values ​​at different quantile levels, a quantile curve of the corporate default time distribution is generated. In the quantile curve, the quantile level is used as the horizontal axis and the estimated default time value at the corresponding quantile level is used as the vertical axis. The curve is plotted to visually show the trend of default time with confidence level. The steeper the curve, the greater the difference in default time at different confidence levels.

[0034] Based on the estimated standard deviation of the quantile regression coefficient matrix, the confidence interval of the default time estimate at each quantile level is calculated: the fluctuation range of each default time estimate is determined by multiplying the coefficient estimated standard deviation by the statistic.

[0035] Based on the censoring point of the censoring quantile loss function, by minimizing the censoring quantile loss function, the quantile regression model distinguishes between defaulted samples and censored samples when estimating variable coefficients.

[0036] The prediction result is formed by integrating the quantile curves, the confidence intervals of each quantile point, and the corrected estimated default point. The prediction result includes the default point corresponding to different confidence levels, the distribution trend of default points, and the fluctuation range of each point.

[0037] Preferably, after calculating the estimated default time points at different quantile levels using the quantile regression model, integrating the results to obtain the corporate default time point distribution, and automatically correcting for data interference by removing quantile loss functions in the objective function, and finally outputting prediction results containing default time points and distribution characteristics at different confidence levels, the method further includes:

[0038] Based on the intermediate parameters of each module, the intermediate parameters include: the standardized core variables, the mean and standard deviation of the core variables, and the coefficients of single-module variables without multi-source integration. Each data provider deploys the model training process on the local node: loading the standardized core variable matrix and the intermediate parameters of the local module, training the model parameters based on the local data, and keeping the original data on the local node throughout the process without transmitting it to the outside.

[0039] After training is completed, each local node only uploads the model parameters to the central node. The model parameters include the quantile regression coefficient matrix, the Contrast penalty parameter, and the Lasso penalty parameter.

[0040] The central node adds dynamic Gaussian noise to the uploaded model parameters and adjusts the noise intensity according to the sensitivity level of the data corresponding to the model parameters, thereby masking the true values ​​of the original parameters through noise perturbation.

[0041] Each local node generates a zero-knowledge proof of the training process and submits it to the central node. The zero-knowledge proof includes the distribution characteristics of the local training data, the convergence index of iterative optimization, and the parameter update trajectory.

[0042] The validity of the zero-knowledge proof is verified by the central node, confirming that the training process of each local node conforms to the preset rules, and eliminating abnormal parameters that fail the verification.

[0043] The validated, noisy model parameters are integrated to form privacy-preserving fusion model parameters, which include quantile regression coefficients, standardized parameters, and penalty parameters after multi-source data co-optimization.

[0044] Secondly, this application discloses a multi-source heterogeneous enterprise credit data fusion device based on MCP, which adopts the following technical solution, including:

[0045] The data acquisition module is used to collect multi-source heterogeneous enterprise credit data to obtain multi-source data. The multi-source data is divided into multiple modules according to the data source. Each module contains original variables and corresponding enterprise identifiers, observation time, censoring status and standardized core variable matrix.

[0046] The variable selection module is used to select the original variables in each module using the Lasso penalty algorithm. By minimizing the censored quantile regression loss function with L1 norm penalty, it retains the core variables that have a significant impact on default risk, removes redundant variables, and outputs standardized modular data.

[0047] The data integration module is used to treat each modular data as an independent censored dataset and construct a multi-source data integration model. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term.

[0048] The regression model module is used to construct a quadratic differentiable substitution function using the MM algorithm to handle the non-smoothness of the censored quantile regression loss function, select important variables and estimate the quantile regression coefficient matrix through forward-backward step-by-step iteration, extract standardized feature data of the enterprise to be evaluated from the standardized core variable matrix, and combine the coefficient matrix with the standardized feature data to obtain the quantile regression model.

[0049] The credit prediction module is used to calculate the estimated default time at different quantile levels using the quantile regression model, integrate the results to obtain the distribution of corporate default time, and automatically correct for data interference by removing the quantile loss function in the objective function, finally outputting prediction results that include the default time and distribution characteristics at different confidence levels.

[0050] Thirdly, this application also provides a control device, the device comprising:

[0051] It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described above for the multi-source heterogeneous enterprise credit data fusion method based on MCP.

[0052] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the MCP-based multi-source heterogeneous enterprise credit data fusion method.

[0053] In summary, this application first divides the data into modules based on their sources, employs Lasso penalty to select and standardize core variables, and addresses the issues of data heterogeneity and high dimensionality. Second, it constructs a multi-source integration model, fusing the quantile loss functions of each module with censoring. Through Contrast penalty to coordinate cross-module variable correlations (sign-consistent coefficient convergence) and Lasso global variable selection, it captures the inherent relationships within the data. Third, it uses the MM algorithm to handle the non-smoothness of the loss function, combines forward-backward iterative optimization of the quantile regression coefficient matrix, and integrates it with enterprise feature data to establish a quantile regression model, outputting the distribution of default time points at different confidence levels. Finally, based on a federated learning framework, local training only transmits parameters, dynamic Gaussian noise is added to protect privacy, and zero-knowledge proofs are used to verify compliance, achieving data usability without visibility. This improves the accuracy of multi-source heterogeneous data fusion, thereby supporting dynamic risk assessment and providing technical support for enterprise credit evaluation. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a multi-source heterogeneous enterprise credit data fusion method based on MCP.

[0055] Figure 2 This is a structural block diagram of a multi-source heterogeneous enterprise credit data fusion device based on MCP. Detailed Implementation

[0056] The following combination Figure 1 - Figure 2 This application will be described in further detail.

[0057] Corporate credit assessment is a core component of financial risk control and resource allocation. With the development of the digital economy, credit data is becoming increasingly multi-sourced and heterogeneous, covering multiple fields such as finance, government affairs, ESG, and supply chain. The data structure includes structured indicators, semi-structured text, and unstructured logs, with a surge in variable dimensions and exhibiting high dimensionality, extreme imbalance, and right censoring. Existing multi-source data fusion technologies have significant shortcomings: insufficient heterogeneity handling leads to "information silos," difficulties in selecting high-dimensional variables cause redundant interference, lack of correlation mining limits prediction accuracy, and the conflict between privacy protection and data sharing is prominent.

[0058] To address the aforementioned issues, this invention provides a method for fusing multi-source heterogeneous enterprise credit data based on MCP, aiming to achieve efficient integration of multi-source data and accurate prediction of default timing, while ensuring data privacy and security.

[0059] Reference Figure 1 The embodiments of this application include at least steps S10 to S50.

[0060] S10: Collect multi-source heterogeneous enterprise credit data to obtain multi-source data. Divide the multi-source data into multiple modules according to the data source. Each module contains original variables and corresponding enterprise identifiers, observation time, censoring status, and standardized core variable matrix.

[0061] S20 uses the Lasso penalty algorithm to filter the original variables in each module. By minimizing the censored quantile regression loss function with L1 norm penalty, it retains the core variables that have a significant impact on default risk, removes redundant variables, and outputs standardized modular data.

[0062] S30 treats each modular data as an independent censored dataset and constructs a multi-source data integration model. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term.

[0063] S40 uses the MM algorithm to construct a quadratic differentiable substitution function to handle the non-smoothness of the censored quantile regression loss function. It then selects important variables and estimates the quantile regression coefficient matrix through forward-backward step-by-step iteration. Standardized feature data of the enterprise to be evaluated are extracted from the standardized core variable matrix, and the coefficient matrix is ​​combined with the standardized feature data to obtain the quantile regression model.

[0064] S50 calculates the estimated default time at different quantile levels using a quantile regression model, integrates the results to obtain the distribution of corporate default time, and automatically corrects for data interference by removing the quantile loss function in the objective function, finally outputting prediction results that include default time and distribution characteristics at different confidence levels.

[0065] Among them, multi-source heterogeneous corporate credit data refers to corporate credit-related data from different sources and in different formats, such as financial statements, government supervision, and ESG ratings; modules are independent data units divided according to data sources, containing original variables and metadata such as corporate identifiers and observation time; censoring status indicates whether a company defaulted during the observation period (right censoring indicates no default was observed); the standardized core variable matrix is ​​a matrix formed by Z-score standardization of important variables after Lasso screening (rows correspond to corporate samples, columns correspond to core variables); the Contrast penalty term is a weighted penalty for the difference in coefficients of the same variable across modules (consistent signs encourage coefficients to converge); the quantile regression coefficient matrix is ​​a set of coefficients of core variables at different confidence levels (quantile levels), used to calculate the distribution of default time points.

[0066] Specifically, the data is first divided into modules based on their sources. Lasso penalty is used to select core variables and standardize them to address the issues of high dimensionality and heterogeneity. Secondly, a multi-source integration model is constructed, which integrates the censored quantile loss functions of each module. Contrast penalty is used to coordinate cross-module variable correlations, and Lasso is used to achieve global variable selection. Then, the MM algorithm is used to handle the non-smoothness of the loss function. The quantile regression coefficient matrix is ​​optimized by forward-backward iteration, and a quantile regression model is constructed with enterprise feature data to output the distribution of default time points at different confidence levels.

[0067] In some embodiments, step S20 specifically includes the following steps: Constructing an original variable matrix based on the original variables and their corresponding enterprise identifiers, observation times, and censoring states, wherein the rows of the original variable matrix correspond to enterprise samples, the columns correspond to the original variables, and the elements are the observed values ​​of the original variables; defining a censoring quantile regression loss function with L1 norm penalty, obtaining the variable coefficient vector by minimizing the censoring quantile regression loss function, identifying original variables with non-zero coefficients as core variables, and identifying original variables with zero coefficients as redundant variables, thus completing variable screening; performing Z-score standardization on the screened core variables to obtain standardized core variables; integrating the standardized core variables, enterprise identifiers, time series observations, and censoring points to form standardized modular data, which includes enterprise identifiers, time series observations, censoring points, and a standardized core variable matrix, where the rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to the standardized core variables.

[0068] In practice, the quantile loss function is:

[0069] ;

[0070] Where u is the error term, reflecting the degree of deviation between the predicted value and the actual value;

[0071] Let the sample size be n, and the observed default time for the i-th firm sample be... The original variable vector is The variable coefficient vector is , which are parameters to be estimated, used to reflect the strength of the influence of each original variable on the default time.

[0072] The initial formula for the censored quantile regression loss function is as follows:

[0073] ;

[0074] These are censored and adjusted linear predictions. For error terms This reflects the deviation between the observed default time and the predicted value. This represents the quantile level.

[0075] To select core variables (remove redundant variables) from high-dimensional original variables, an L1 norm penalty (Lasso penalty) is introduced for the coefficient vector. The summation of absolute values ​​is penalized, forcing the coefficients of minor variables to be compressed to 0.

[0076] The complete formula for the censored quantile regression loss function is as follows:

[0077] ;

[0078] in, A value greater than 0 indicates a Lasso penalty parameter that controls the screening intensity. The larger the value, the more significant the coefficient compression, and the fewer core variables are selected. The coefficients are L1 norm (the sum of the absolute values ​​of the coefficient vectors), and the coefficients of redundant variables are compressed to 0, while the coefficients of core variables are non-zero.

[0079] By minimizing the regression loss function of the censored quantiles The optimal coefficient vector is obtained by solving the problem. If the optimal coefficient vector is 0, the corresponding original variable is a redundant variable and is removed from the model. If the optimal coefficient vector is not 0, the corresponding original variable is a core variable.

[0080] Specifically, in corporate credit assessment scenarios, various data sources contain a large number of raw variables, most of which are redundant information. Modular processing is used to achieve precise variable selection and standardization: first, the corporate samples and raw variables from each data source are organized into a matrix (rows correspond to companies, columns correspond to variables); then, Lasso-penalized censored quantile regression is used to automatically compress the coefficients of redundant variables to zero by minimizing the loss function with L1 norm penalty, retaining the core variables that significantly impact default risk; finally, the core variables are Z-score standardized, integrating them into modular data containing corporate identifiers, time series, censored status, and a standardized core variable matrix. This lays the foundation for subsequent cross-module data fusion and default prediction, effectively solving the problems of high-dimensional variable redundancy and data heterogeneity.

[0081] In some embodiments, step S30 specifically includes the following steps: Labeling each modular data as an independent censored dataset, each dataset containing: a standardized core variable matrix, observation default time, and censoring point, where the censoring point is the observation cutoff time under right censoring; Censored quantile regression loss function for each module: For each independent censored dataset, calculating and summing the quantile losses for all enterprise samples, where the quantile loss function is: when the error term is greater than or equal to 0, the loss is the product of the quantile level and the error term; when the error term is less than 0, the loss is the product of the difference between the quantile level and 1 and the error term, where the quantile level is a confidence level parameter between 0 and 1; Contrast penalty term: For each core variable, under different independent censoring... Pairwise comparisons are performed between datasets. When the coefficients of the same core variable have the same sign in two independent censored datasets, a weighted penalty is applied to the squared difference between the two coefficients. When the signs are inconsistent, no penalty is applied to the difference in the coefficients of the core variable. The global Lasso penalty term sums the absolute values ​​of the coefficients of all core variables in all independent censored datasets and multiplies them by the Lasso penalty parameter to achieve global variable selection across datasets and eliminate redundant core variables. The censoring quantile regression loss function, the Contrast penalty term, and the global Lasso penalty term of all independent censored datasets are added together to obtain the objective function that integrates the heterogeneity and correlation of multi-source data. The variable coefficient vector set contains the coefficients of the core variables in all independent censored datasets.

[0082] Specifically, in the context of corporate credit assessment, an integrated model is constructed to combine heterogeneity and correlation for multi-source modular data, including financial, government, and ESG data. First, the censored quantile regression loss is calculated for each module to capture the time distribution characteristics of defaults within a single module. Second, the correlation between cross-module variables is handled through Contrast penalty—when the coefficients of the same core variable have the same sign in different modules, a weighted penalty is applied to the squared differences in coefficients to encourage similarity, while retaining heterogeneity when there is a sign conflict. Finally, a global Lasso penalty is applied to sum and weight the absolute values ​​of the coefficients of all module variables, eliminating redundant variables that are significant only in a few modules. The sum of these three factors forms the objective function, achieving collaborative modeling of multi-source data.

[0083] Here, there are M independent censored datasets, and the variable coefficient vector of the m-th dataset is... , Let J be the coefficient of the j-th core variable in the m-th dataset. The core objective function for fusing multi-source data is:

[0084] ;

[0085] This represents the sum of the regression loss functions of the censored quantiles of each module. For Contrast penalty terms, Global Lasso penalty item, The penalty parameter (a non-negative real number, set through cross-validation or experience to control the penalty strength) is used. Indicates the characteristic function (the same variable j in datasets k and k). The value is 1 if the signs of the coefficients are the same, otherwise it is 0.

[0086] In some embodiments, step S40 specifically includes the following steps: Using the MM algorithm to handle the non-smoothness of the censored quantile regression loss function: transforming the non-smooth part of the objective function containing quantile loss into a quadratically differentiable substitution function, where the substitution function is tangent to the censored quantile regression loss function at the current iteration point and its value is greater than or equal to the censored quantile regression loss function; approximating the censored quantile regression loss function by optimizing the substitution function; based on the substitution function, selecting important variables and estimating the quantile regression coefficient matrix through forward-backward step-by-step iteration; the rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to standardized core variables; extracting the standardized observation values ​​of the standardized core variables in the row of the enterprise to be evaluated to form standardized feature data; using the coefficients of the core variables at different quantile levels in the quantile regression coefficient matrix as weights, weighted summing of the standardized feature data to obtain the quantile regression model; the quantile regression model can output the estimated default time at different quantile levels based on the input standardized feature data.

[0087] Furthermore, step S40 also includes the following steps: based on the variable coefficient vector of the current substitution function iteration, select the core variable that causes the objective function to decrease the most and add it to the model; compress the coefficients of the variables already selected into the integrated model, identify variables with coefficients close to zero as redundant variables and remove them, and iterate multiple times until the objective function converges to obtain the quantile regression coefficient matrix. The rows of the quantile regression coefficient matrix correspond to the quantile level, the columns correspond to the standardized core variables, and the elements are the coefficients of the standardized core variables at different quantile levels.

[0088] Specifically, in corporate credit assessment, for the non-smooth objective function (including quantile loss) of the multi-source data integration model, the MM algorithm is used to construct a quadratic differentiable substitution function (tangent to the original loss function at the iteration point and with a larger value), transforming the non-smooth optimization into a smooth approximation problem. Through forward-backward step-by-step iteration: the core variables that reduce the objective function the most are selected and added to the model in the forward direction, and redundant variables whose coefficients approach zero are compressed in the backward direction. The iteration is repeated until convergence, resulting in a quantile regression coefficient matrix (rows correspond to quantile levels τ, and columns correspond to core variables). Standardized feature data of the company to be assessed are extracted, and weighted summation is performed with the coefficient matrix to construct a quantile regression model, outputting the estimated default time value corresponding to different τ, thus realizing dynamic risk assessment.

[0089] The formula for prediction using the quantile regression model is: ; Here, X represents the time of default at the τ quantile, and X is the standardized feature data. Let τ be the coefficient vector corresponding to the quantile τ.

[0090] In some embodiments, step S50 specifically includes the following steps: using the coefficients of the core variables corresponding to different quantile levels in the quantile regression model as weights, weighted summation of standardized feature data is performed to calculate the estimated default time at each quantile level; the estimated default time at different quantile levels is integrated to generate a quantile curve of the corporate default time distribution, with the quantile level as the horizontal axis and the estimated default time at the corresponding quantile level as the vertical axis, to visually demonstrate the changing trend of default time with confidence level; the steeper the curve, the greater the difference in default time at different confidence levels; based on The estimated standard deviation of the quantile regression coefficient matrix is ​​used to calculate the confidence interval of the default time estimate at each quantile level. The fluctuation range of each default time estimate is determined by multiplying the coefficient estimation standard deviation by the statistic. Based on the censoring point of the censored quantile loss function, the quantile regression model is used to distinguish between defaulted samples and censored samples when estimating variable coefficients by minimizing the censored quantile loss function. The quantile curve, the confidence interval of each quantile time point, and the corrected default time estimate are integrated to form the prediction result. The prediction result includes the default time points corresponding to different confidence levels, the distribution trend of default time points, and the fluctuation range of each time point.

[0091] Specifically, in the corporate credit assessment scenario, firstly, the core variable coefficients corresponding to the quantile level τ are used as weights to weight the standardized characteristic data of the enterprise (weighted summation is performed to obtain the estimated default time at each τ). Next, a quantile curve is plotted with τ as the horizontal axis and the estimated values ​​as the vertical axis, visually demonstrating the trend of default time with confidence level (a steep curve indicates significant risk differences under different scenarios). Simultaneously, the confidence interval for each time point is calculated by combining the standard deviation of the coefficient estimates, reflecting the reliability of the results. A censored quantile loss function automatically distinguishes between defaulted samples and censored samples (enterprises whose defaults were not observed), avoiding estimation bias caused by censored data. Finally, the quantile curve, confidence intervals, and corrected estimates are integrated to form a prediction result that includes default time, distribution trends, and fluctuation ranges across multiple scenarios, providing a quantitative basis for dynamic risk control decisions such as bank loan term allocation and interest rate pricing.

[0092] In some embodiments, considering the principle of data confidentiality, the corresponding processing steps are as follows: Based on the intermediate parameters of each module, which include: standardized core variables, the mean and standard deviation of the core variables, and the coefficients of single-module variables without multi-source integration, each data provider deploys the model training process on its local node: loading the standardized core variable matrix and intermediate parameters of the local module, training the model parameters based on the local data, and keeping the original data on the local node throughout the process without transmitting it to the outside; after training, each local node only uploads the model parameters to the central node, which include the quantile regression coefficient matrix, Contrast penalty parameters, and Lasso penalty parameters; the central node processes the uploaded data... Dynamic Gaussian noise is added to the model parameters, and the noise intensity is adjusted according to the sensitivity level of the data corresponding to the model parameters, thus masking the true values ​​of the original parameters through noise perturbation. Each local node generates a zero-knowledge proof of the training process and submits it to the central node. The zero-knowledge proof includes the distribution characteristics of the local training data, the convergence index of iterative optimization, and the parameter update trajectory. The validity of the zero-knowledge proof is verified by the central node to confirm that the training process of each local node conforms to the preset rules and to remove abnormal parameters that fail the verification. The verified noisy model parameters are integrated to form privacy-preserving fusion model parameters, which include quantile regression coefficients, standardized parameters, and penalty parameters after multi-source data co-optimization.

[0093] Specifically, in the scenario of multi-source credit data fusion in financial institutions, to protect sensitive corporate data, each institution loads a standardized core variable matrix and intermediate parameters to train the model on its local node, ensuring that the original data remains within its domain throughout the process; only model parameters such as quantile regression coefficients and penalty parameters are uploaded to the central node to avoid leakage of the original data. The central node adds dynamic Gaussian noise according to the sensitivity level of the parameters to mask the true values; each node generates zero-knowledge proofs (including data distribution and iterative convergence indicators), and after the central node verifies compliance, it removes abnormal parameters, finally integrating the noisy parameters to form a fusion model.

[0094] The implementation principle of the multi-source heterogeneous enterprise credit data fusion method based on MCP in this application is as follows: First, the data is divided into modules according to its source. Lasso penalty is used to screen and standardize core variables, addressing the issues of data heterogeneity and high dimensionality. Second, a multi-source integration model is constructed, fusing the quantile loss functions of each module. Contrast penalty is used to coordinate cross-module variable correlation (sign consistency coefficient convergence) and Lasso global variable selection to capture the inherent relationships within the data. Then, the non-smoothness of the loss function is addressed using the MM algorithm, and the quantile regression coefficient matrix is ​​optimized through forward-backward iteration. This matrix is ​​then combined with enterprise feature data to build a quantile regression model, outputting the distribution of default time points at different confidence levels. Finally, based on a federated learning framework, local training only transmits parameters, dynamic Gaussian noise is added to protect privacy, and zero-knowledge proofs are used to verify compliance, achieving data usability without visibility. This improves the accuracy of multi-source heterogeneous data fusion, thereby supporting dynamic risk assessment and providing technical support for enterprise credit evaluation.

[0095] Figure 1 This is a flowchart illustrating a multi-source heterogeneous enterprise credit data fusion method based on MCP in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0096] Based on the same technical concept, referring to Figure 2 This application also provides a device for fusing multi-source heterogeneous enterprise credit data based on MCP, which adopts the following technical solution: The device includes:

[0097] The data acquisition module is used to collect multi-source heterogeneous enterprise credit data to obtain multi-source data. The multi-source data is divided into multiple modules according to the data source. Each module contains the original variables and the corresponding enterprise identifiers, observation time, censoring status and standardized core variable matrix.

[0098] The variable selection module is used to select the original variables in each module using the Lasso penalty algorithm. By minimizing the censored quantile regression loss function with L1 norm penalty, it retains the core variables that have a significant impact on default risk, removes redundant variables, and outputs standardized modular data.

[0099] The data integration module is used to treat each modular data as an independent censored dataset and build a multi-source data integration model. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term.

[0100] The regression model module is used to construct a quadratic differentiable substitution function using the MM algorithm to handle the non-smoothness of the censored quantile regression loss function. It selects important variables and estimates the quantile regression coefficient matrix through forward-backward step-by-step iteration. It extracts the standardized feature data of the enterprise to be evaluated from the standardized core variable matrix and combines the coefficient matrix with the standardized feature data to obtain the quantile regression model.

[0101] The credit prediction module is used to calculate the estimated default time at different quantile levels using a quantile regression model, integrate the results to obtain the distribution of corporate default time, and automatically correct for data interference by removing the quantile loss function in the objective function. Finally, it outputs prediction results that include the default time and distribution characteristics at different confidence levels.

[0102] In some embodiments, the variable filtering module is specifically used to construct an original variable matrix based on the original variables and their corresponding enterprise identifiers, observation times and censoring statuses, wherein the rows of the original variable matrix correspond to enterprise samples, the columns correspond to the original variables, and the elements are the observed values ​​of the original variables;

[0103] Define a censored quantile regression loss function with L1 norm penalty, and obtain the variable coefficient vector by minimizing the censored quantile regression loss function. Identify the original variables with non-zero coefficients in the variable coefficient vector as core variables and the original variables with zero coefficients as redundant variables, thus completing the variable selection.

[0104] The selected core variables are Z-score standardized to obtain standardized core variables. The standardized core variables, enterprise identifiers, time series observations, and censoring points are integrated to form standardized modular data. The modular data includes enterprise identifiers, time series observations, censoring points, and a standardized core variable matrix. The rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to standardized core variables.

[0105] In some embodiments, the data integration module is specifically used to label each modular data as an independent censored dataset. Each dataset includes: a standardized core variable matrix, observation default time, and censoring point, where the censoring point is the observation deadline in the right censored state.

[0106] Censored quantile regression loss function for each module: For each independent censored dataset, calculate and sum the quantile losses for all enterprise samples. The quantile loss function is as follows: When the error term is greater than or equal to 0, the loss is the product of the quantile level and the error term; when the error term is less than 0, the loss is the product of the difference between the quantile level and 1 and the error term. The quantile level is a confidence level parameter between 0 and 1.

[0107] Contrast penalty: For each core variable, pairwise comparisons are performed between different independent censored datasets. When the coefficients of the same core variable have the same sign in two independent censored datasets, a weighted penalty is applied to the squared difference between the two coefficients; when the signs are inconsistent, no penalty is applied to the difference in the coefficients of that core variable.

[0108] Global Lasso penalty term: Summing the absolute values ​​of the coefficients of all core variables across all independent censored datasets and multiplying by the Lasso penalty parameter to achieve global variable selection across datasets and eliminate redundant core variables;

[0109] The objective function that integrates the heterogeneity and correlation of multi-source data is obtained by summing the censored quantile regression loss function, the Contrast penalty term, and the global Lasso penalty term of all independent censored datasets. The set of variable coefficient vectors contains the core variable coefficients of all independent censored datasets.

[0110] In some embodiments, the regression model module is specifically used to handle the non-smoothness of the censored quantile regression loss function using the MM algorithm: the non-smooth part of the objective function containing quantile loss is transformed into a quadratically differentiable substitution function, the substitution function is tangent to the censored quantile regression loss function at the current iteration point and its value is greater than or equal to the censored quantile regression loss function, and the censored quantile regression loss function is approximately minimized by optimizing the substitution function;

[0111] Based on the substitution function, the important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration;

[0112] The rows of the standardized core variable matrix correspond to the enterprise samples, and the columns correspond to the standardized core variables. The standardized observations of the standardized core variables in the row of the enterprise to be evaluated are extracted to form standardized feature data.

[0113] By using the coefficients of the core variables at different quantile levels in the quantile regression coefficient matrix as weights, the standardized feature data are weighted and summed to obtain the quantile regression model. The quantile regression model can output the estimated default time at different quantile levels based on the input standardized feature data.

[0114] In some embodiments, the regression model module is also used to select the core variable that causes the objective function to decrease the most and add it to the model based on the variable coefficient vector of the current substitution function iteration;

[0115] The coefficients of the variables selected for integration are compressed, and variables with coefficients close to zero are identified as redundant and removed. Through multiple iterations until the objective function converges, the quantile regression coefficient matrix is ​​obtained. The rows of the quantile regression coefficient matrix correspond to the quantile level, the columns correspond to the standardized core variables, and the elements are the coefficients of the standardized core variables at different quantile levels.

[0116] In some embodiments, the credit prediction module is specifically used to perform weighted summation on standardized feature data based on the coefficients of the core variables corresponding to different quantile levels in the quantile regression model as weights, and to calculate the estimated default time value at each quantile level.

[0117] By integrating the estimated default time at different quantile levels, a quantile curve of the corporate default time distribution is generated. The quantile level is plotted on the horizontal axis, and the estimated default time at the corresponding quantile level is plotted on the vertical axis. The curve is plotted to visually show the trend of default time with confidence level. The steeper the curve, the greater the difference in default time at different confidence levels.

[0118] Based on the estimated standard deviation of the quantile regression coefficient matrix, the confidence interval of the default time estimate at each quantile level is calculated: the fluctuation range of each default time estimate is determined by multiplying the coefficient estimated standard deviation by the statistic.

[0119] Based on the censoring point of the censoring quantile loss function, by minimizing the censoring quantile loss function, the quantile regression model can distinguish between defaulted samples and censored samples when estimating variable coefficients.

[0120] The prediction results are formed by integrating the quantile curves, the confidence intervals of each quantile point, and the corrected default point estimates. The prediction results include the default point corresponding to different confidence levels, the distribution trend of default points, and the fluctuation range of each point.

[0121] In some embodiments, the credit prediction module is also used to deploy the model training process on local nodes based on the intermediate parameters of each module, including: standardized core variables, the mean and standard deviation of the core variables, and the coefficients of single-module variables without multi-source integration. The process involves loading the standardized core variable matrix and intermediate parameters of the local module, training the model parameters based on the local data, and keeping the original data on the local node throughout the process without transmitting it to the outside.

[0122] After training is complete, each local node only uploads the model parameters to the central node. The model parameters include the quantile regression coefficient matrix, the Contrast penalty parameter, and the Lasso penalty parameter.

[0123] The central node adds dynamic Gaussian noise to the uploaded model parameters and adjusts the noise intensity according to the sensitivity level of the data corresponding to the model parameters, thereby masking the true values ​​of the original parameters through noise perturbation.

[0124] Each local node generates a zero-knowledge proof of the training process and submits it to the central node. The zero-knowledge proof includes the distribution characteristics of the local training data, the convergence index of iterative optimization, and the parameter update trajectory.

[0125] The validity of zero-knowledge proofs is verified by the central node, confirming that the training process of each local node conforms to the preset rules, and eliminating abnormal parameters that fail the verification.

[0126] The validated noisy model parameters are integrated to form privacy-preserving fusion model parameters, which include quantile regression coefficients, standardized parameters, and penalty parameters after multi-source data co-optimization.

[0127] This application also discloses a control device.

[0128] Specifically, the control device includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed to perform the aforementioned MCP-based multi-source heterogeneous enterprise credit data fusion method.

[0129] This application also discloses a computer-readable storage medium.

[0130] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the MCP-based multi-source heterogeneous enterprise credit data fusion method described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for fusing multi-source heterogeneous enterprise credit data based on MCP, characterized in that, include: Collect multi-source heterogeneous enterprise credit data to obtain multi-source data. Divide the multi-source data into multiple modules according to the data source. Each module contains original variables and corresponding enterprise identifiers, observation time, censoring status and standardized core variable matrix. The Lasso penalty algorithm is used to filter the original variables in each module. By minimizing the censored quantile regression loss function with L1 norm penalty, the core variables that have a significant impact on default risk are retained, redundant variables are removed, and standardized modular data are output. Each modular data is treated as an independent censored dataset, and a multi-source data integration model is constructed. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term. The non-smoothness of the censored quantile regression loss function is handled by constructing a quadratic differentiable substitution function using the MM algorithm. Important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration. Standardized feature data of the enterprise to be evaluated is extracted from the standardized core variable matrix. The coefficient matrix is ​​combined with the standardized feature data to obtain the quantile regression model. The estimated default time at different quantile levels is calculated using the quantile regression model, and the distribution results of corporate default time are integrated. The data interference is automatically corrected by censoring the quantile loss function in the objective function, and the final output is a prediction result containing the default time and distribution characteristics at different confidence levels.

2. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 1, characterized in that, The Lasso penalty algorithm is used to filter the original variables in each module. By minimizing the censored quantile regression loss function with L1 norm penalty, core variables that have a significant impact on default risk are retained, redundant variables are removed, and standardized modular data is output, including: Based on the original variables and their corresponding enterprise identifiers, observation times, and censoring states, an original variable matrix is ​​constructed, where the rows of the original variable matrix correspond to enterprise samples, the columns correspond to the original variables, and the elements are the observed values ​​of the original variables. Define a censored quantile regression loss function with L1 norm penalty, and obtain the variable coefficient vector by minimizing the censored quantile regression loss function. Identify the original variables with non-zero coefficients in the variable coefficient vector as core variables and the original variables with zero coefficients as redundant variables to complete the variable screening. The selected core variables are Z-score standardized to obtain standardized core variables. The standardized core variables, enterprise identifiers, time series observations, and censoring points are integrated to form the standardized modular data. The modular data includes enterprise identifiers, time series observations, censoring points, and a standardized core variable matrix. The rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to the standardized core variables.

3. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 2, characterized in that, Each modular data set is treated as an independent censored dataset, and a multi-source data integration model is constructed. The objective function of the integration model includes: the censoring quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term, including: Each of the modular data is labeled as an independent censored dataset. Each dataset includes: the standardized core variable matrix, the observation default time, and the censoring point, where the censoring point is the observation deadline under the right censoring state. The quantile regression loss function for each module is as follows: For each independent censored dataset, the quantile loss of all enterprise samples is calculated and summed. The quantile loss function is as follows: When the error term is greater than or equal to 0, the loss is the product of the quantile level and the error term; when the error term is less than 0, the loss is the product of the difference between the quantile level and 1 and the error term. The quantile level is a confidence level parameter between 0 and 1. Contrast penalty: For each core variable, pairwise comparisons are performed between different independent censored datasets. When the coefficients of the same core variable have the same sign in two independent censored datasets, a weighted penalty is applied to the squared difference between the two coefficients; when the signs are inconsistent, no penalty is applied to the difference in the coefficients of the core variable. Global Lasso penalty term: Summing the absolute values ​​of the coefficients of all core variables in all the independent censored datasets and multiplying by the Lasso penalty parameter to achieve global variable selection across datasets and eliminate redundant core variables; The censoring quantile regression loss function, the Contrast penalty term, and the global Lasso penalty term of all the independent censored datasets are added together to obtain the objective function that integrates the heterogeneity and correlation of multi-source data, wherein the variable coefficient vector set contains the core variable coefficients of all the independent censored datasets.

4. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 3, characterized in that, The non-smoothness of the censored quantile regression loss function is addressed by constructing a quadratic differentiable substitution function using the MM algorithm. Important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration. Standardized feature data of the enterprise to be evaluated is extracted from the standardized core variable matrix, and the coefficient matrix is ​​combined with the standardized feature data to obtain the quantile regression model, including: The MM algorithm is used to handle the non-smoothness of the censored quantile regression loss function: the non-smooth part of the objective function containing quantile loss is transformed into a quadratically differentiable substitution function. The substitution function is tangent to the censored quantile regression loss function at the current iteration point and its value is greater than or equal to that of the censored quantile regression loss function. The censored quantile regression loss function is approximately minimized by optimizing the substitution function. Based on the substitution function, the important variables are selected and the quantile regression coefficient matrix is ​​estimated through forward-backward step-by-step iteration; The rows of the standardized core variable matrix correspond to enterprise samples, and the columns correspond to the standardized core variables. The standardized observations of the standardized core variables in the row of the enterprise to be evaluated are extracted to form the standardized feature data. Using the coefficients of the core variables at different quantile levels in the quantile regression coefficient matrix as weights, the standardized feature data are weighted and summed to obtain the quantile regression model. The quantile regression model can output the estimated value of the default time point at different quantile levels based on the input standardized feature data.

5. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 4, characterized in that, Based on the substitution function, the quantile regression coefficient matrix is ​​obtained by selecting important variables and estimating them through forward-backward step-by-step iterations, including: Based on the variable coefficient vector of the current substitution function iteration, the core variable that causes the objective function to decrease the most is selected and added to the model; The variables selected for the integrated model are subjected to coefficient compression. Variables with coefficients close to zero are identified as redundant variables and removed. Through multiple iterations until the objective function converges, the quantile regression coefficient matrix is ​​obtained. The rows of the quantile regression coefficient matrix correspond to the quantile level, the columns correspond to the standardized core variables, and the elements are the coefficients of the standardized core variables at different quantile levels.

6. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 4, characterized in that, The method involves calculating the estimated default time points at different quantile levels using the quantile regression model, integrating these estimates to obtain the distribution results of corporate default time points, and automatically correcting for data interference by removing quantile loss functions in the objective function. The final output includes prediction results containing default time points and distribution characteristics at different confidence levels, including: Based on the coefficients of the core variables corresponding to different quantile levels in the quantile regression model as weights, the standardized feature data are weighted and summed to calculate the estimated value of the default time point at each quantile level. By integrating the estimated default time values ​​at different quantile levels, a quantile curve of the corporate default time distribution is generated. In the quantile curve, the quantile level is used as the horizontal axis and the estimated default time value at the corresponding quantile level is used as the vertical axis. The curve is plotted to visually show the trend of default time with confidence level. The steeper the curve, the greater the difference in default time at different confidence levels. Based on the estimated standard deviation of the quantile regression coefficient matrix, the confidence interval of the default time estimate at each quantile level is calculated: the fluctuation range of each default time estimate is determined by multiplying the coefficient estimated standard deviation by the statistic. Based on the censoring point of the censoring quantile loss function, by minimizing the censoring quantile loss function, the quantile regression model distinguishes between defaulted samples and censored samples when estimating variable coefficients. The prediction result is formed by integrating the quantile curves, the confidence intervals of each quantile point, and the corrected estimated default point. The prediction result includes the default point corresponding to different confidence levels, the distribution trend of default points, and the fluctuation range of each point.

7. The method for fusing multi-source heterogeneous enterprise credit data based on MCP according to claim 1, characterized in that, After calculating the estimated default time points at different quantile levels using the quantile regression model, integrating the results to obtain the corporate default time point distribution, and automatically correcting for data interference by using the quantile loss function in the objective function to finally output the prediction results containing default time points and distribution characteristics at different confidence levels, the process further includes: Based on the intermediate parameters of each module, the intermediate parameters include: the standardized core variables, the mean and standard deviation of the core variables, and the coefficients of single-module variables without multi-source integration. Each data provider deploys the model training process on the local node: loading the standardized core variable matrix and the intermediate parameters of the local module, training the model parameters based on the local data, and keeping the original data on the local node throughout the process without transmitting it to the outside. After training is completed, each local node only uploads the model parameters to the central node. The model parameters include the quantile regression coefficient matrix, the Contrast penalty parameter, and the Lasso penalty parameter. The central node adds dynamic Gaussian noise to the uploaded model parameters and adjusts the noise intensity according to the sensitivity level of the data corresponding to the model parameters, thereby masking the true values ​​of the original parameters through noise perturbation. Each local node generates a zero-knowledge proof of the training process and submits it to the central node. The zero-knowledge proof includes the distribution characteristics of the local training data, the convergence index of iterative optimization, and the parameter update trajectory. The validity of the zero-knowledge proof is verified by the central node, confirming that the training process of each local node conforms to the preset rules, and eliminating abnormal parameters that fail the verification. The validated, noisy model parameters are integrated to form privacy-preserving fusion model parameters, which include quantile regression coefficients, standardized parameters, and penalty parameters after multi-source data co-optimization.

8. A device for fusing multi-source heterogeneous enterprise credit data based on MCP, characterized in that, The device includes: The data acquisition module is used to collect multi-source heterogeneous enterprise credit data to obtain multi-source data. The multi-source data is divided into multiple modules according to the data source. Each module contains original variables and corresponding enterprise identifiers, observation time, censoring status and standardized core variable matrix. The variable selection module is used to select the original variables in each module using the Lasso penalty algorithm. By minimizing the censored quantile regression loss function with L1 norm penalty, it retains the core variables that have a significant impact on default risk, removes redundant variables, and outputs standardized modular data. The data integration module is used to treat each modular data as an independent censored dataset and construct a multi-source data integration model. The objective function of the integration model includes: the censored quantile regression loss function of each module, the Contrast penalty term, and the Lasso penalty term. The regression model module is used to construct a quadratic differentiable substitution function using the MM algorithm to handle the non-smoothness of the censored quantile regression loss function, select important variables and estimate the quantile regression coefficient matrix through forward-backward step-by-step iteration, extract standardized feature data of the enterprise to be evaluated from the standardized core variable matrix, and combine the coefficient matrix with the standardized feature data to obtain the quantile regression model. The credit prediction module is used to calculate the estimated default time at different quantile levels using the quantile regression model, integrate the results to obtain the distribution of corporate default time, and automatically correct for data interference by removing the quantile loss function in the objective function, finally outputting prediction results that include the default time and distribution characteristics at different confidence levels.

9. A control device, characterized in that, The device includes: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images