Credit score card construction method, system and device and storage medium

Abstract

The invention belongs to the field of artificial intelligence and machine learning, particularly relates to a credit score card construction method, system and device and a storage medium, and aims to solve the problems that an existing credit evaluation model is difficult to process high-dimensional heterogeneous data and neural network training is unstable. The method comprises the steps of performing preprocessing and feature screening on multi-source heterogeneous credit data, and generating an initial population by adopting chaotic mapping improved by designing an adaptive disturbance strategy so as to improve diversity and uniformity during population initialization and optimize initial parameters of a neural network; a gradient estimation deviation correction mechanism is introduced, gradient noise interference in the training process is suppressed, and the convergence stability of the model is improved; and finally, the output of the neural network is converted into interpretable credit scores and risk levels. The method effectively solves the problems that when a traditional credit evaluation model processes high-dimensional heterogeneous data, the feature extraction capacity is insufficient, an optimization algorithm is prone to premature convergence, and the training process is unstable.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning, and specifically relates to a method, system, device and storage medium for constructing a credit scoring card. Background Technology

[0002] With the deepening of the digital inclusive finance strategy, commercial banks are continuously expanding the breadth and depth of their services to SMEs. Traditional credit risk assessment models are facing unprecedented challenges: data features exhibit significant high dimensionality, multiple sources, and unstructured complexity. Existing credit scoring systems are mainly based on traditional statistical learning methods such as logistic regression and decision trees. While these methods have advantages such as simple model structure and strong interpretability, they are essentially linear or shallow models, making it difficult to effectively capture the complex nonlinear relationships between multidimensional features. This severely restricts the accuracy of risk assessment.

[0003] To overcome the limitations of traditional models, researchers have begun exploring the application of artificial intelligence technologies such as deep learning in the field of credit risk assessment. Neural network models, with their powerful feature self-learning and nonlinear mapping capabilities, are theoretically better suited for handling high-dimensional heterogeneous data. However, in practical applications, deep neural networks face three major technical bottlenecks: First, the model training process is extremely sensitive to initial parameter settings; inappropriate initialization can lead to vanishing or exploding gradients. Second, the non-convexity of the loss function makes the optimization process prone to getting trapped in local optima, especially in high-noise, high-dimensional scenarios such as financial data. Finally, the optimization of hyperparameters (such as learning rate, batch size, and number of network layers) lacks a systematic methodology and often relies on trial and error.

[0004] To address the optimization problem in neural network training, researchers have attempted to introduce swarm intelligence algorithms. These algorithms simulate the behavior of biological groups in nature, achieving global optimization through information sharing and cooperation among individuals within the population. Among them, the Hippo Algorithm (HO), as an emerging swarm intelligence algorithm, demonstrates strong global exploration capabilities by simulating the social hierarchy and territorial defense mechanisms of a hippopotamus population. This algorithm performs excellently in low-dimensional optimization problems, but it reveals significant shortcomings when dealing with high-dimensional parameter optimization in neural networks: population diversity rapidly declines with increasing iterations, leading to premature convergence; and the parameter search lacks directional guidance, causing the convergence speed to decrease exponentially with increasing dimensionality.

[0005] On the other hand, at the level of neural network training algorithms, the Adam optimizer is widely used due to its adaptive learning rate characteristics. However, this algorithm has inherent defects in financial data applications: its first-order momentum estimation uses an exponential moving average method, which assigns too high a weight to recent gradients, resulting in gradient estimation bias when encountering abnormal fluctuations common in financial data; secondly, the learning rate decay strategy lacks adaptability to data characteristics and cannot effectively balance the contradiction between exploration and utilization.

[0006] At present, domestic and foreign research has not fully solved the above problems. There are three key technical bottlenecks: (1) The Hippo algorithm has poor population initialization quality and weak diversity maintenance ability in high-dimensional space, resulting in poor optimization effect of neural network parameters; (2) The gradient estimation bias problem of Adam optimizer in noisy environment affects the convergence stability of model; (3) The contradiction between black box model and interpretability requirements hinders the application of artificial intelligence technology in financial risk control.

[0007] Therefore, there is an urgent need for an innovative technical solution that can systematically address the optimization challenges faced by neural networks in credit risk assessment, while simultaneously meeting the financial industry's multiple requirements for model accuracy, stability, and interpretability. Summary of the Invention

[0008] To address the aforementioned problems in the prior art, namely the difficulty of existing credit assessment models in handling high-dimensional heterogeneous data and the instability of neural network training, this invention provides a method, system, device, and storage medium for constructing a credit scoring card.

[0009] In a first aspect, the present invention provides a method for constructing a credit scoring card, the method comprising:

[0010] Preprocessing and feature selection are performed on the multi-source heterogeneous raw credit data to form the feature set required for model training;

[0011] An initial population is generated using a chaotic mapping mechanism, where a random perturbation factor is introduced to perturb the chaotic sequence to eliminate the short-period phenomenon of chaotic mapping and improve the distribution uniformity of the population in the solution space; and an adaptive perturbation strategy based on hippo group behavior simulation is introduced to construct a swarm intelligence optimization algorithm; based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment to obtain an optimized parameter set;

[0012] An adaptive learning rate optimization algorithm with a gradient estimation bias correction mechanism is adopted. The LSTM network is initialized and trained using the optimization parameter set. The historical gradient weights are dynamically adjusted to suppress gradient noise interference during the training process, thereby obtaining a credit risk assessment model.

[0013] The feature data of the object to be evaluated is input into the trained credit risk assessment model to obtain the predicted output, and the predicted output is converted into an interpretable credit score and risk level through a mapping mechanism.

[0014] Furthermore, the multi-source heterogeneous raw credit data is preprocessed and its features are selected to form the feature set required for model training. The method is as follows:

[0015] The missing and outlier values ​​in the original credit data are identified and processed, including filling missing values ​​with the median and removing outliers.

[0016] A feature selection method based on variance filtering is used to select feature variables with predictive power higher than a preset threshold from the processed data;

[0017] The selected feature variables are fused and dimensionality reduced using principal component analysis, and the number of principal components is determined based on a preset cumulative variance contribution rate threshold to form the feature set.

[0018] Furthermore, a chaotic mapping mechanism is used to generate the initial population, the method of which is as follows:

[0019] A chaotic sequence with ergodicity and randomness is generated by iteratively generating a piecewise linear chaotic map, and the chaotic sequence is used to generate a uniformly distributed initial population in the solution space.

[0020] Furthermore, an adaptive perturbation strategy based on hippocampal group behavior simulation is introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment to obtain the optimization parameter set. The method is as follows:

[0021] The social hierarchy and territorial defense mechanisms of hippopotamus groups were simulated, and a population diversity index based on information entropy was calculated. The population diversity index was obtained by statistically analyzing the distribution density of individual hippopotamus in the solution space and calculating the relationship based on information entropy.

[0022] The strength of the perturbation term in the chaotic map is dynamically adjusted based on the value of the population diversity index.

[0023] When the population diversity index reaches or exceeds a preset aggregation threshold, the strength of the perturbation term is increased;

[0024] When the population diversity index is lower than a preset dispersion threshold, the intensity of the perturbation term is reduced;

[0025] The adaptive perturbation strategy is used to perform a global optimization search on the initial weights and biases, thereby obtaining the optimized parameter set.

[0026] Furthermore, a credit risk assessment model is obtained by dynamically adjusting historical gradient weights to suppress gradient noise interference during training. The method is as follows:

[0027] Real-time monitoring of gradient statistics during training, and calculation of the gradient variance at the current time step as an evaluation index of noise level.

[0028] A gradient bias correction factor is generated based on the gradient variance value through a negative correlation function, wherein the gradient bias correction factor adaptively decreases as the gradient variance increases.

[0029] The gradient bias correction factor is introduced into the momentum estimation process, and the contribution weight of historical gradients in parameter updates is dynamically adjusted through an adaptive scaling mechanism.

[0030] The trainable parameters of the LSTM network are updated using the backpropagation algorithm based on the bias-corrected momentum estimate.

[0031] The process of gradient variance monitoring, correction factor calculation, momentum weight adjustment and parameter update is executed iteratively until the model loss function converges to a stable state, and finally the optimized credit risk assessment model is obtained.

[0032] Furthermore, the predicted output is converted into an interpretable credit score and risk level through a mapping mechanism, the method of which is as follows:

[0033] Based on the statistical quantiles of historical default probability distribution, multiple mapping interval boundary points are determined, and a nonlinear transformation framework from default probability to credit score is established.

[0034] A multi-segment linear interpolation method is used to map the predicted default probability values ​​output by the LSTM network to a standardized credit score range according to the interval in which their values ​​lie.

[0035] Based on a preset risk threshold matrix, continuous credit scores are discretized into rating labels with clear risk implications.

[0036] Furthermore, when generating level labels, a complete scorecard system is also generated, including variable weight coefficients, segmented scoring rules, and decision boundary values.

[0037] A second aspect of the present invention provides a credit scoring card construction system, based on a credit scoring card construction method, the system comprising:

[0038] The feature set construction module is configured to preprocess and select features from multi-source heterogeneous raw credit data to form the feature set required for model training.

[0039] The parameter set generation module is configured to generate an initial population using a chaotic mapping mechanism. This involves introducing a random perturbation factor to perturb the chaotic sequence, thereby eliminating the short-period phenomenon of the chaotic mapping and improving the uniformity of the population distribution in the solution space. An adaptive perturbation strategy based on hippocampal group behavior simulation is also introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment, obtaining the optimized parameter set.

[0040] The credit risk assessment model acquisition module is configured to use an adaptive learning rate optimization algorithm with a gradient estimation bias correction mechanism to initialize the LSTM network and train it using the optimization parameter set, and to obtain the credit risk assessment model by dynamically adjusting the historical gradient weights to suppress gradient noise interference during the training process.

[0041] The scoring and grading module is configured to input the feature data of the object to be evaluated into the trained credit risk assessment model to obtain the predicted output, and convert the predicted output into an interpretable credit score and risk level through a mapping mechanism.

[0042] A third aspect of the present invention provides an electronic device comprising:

[0043] At least one processor; and

[0044] A memory communicatively connected to at least one of the processors; wherein,

[0045] The memory stores instructions that can be executed by the processor to implement a method for constructing a credit scoring card.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer instructions for execution by a computer to implement a method for constructing a credit scoring card.

[0047] The beneficial effects of this invention are:

[0048] (1) By preprocessing and feature screening of multi-source heterogeneous data, this invention effectively improves the model’s ability to extract features from high-dimensional complex data, enabling the model to better capture nonlinear feature relationships in the data, thereby significantly improving the accuracy of credit risk assessment.

[0049] (2) This invention enhances the global search capability of swarm intelligence optimization algorithm in high-dimensional parameter space through the synergistic effect of chaotic mapping mechanism and adaptive perturbation strategy, effectively avoids the problem of premature convergence of algorithm, improves the quality of neural network parameter initialization, and lays a good foundation for subsequent model training.

[0050] (3) By applying the gradient estimation bias correction mechanism, this invention effectively suppresses gradient noise interference during the training process, enhances the stability and convergence efficiency of model training, enables the neural network to converge to the optimal solution more quickly, and improves the reliability of the training process.

[0051] (4) By converting the model output into interpretable credit scores and risk levels, this invention maintains the high accuracy of deep learning models and meets the requirements of the financial field for the interpretability of model decision-making processes, enabling risk control personnel to better understand and use the model output results.

[0052] (5) The overall technical solution of the present invention has good practicality and scalability. The optimized feature processing flow reduces the computational complexity and improves the system operating efficiency, enabling it to better adapt to the actual application needs of various financial risk control scenarios. Attached Figure Description

[0053] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0054] Figure 1 This is an overall structural diagram of a method for constructing a credit scoring card according to the present invention;

[0055] Figure 2 This is a flowchart of a method for constructing a credit scoring card according to the present invention;

[0056] Figure 3 This is a schematic diagram of the structure of a computer system used to implement the methods, systems, and devices of this application. Detailed Implementation

[0057] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] The first embodiment of the present invention provides a method for constructing a credit scoring card, the method comprising:

[0060] Step S10: Preprocess and feature-select the multi-source heterogeneous raw credit data to form the feature set required for model training.

[0061] Step S20: An initial population is generated using a chaotic mapping mechanism. A random perturbation factor is introduced to perturb the chaotic sequence to eliminate the short-period phenomenon of chaotic mapping and improve the distribution uniformity of the population in the solution space. An adaptive perturbation strategy based on hippo group behavior simulation is introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment to obtain an optimized parameter set.

[0062] Step S30: An adaptive learning rate optimization algorithm with a gradient estimation bias correction mechanism is adopted to initialize the LSTM network and train it using the optimization parameter set. The historical gradient weights are dynamically adjusted to suppress gradient noise interference during the training process, thereby obtaining a credit risk assessment model.

[0063] Step S40: Input the feature data of the object to be evaluated into the trained credit risk assessment model to obtain the prediction output, and convert the prediction output into an interpretable credit score and risk level through a mapping mechanism.

[0064] To more clearly illustrate the method for constructing a credit scoring card according to the present invention, the following will be combined with... Figure 1 and Figure 2 The steps in the embodiments of the present invention are described in detail below, including steps S10-S40.

[0065] Step S10: Preprocess and feature-select the multi-source heterogeneous raw credit data to form the feature set required for model training.

[0066] In this embodiment, step S10 includes the following specific steps:

[0067] Step S11: Identify and process missing values ​​and outliers in the original credit data. The processing includes filling missing values ​​with the median and removing outliers.

[0068] Step S12: Use a feature selection method based on variance filtering to select feature variables with predictive power higher than a preset threshold from the processed data.

[0069] Step S13: The selected feature variables are fused and dimensionality reduced using principal component analysis, and the number of principal components is determined based on a preset cumulative variance contribution rate threshold to form the feature set.

[0070] Specifically, the process begins by identifying and processing missing values ​​in the multi-source, heterogeneous raw credit data from the banking system. This data includes, but is not limited to, time-series structured data such as whether a payment is overdue, number of employees, annual sales, registered capital, whether the entity is a listed company, average cumulative overdue interest periods, average consecutive overdue interest periods, and account credit limits. For numerical feature variables, median imputation is used to handle missing values. Specifically, the median of the feature across all samples is calculated and used to fill in the missing positions. For categorical feature variables, mode imputation is employed. Outlier handling utilizes statistical analysis-based methods. The mean μ and standard deviation σ of each feature variable are calculated, and data points exceeding the range [μ-3σ, μ+3σ] are identified as outliers and removed to ensure data quality meets subsequent modeling requirements. The entire process is implemented through a standardized workflow of a rule engine, ensuring consistency and repeatability in data processing.

[0071] A variance-based feature selection method is employed. First, the variance value of each feature variable is calculated, and a variance threshold is set. In this embodiment, the variance threshold is empirically set to 0.01. Feature variables with variance values ​​less than the threshold are eliminated, as these low-variance features typically contain less information. Based on this, a recursive feature elimination method is further used. By constructing a base model and iteratively removing the least important features, feature variables with predictive power higher than a preset threshold are selected. Simultaneously, the Pearson correlation coefficient between each feature and the target variable is calculated, and feature variables with absolute correlation coefficient values ​​greater than a preset threshold are retained. In this embodiment, a value preferably of 0.05 is used. This process ensures that the selected features have sufficient variability and maintain a significant correlation with the target variable, thereby improving the predictive performance of subsequent models.

[0072] Principal component analysis (PCA) is used for feature fusion and dimensionality reduction on the selected feature variables. First, the feature data is standardized so that each feature has a mean of 0 and a variance of 1. Then, the covariance matrix of the features is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​are sorted in descending order, and the cumulative variance contribution rate is calculated using the formula: Cumulative variance contribution rate = Σ(first k eigenvalues) / Σ(all eigenvalues). A cumulative variance contribution rate threshold of 85% is set. Through iterative calculation, the smallest value of k is selected so that the cumulative variance contribution rate first reaches or exceeds this threshold, thus determining the number of principal components. Finally, the original features are projected onto the selected principal component directions to form a dimensionality-reduced feature set. This method reduces feature dimensionality while ensuring that more than 85% of the information in the original data is retained, providing high-quality feature input for subsequent model training.

[0073] Step S20: An initial population is generated using a chaotic mapping mechanism. A random perturbation factor is introduced to perturb the chaotic sequence to eliminate the short-period phenomenon of chaotic mapping and improve the distribution uniformity of the population in the solution space. An adaptive perturbation strategy based on hippo group behavior simulation is introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment to obtain an optimized parameter set.

[0074] In the specific implementation of this invention, the deep learning model used for credit risk assessment employs a bidirectional Long Short-Term Memory (LSTM) network structure incorporating an attention mechanism. This network model is specifically composed of an input layer, a bidirectional LSTM layer, an attention mechanism layer, a fully connected layer, and an output layer connected sequentially. The network first receives a preprocessed and dimensionality-reduced feature set as input. Subsequently, the input data enters a two-layer bidirectional LSTM structure, each layer of which contains 64 hidden units. The purpose of this structure is to fully capture long-term dependencies and contextual information in the input sequence data from both forward and backward perspectives. The LSTM unit internally uses the hyperbolic tangent function as its activation function by default.

[0075] To further enhance the model's ability to focus on information at key time steps, an attention mechanism layer is introduced after the bidirectional LSTM layer. This attention layer, with 32 units, dynamically calculates and assigns importance weights to the hidden state outputs at different time steps, generating a weighted context vector as the layer's output. This weighted vector is then fed into a fully connected layer with 32 neurons for nonlinear transformation, using a linear rectified unit as its activation function to introduce nonlinear expressiveness into the model. Finally, an output layer with a single neuron, applying a sigmoid activation function, compresses the final output value to the range of 0 to 1 to represent the customer's default probability.

[0076] For network parameter initialization, the weight matrices of all LSTM and fully connected layers are set using the Xavier initialization method. This method adaptively adjusts the scale of the initial weights according to the input and output dimensions of each network layer, effectively promoting stable gradient flow and mitigating the vanishing or exploding gradient problem in the early stages of training. To prevent overfitting, a dropout layer is set before the fully connected layers, with a dropout ratio set to 0.3, meaning that a certain proportion of network nodes are randomly masked in each training iteration. For model training, binary cross-entropy is chosen as the loss function, which is suitable for binary classification tasks and can effectively measure the difference between the model's predicted default probability and the true label. Through the carefully designed structure, initialization strategy, and regularization techniques described above, the entire network ensures that it maintains strong feature learning capabilities while possessing good training stability and generalization performance when processing high-dimensional heterogeneous credit data.

[0077] In this embodiment, a chaotic mapping mechanism is used to generate the initial population. The method is as follows:

[0078] Step S21: A chaotic sequence with ergodicity and randomness is generated through piecewise linear chaotic mapping iteration, and the chaotic sequence is used to generate a uniformly distributed initial population in the solution space.

[0079] An adaptive perturbation strategy based on hippocampal group behavior simulation is introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment to obtain the optimization parameter set. The method is as follows:

[0080] Step S22: Simulate the social hierarchy and territorial defense mechanism of the hippopotamus group, and calculate the population diversity index based on information entropy. The population diversity index is obtained by statistically analyzing the distribution density of individual hippopotamus in the solution space and calculating the relationship based on information entropy.

[0081] Step S23: Dynamically adjust the strength of the perturbation term in the chaotic mapping according to the value of the population diversity index.

[0082] Step S24: When the population diversity index reaches or exceeds a preset aggregation threshold, the strength of the perturbation term is increased.

[0083] Step S25: When the population diversity index is lower than a preset dispersion threshold, reduce the intensity of the perturbation term;

[0084] Step S26: Through the adaptive perturbation strategy, complete the global optimization search of the initial weights and biases to obtain the optimized parameter set.

[0085] In this embodiment, the Tent chaotic map is used as the specific implementation of the piecewise linear chaotic map, and its mathematical expression is: when hour, ;when hour, .in, This represents the current value of the chaotic sequence at the nth iteration. This represents the next value of the chaotic sequence in the (n+1)th iteration.

[0086] First, an initial value x0 is randomly generated within the interval [0, 1]. A chaotic sequence of length N is generated through iterative calculation, where N is the population size. This sequence exhibits uniform traversal characteristics and good randomness. Then, the chaotic sequence is mapped to the parameter solution space of a neural network, with the parameter search range set as follows: The initial value of the i-th parameter is obtained through the formula. The calculations show that an initial population is generated that is uniformly distributed in the solution space, providing a high-quality initial solution for subsequent optimization algorithms. This is the lower bound of the search range for the i-th parameter of the neural network. This represents the upper bound of the search range for the i-th parameter of the neural network. These are the initial values ​​of the i-th parameter of the neural network generated through chaotic mapping;

[0087] To simulate the social hierarchy of a hippopotamus population, each individual in the optimization algorithm is treated as a hippopotamus. First, the solution space is divided into M grids, and the number of individuals in each grid is counted, calculating the distribution density of each grid. Then, the population diversity index is calculated based on information entropy theory, using the following formula: ;in This represents the distribution density of individuals in the solution space. This index effectively reflects the dispersion of the population; a larger value indicates a more uniform population distribution and better diversity, while a smaller value indicates a higher population aggregation and poorer diversity.

[0088] A population aggregation threshold is set. When the population diversity index reaches or exceeds this threshold, it indicates that the population aggregation is too high, and the intensity of chaotic perturbation needs to be increased to simulate the territorial expansion behavior of hippopotamus groups. The adjustment formula for the perturbation intensity is:

[0089] ,in, Based on the basic disturbance strength, The adjusted disturbance strength, The expansion coefficient is preferably 0.5 in this embodiment. By increasing the perturbation intensity, some individuals can escape the current local optimum, expanding the search range and improving the global exploration capability.

[0090] A population dispersion threshold is set. When the population diversity index falls below this threshold, it indicates that the population dispersion is too high, and the perturbation intensity needs to be reduced to simulate the resource optimization behavior of the hippopotamus population. The adjustment formula for the perturbation intensity is: By reducing the perturbation intensity, the algorithm can concentrate resources on fine-grained searches in promising regions, improving local development efficiency and convergence speed.

[0091] The adaptive perturbation strategy based on hippocampal group behavior simulation described above dynamically adjusts the search strategy during algorithm iteration. In each iteration, the diversity index of the current population is first calculated, then the perturbation intensity is adjusted based on the comparison between the index value and a threshold, and finally, the perturbation operation is applied to generate a new generation of population. This process is repeated until the termination condition is met (such as reaching the maximum number of iterations or the quality of the solution no longer significantly improves), ultimately obtaining the optimal initial weights and bias parameter set that minimizes the neural network loss function, completing the global optimization search process.

[0092] In this embodiment, another method for adjusting the disturbance intensity is provided, as follows:

[0093] Preset baseline, upper limit, and lower limit values ​​for disturbance intensity; preset population aggregation and dispersion thresholds;

[0094] The current population diversity index is compared with the threshold. When the index reaches or exceeds the aggregation threshold, a nonlinear enhancement coefficient is determined based on the excess ratio. The current intensity is multiplied by the coefficient and the smaller value between the coefficient and the upper limit is taken as the new intensity value.

[0095] When the index is below the dispersion threshold, a nonlinear attenuation coefficient is determined based on the proportion below the threshold. The current intensity is multiplied by this coefficient, and the larger of the result and the lower limit value is taken as the new intensity value.

[0096] Among them, the enhancement coefficient increases exponentially with the increase of the excess ratio, and the decay coefficient decreases logarithmically with the increase of the under-proportion; it also includes the correction of the adjustment results based on the trend of population diversity change: if the trend shows that diversity continues to deteriorate, the adjustment range is increased, and if the trend shows that it begins to improve, the adjustment range is decreased.

[0097] In implementing this scheme, the initial perturbation intensity is set to a baseline value of 0.1, an upper limit of 0.5, and a lower limit of 0.02. The population aggregation threshold is set at 0.85, and the dispersion threshold at 0.65. During algorithm iteration, the current population diversity index is calculated in real time using the information entropy formula. When the current population diversity index reaches or exceeds 0.85, the excess ratio a = (current index - 0.85) / 0.15 is calculated, and the nonlinear enhancement coefficient is calculated using an exponential function. The current disturbance intensity value is multiplied by k1, and the smaller value is taken as the actual disturbance intensity when compared with the upper limit value of 0.5. When the current population diversity index is detected to be below 0.65, the proportion below is calculated as a = (0.65 - current index) / 0.65. The nonlinear decay coefficient k2 = 1 / (1 + ln(1 + 10a)) is calculated using a logarithmic function. The current disturbance intensity value is multiplied by k2, and the larger value is taken as the actual disturbance intensity when compared with the lower limit value of 0.02. At the same time, based on the population diversity index values ​​of the last 5 iterations, the slope of the change trend is calculated using linear regression: if the slope of the change trend is < -0.1, it is determined to be a continuous deterioration, and the adjusted disturbance intensity value is multiplied by 1.2; if the slope of the change trend is > 0.1, it is determined to be an improvement, and the adjusted disturbance intensity value is multiplied by 0.8. All parameter adjustment processes are automatically executed in each iteration cycle, forming a closed-loop control.

[0098] Step S30: An adaptive learning rate optimization algorithm with a gradient estimation bias correction mechanism is adopted to initialize the LSTM network and train it using the optimization parameter set. The historical gradient weights are dynamically adjusted to suppress gradient noise interference during the training process, thereby obtaining a credit risk assessment model.

[0099] In this embodiment, a credit risk assessment model is obtained by dynamically adjusting historical gradient weights to suppress gradient noise interference during the training process. The method is as follows:

[0100] Step S31: Monitor the gradient statistical characteristics in real time during the training process and calculate the gradient variance value at the current time step as a noise level evaluation index.

[0101] Step S32: Generate a gradient bias correction factor based on the gradient variance value through a negative correlation function relationship, wherein the gradient bias correction factor adaptively decreases as the gradient variance increases;

[0102] Step S33: The gradient bias correction factor is introduced into the momentum estimation process, and the contribution weight of the historical gradient in the parameter update is dynamically adjusted through an adaptive scaling mechanism.

[0103] Step S34: Update the trainable parameters of the LSTM network using the backpropagation algorithm based on the momentum estimate after bias correction.

[0104] Step S35: Iteratively execute the above gradient variance monitoring, correction factor calculation, momentum weight adjustment and parameter update process until the model loss function converges to a stable state, and finally obtain the optimized credit risk assessment model.

[0105] Specifically, at each training time step t, i.e. after processing one mini-batch, the gradient variance of the current mini-batch is calculated. This is used as an evaluation index for noise level. The specific calculation method is as follows: first, calculate the gradient of each sample within the current mini-batch; then, calculate the variance of the gradient values ​​of all samples within the mini-batch. The formula for calculating the variance is... ,in, It is the gradient of a single sample. This variance is the mean of the gradients of all samples in the mini-batch, where n is the size of the mini-batch. This variance effectively reflects the volatility and uncertainty of the current gradient estimate.

[0106] Based on the gradient variance value calculated in step S31 A dynamic gradient bias correction factor is generated through a negative exponential function relationship. The calculation formula is as follows:

[0107] ;in, It is the preset original momentum coefficient. This is a negative attenuation coefficient hyperparameter used to adjust the sensitivity of the correction factor to the gradient variance. This formula ensures... and They are negatively correlated; that is, the larger the gradient variance, the greater the noise. The smaller the value, the more adaptive the decay of historical gradient weights becomes.

[0108] The dynamic correction factor calculated in step S32 This is introduced into the update process of the first-order momentum estimate. Specifically, it uses... Instead of the fixed coefficient β in the standard Adam algorithm, according to the formula To update the first-order momentum .in, It is the historical momentum of the previous time step. It is the mean gradient at the current time step. The core of this step is through... This adaptive scaling mechanism dynamically adjusts the historical gradients. The contribution weight in the current momentum estimate.

[0109] Using the first-order momentum estimate after deviation correction in step S33 Combining second-order momentum estimation, i.e., the exponential moving average of the squared gradient, with the same update method as standard Adam, the update amount of the current network parameters is calculated. Subsequently, the backpropagation algorithm is used to propagate this update amount back along the network, updating all trainable parameters (including weights and biases) of the LSTM network layer by layer.

[0110] The process of steps S31 to S34 is repeated iteratively. In each iteration, a new mini-batch of data is used, and gradient variance monitoring, correction factor calculation, momentum weight adjustment, and parameter update operations are performed sequentially. This iterative process continues until the loss function value of the model on the validation set no longer decreases significantly or the preset maximum number of iterations is reached, indicating that the model has converged to a stable state. Finally, a trained and optimized credit risk assessment model is obtained.

[0111] Step S40: Input the feature data of the object to be evaluated into the trained credit risk assessment model to obtain the prediction output, and convert the prediction output into an interpretable credit score and risk level through a mapping mechanism.

[0112] In this embodiment, the predicted output is converted into an interpretable credit score and risk level through a mapping mechanism, and the method is as follows:

[0113] Step S41: Based on the statistical quantiles of the historical default probability distribution, determine multiple mapping interval boundary points and establish a nonlinear transformation framework from default probability to credit score.

[0114] Step S42: Using a multi-segment linear interpolation method, the predicted default probability value output by the LSTM network is mapped to a standardized credit score range according to the interval in which its value lies.

[0115] Step S43: Discretize the continuous credit scores into level labels with clear risk meanings according to the preset risk threshold matrix, and generate a complete scoring card system including variable weight coefficients, segmented scoring rules and decision boundary values.

[0116] In this embodiment, the preprocessed and feature-engineered feature data of the SME customer to be evaluated is used as input and fed into the pre-trained credit risk assessment model, namely the aforementioned optimized LSTM network. After forward propagation, the network outputs a scalar value between 0 and 1, which is the customer's predicted output, representing its estimated default probability p. Subsequently, the mapping and transformation module is called to transform this probability value p into two intuitive outputs through a set of preset rules: one is a credit score within a standardized range (e.g., 300-850 points), and the other is a discrete risk level label, such as A, B, C, or D.

[0117] Before model deployment, the boundary points of the mapping interval are determined based on the true distribution of default probabilities of the training set samples. Specifically, the statistical quantiles of the true default probabilities (or model-predicted probabilities) of all samples in the training set are calculated. The choice of key quantiles depends on business requirements and typically includes a minimum value p. min 25th percentile p 25 The 50th percentile, also known as the median, is p.50 and maximum value p max These four quantiles will cover the entire probability range [p] min p max Divided into three intervals: [p] min p 25 ), [p 25 p 50 ), [p 50 p max These three intervals correspond to low, medium, and high levels of credit risk, respectively, thus constructing a nonlinear transformation framework that maps continuous probability to discrete risk intervals and finally to continuous credit scores.

[0118] A multi-segment linear interpolation method is used to map the predicted default probability p output by the neural network to a preset standard credit score range, such as 300-850. Its mathematical expression and calculation rules are as follows:

[0119] ;

[0120] This mapping rule ensures that the lower a customer's probability of default, the higher their credit score, thus achieving a monotonic, continuous conversion from default probability to credit score.

[0121] This embodiment defines a risk threshold matrix to discretize continuous credit scores into a finite number of risk levels. For example, scores in the range [300, 550) can be classified as high-risk (Level D), [550, 650) as medium-risk (Level C), [650, 750) as low-risk (Level B), and [750, 850] as excellent (Level A). Simultaneously, the system automatically generates a complete scoring card system, which is a structured document or data table that should include:

[0122] Variable weight coefficients: show the contribution or importance ranking of each feature (or principal component) to the credit score in the final model.

[0123] Piecewise scoring rules: Clearly list the piecewise linear mapping formula and its parameters in step S42.

[0124] Decision boundary values: Clearly record the specific score thresholds (e.g., 550, 650, 750) used to classify risk levels and the probability quantile boundaries (p) determined in step S41. min p 25 p 50 p max ).

[0125] This scoring card system makes the model's decision-making process completely transparent and rule-based, meeting the stringent requirements of financial regulatory agencies for the interpretability and auditability of the model.

[0126] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple variations are all within the protection scope of this invention.

[0127] A credit scoring card construction system according to a second embodiment of the present invention, based on a credit scoring card construction method, the system comprising:

[0128] The feature set construction module is configured to preprocess and select features from multi-source heterogeneous raw credit data to form the feature set required for model training.

[0129] The parameter set generation module is configured to generate an initial population using a chaotic mapping mechanism and introduce a swarm intelligence optimization algorithm based on an adaptive perturbation strategy simulated by hippopotamus groups. Based on the feature set, the initial weights and biases of the LSTM network used for credit risk assessment are globally optimized and searched to obtain the optimized parameter set.

[0130] The credit risk assessment model acquisition module is configured to generate an initial population using a chaotic mapping mechanism. This involves introducing a random perturbation factor to perturb the chaotic sequence, thereby eliminating the short-period phenomenon of the chaotic mapping and improving the uniformity of the population distribution in the solution space. An adaptive perturbation strategy based on hippocampal group behavior simulation is also introduced to construct a swarm intelligence optimization algorithm. Based on the feature set, the swarm intelligence optimization algorithm is used to perform a global optimization search on the initial weights and biases of the LSTM network used for credit risk assessment, obtaining an optimized parameter set.

[0131] The scoring and grading module is configured to input the feature data of the object to be evaluated into the trained credit risk assessment model to obtain a predicted output, and then convert the predicted output into an interpretable credit score and risk level through a mapping mechanism. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the system described above can be found in the corresponding processes of the foregoing method embodiments, and will not be repeated here.

[0132] It should be noted that the credit scoring card construction system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0133] An electronic device according to a third embodiment of the present invention includes:

[0134] At least one processor; and

[0135] A memory communicatively connected to at least one of the processors; wherein,

[0136] The memory stores instructions that can be executed by the processor to implement the above-described method for constructing a credit scoring card.

[0137] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are executed by the computer to implement the above-described method for constructing a credit scoring card.

[0138] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0139] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0140] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system for implementing the methods, systems, and devices of this application. Figure 3 The server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0141] like Figure 3 As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0142] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0143] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0144] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0146] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0147] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0148] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method of constructing a credit scorecard, characterized by, The method comprises: Preprocessing and feature screening of multi-source heterogeneous original credit data to form a feature set required for model training; An initial population is generated using a chaotic mapping mechanism, wherein a random disturbance factor is introduced to disturb the chaotic sequence to eliminate the short cycle phenomenon of chaotic mapping and improve the uniformity of the population distribution in the solution space; and an adaptive disturbance strategy based on the simulation of the group behavior of a hippo population is introduced to construct a swarm intelligence optimization algorithm; based on the feature set, the initial weights and biases of the LSTM network for credit risk assessment are globally optimized and searched using the swarm intelligence optimization algorithm to obtain an optimized parameter set; An adaptive learning rate optimization algorithm with gradient estimation bias correction mechanism is used to initialize the LSTM network using the optimized parameter set and train it, and by dynamically adjusting the historical gradient weight to suppress the gradient noise interference in the training process, a credit risk assessment model is obtained; The feature data of the object to be evaluated is input into the trained credit risk assessment model to obtain a prediction output, and the prediction output is converted into an interpretable credit score and risk level through a mapping mechanism.

2. The method of claim 1, wherein, The method for preprocessing and feature screening of multi-source heterogeneous original credit data to form a feature set required for model training is as follows: The missing values and abnormal values in the original credit data are identified and processed, and the processing includes filling the missing values with the median and removing the abnormal values; A feature selection method based on variance filtering is used to select feature variables with prediction ability higher than a preset threshold from the processed data; Principal component analysis is used to fuse and reduce the dimension of the selected feature variables, and the number of principal components is determined according to a preset cumulative variance contribution rate threshold to form the feature set.

3. The method of claim 1, wherein, The method for generating an initial population using a chaotic mapping mechanism is as follows: A chaotic sequence with ergodicity and randomness is generated by piecewise linear chaotic mapping iteration, and the initial population is generated uniformly in the solution space using the chaotic sequence.

4. The method of claim 1, wherein, An adaptive disturbance strategy based on the simulation of the group behavior of a hippo population is introduced to construct a swarm intelligence optimization algorithm; based on the feature set, the initial weights and biases of the LSTM network for credit risk assessment are globally optimized and searched using the swarm intelligence optimization algorithm to obtain an optimized parameter set, and the method is as follows: The social class structure and territory defense mechanism of a hippo population are simulated, and a population diversity index based on information entropy is calculated, which is obtained by calculating the relationship between the information entropy and the distribution density of hippo individuals in the solution space; The strength of the disturbance term in the chaotic mapping is dynamically adjusted according to the numerical value of the population diversity index; When the population diversity index reaches or exceeds a preset aggregation threshold, the strength of the disturbance term is increased; When the population diversity index is lower than a preset dispersion threshold, the strength of the disturbance term is reduced; Through the adaptive disturbance strategy, the global optimization search of the initial weights and biases is completed, and the optimized parameter set is obtained.

5. The method of claim 1, wherein, The method for obtaining a credit risk assessment model by dynamically adjusting the historical gradient weight to suppress the gradient noise interference in the training process is as follows: Real-time monitoring of gradient statistical characteristics in the training process, calculating the gradient variance value at the current time step as the noise level evaluation index; Based on the gradient variance value, a gradient deviation correction factor is generated through a negative correlation function relationship, wherein the gradient deviation correction factor is adaptively attenuated with the increase of the gradient variance; The gradient deviation correction factor is introduced into the momentum estimation process, and the contribution weight of the historical gradient in the parameter update is dynamically adjusted through an adaptive scaling mechanism; Based on the momentum estimation value after deviation correction, the trainable parameters of the LSTM network are updated using the back propagation algorithm; The gradient variance monitoring, correction factor calculation, momentum weight adjustment and parameter update processes are iteratively performed until the model loss function converges to a stable state, and an optimized credit risk evaluation model is obtained.

6. The method of claim 1, wherein, The predicted output is converted into an interpretable credit score and risk level through a mapping mechanism, and the method is as follows: Based on the statistical quantile of the historical default probability distribution, a plurality of mapping interval boundary points are determined, and a nonlinear transformation framework from default probability to credit score is established; A multi-segment linear interpolation method is used to map the default probability prediction value output by the LSTM network to the standardized credit score range according to the interval in which the value is located; According to the preset risk threshold matrix, the continuous credit score is discretized into level labels with clear risk implications.

7. The method of claim 6, wherein, When generating the level labels, a complete scoring card system including variable weight coefficients, segmented scoring rules and decision boundary values is also generated.

8. A system for constructing a credit scorecard based on the method for constructing a credit scorecard according to any one of claims 1 to 7, characterized in that, The system comprises: a feature set construction module configured to preprocess and feature screen the multi-source heterogeneous original credit data to form a feature set required for model training; an optimized parameter set generation module configured to generate an initial population using a chaotic mapping mechanism, wherein a random disturbance factor is introduced to disturb the chaotic sequence to eliminate the short period phenomenon of chaotic mapping and improve the uniformity of the population distribution in the solution space; and an adaptive disturbance strategy based on the simulation of the group behavior of a river horse population is introduced to construct a swarm intelligence optimization algorithm; based on the feature set, the initial weights and biases of the LSTM network used for credit risk evaluation are globally optimized and searched using the swarm intelligence optimization algorithm to obtain an optimized parameter set; a credit risk evaluation model acquisition module configured to use an adaptive learning rate optimization algorithm with gradient estimation deviation correction mechanism to initialize the LSTM network using the optimized parameter set and train the LSTM network, suppress gradient noise interference in the training process by dynamically adjusting the historical gradient weight, and obtain a credit risk evaluation model; a scoring and grading module configured to input the feature data of the object to be evaluated into the trained credit risk evaluation model to obtain a predicted output, and convert the predicted output into an interpretable credit score and risk level through a mapping mechanism.

9. An electronic device, comprising: It comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement the credit scoring card construction method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for being executed by the computer to implement the method for constructing a credit scorecard according to any one of claims 1-7.

Images