Credit risk assessment method and device based on dynamic migration and incremental learning

Abstract

The invention provides a credit risk assessment method and device based on dynamic migration and incremental learning. The method comprises the steps of obtaining target domain credit data of a borrower; and inputting the target domain credit data into a trained target domain credit risk assessment model to obtain the default probability of the borrower output by the target domain credit risk assessment model. In the application, the target domain credit risk assessment model is obtained by carrying out transfer learning and incremental learning on the source domain credit risk assessment model. The knowledge migration weights of the source domain and the target domain are dynamically adjusted in the migration learning process according to the data volume and data distribution change condition of the target domain, so that the dependence on source domain knowledge is enhanced when the data of the target domain is less, the influence of the source domain is gradually reduced along with the accumulation of the data of the target domain, and the negative migration problem is prevented; and the long-term adaptability and generalization ability of the model are improved, so that a more accurate and efficient risk control solution can be provided for financial institutions.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of credit risk assessment, in particular to a credit risk assessment method and device based on dynamic migration and incremental learning. BACKGROUND

[0002] Credit risk assessment plays a vital role in the financial field, and it is a key link for financial institutions to effectively manage credit business and reduce the rate of non-performing loans. Accurate credit risk assessment can help financial institutions to rationally allocate credit resources, protect the safety of funds, and also help to maintain the stability and healthy development of the financial market. Credit risk assessment technology has mainly experienced two stages of traditional statistical models such as logistic regression and linear discrimination, and machine learning models such as decision tree and random forest, which establish a mathematical relationship between borrower characteristics and default probability to assess risk.

[0003] However, when facing new scenarios, traditional models are difficult to adapt to changes in data distribution, resulting in insufficient model generalization ability. In order to alleviate the problem of insufficient data, transfer learning technology is gradually introduced into the field of financial risk control, which transfers the knowledge of the source domain (a domain with a large amount of labeled data) to the target domain (a domain with less data) to improve the initial performance of the model. However, existing transfer learning methods mostly use fixed transfer weights, which may cause the model to excessively rely on source domain knowledge in the early stage of data, and still be affected by source domain information when the target domain data gradually becomes rich, thereby causing negative transfer risk and affecting the generalization ability of the model. In addition, the credit market is a dynamic environment, and the behavior patterns of borrowers will change over time, and existing models are prone to making wrong decisions when facing fraud or adversarial attacks, thereby increasing the risk of financial institutions. SUMMARY

[0004] The purpose of the present application is to provide a credit risk assessment method and device based on dynamic migration and incremental learning, which can improve the adaptability, generalization ability and security of the credit risk assessment model.

[0005] The first aspect of the present application provides a credit risk assessment method based on dynamic migration and incremental learning, comprising:

[0006] obtaining target domain credit data of a borrower;

[0007] inputting the target domain credit data into a trained target domain credit risk assessment model to obtain a default probability of the borrower output by the target domain credit risk assessment model;

[0008] The target domain credit risk assessment model is obtained by migration learning and incremental learning on the source domain credit risk assessment model, and in the migration learning process, the knowledge migration weights of the source domain and the target domain are dynamically adjusted according to the target domain data volume and data distribution change.

[0009] In a possible implementation, the training process of the target domain credit risk assessment model comprises:

[0010] The source domain modeling step is to construct a deep neural network as a source domain model S on the source domain data, to use Z-score standardization to pre-process the data of the source domain data, to obtain source domain data A, to train the source domain model S on the source domain data A, to identify key features in combination with an attention mechanism, to optimize the target to minimize the Binary Cross-Entropy loss, and to output the default probability of a borrower;

[0011] The target domain migration step is to extract the target domain data using the source domain model S after the source domain training is completed, to obtain a target domain model T1, in the migration process, to reduce the feature distribution difference between the source domain and the target domain by using the maximum mean difference loss, and to dynamically adjust the migration weights of the source domain and the target domain according to the target domain data volume and data distribution change; the dynamic adjustment process of the migration weight is to increase the weight of the source domain when the target domain data volume is small, and to gradually reduce the weight of the source domain as the target domain data gradually accumulates, so as to reduce the influence of negative migration;

[0012] The incremental learning step is to perform incremental learning on the target domain model T1 to obtain a target domain model T2, and to use the target domain model T2 as the target domain credit risk assessment model; in each incremental learning, only the target domain data of the first preset number of days is used, and the data of the first day is removed every day, so that the training set always maintains the data of the first preset number of days; if a piece of data is not updated for more than a second preset number of days, the piece of data is removed.

[0013] In a possible implementation, the target domain migration step specifically comprises:

[0014] According to the target domain data volume and data distribution change, the migration weights of the source domain and the target domain are dynamically adjusted by using an exponential decay function.

[0015] In a possible implementation, the training process of the target domain credit risk assessment model further comprises:

[0016] Adversarial training step: using the adversarial training method based on fast gradient sign method and projected gradient descent to train the target domain model T2, mixing the generated adversarial samples with the original samples for training during the training process, using the parameters of the target domain model T2 as the initial parameters, calculating the loss on the mini-batch and updating the parameters by back propagation to obtain the target domain model T3, taking the target domain model T3 as the target domain credit risk assessment model.

[0017] In a possible implementation, the training process of the target domain credit risk assessment model further includes:

[0018] Feedback optimization step: load the target domain model T3 and the source domain model S, copy the first n layers of the target domain model T3 to the corresponding layers of the source domain model S, freeze the migrated layers, and then set a learning rate less than a preset value on the source domain data A to fine-tune the source domain model S.

[0019] The second aspect of the application provides a credit risk assessment device based on dynamic migration and incremental learning, comprising:

[0020] The acquisition module is configured to acquire target domain credit data of a borrower.

[0021] The evaluation module is configured to input the target domain credit data into the trained target domain credit risk assessment model to obtain a default probability of the borrower output by the target domain credit risk assessment model.

[0022] The target domain credit risk assessment model is obtained by performing migration learning and incremental learning on a source domain credit risk assessment model, and the knowledge migration weight of the source domain and the target domain is dynamically adjusted according to the target domain data volume and the data distribution change during the migration learning process.

[0023] In a possible implementation, the training process of the target domain credit risk assessment model includes:

[0024] Source domain modeling step: constructing a deep neural network as a source domain model S on source domain data, performing data preprocessing on the source domain data using Z-score standardization to obtain source domain data A; training the source domain model S on the source domain data A, combining an attention mechanism to identify key features, and optimizing the target to minimize the Binary Cross-Entropy loss, and outputting the default probability of the borrower;

[0025] Target domain migration step: after the source domain is trained, the target domain data is extracted using the source domain model S to obtain a target domain model T1; in the migration process, the maximum mean difference loss is used to reduce the feature distribution difference between the source domain and the target domain, and meanwhile, the migration weight of the source domain and the target domain is dynamically adjusted according to the target domain data volume and the data distribution change; the dynamic adjustment process of the migration weight is as follows: when the target domain data volume is small, the weight of the source domain is increased; as the target domain data gradually accumulates, the weight of the source domain is gradually reduced to reduce the influence of negative migration.

[0026] Incremental learning step: the target domain model T1 is subjected to incremental learning to obtain a target domain model T2, and the target domain model T2 is taken as a target domain credit risk assessment model; in each incremental learning, only the target domain data of the first preset number of days is used, and the data of the first day is removed every day, so that the whole training set always maintains the data of the first preset number of days; if a piece of data is not updated for more than a second preset number of days, the piece of data is removed.

[0027] In a possible implementation, the target domain migration step specifically includes:

[0028] According to the target domain data volume and the data distribution change, the migration weight of the source domain and the target domain is dynamically adjusted by using an exponential decay function.

[0029] The third aspect of the present application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the method of the first aspect of the present application.

[0030] The fourth aspect of the present application provides a computer readable medium having computer readable instructions stored thereon, and the computer readable instructions are executable by a processor to implement the method of the first aspect of the present application.

[0031] Compared with the prior art, the credit risk assessment method, device, electronic device and storage medium based on dynamic migration and incremental learning provided by the present application obtain target domain credit data of a borrower; the target domain credit data is input into a trained target domain credit risk assessment model to obtain the default probability of the borrower output by the target domain credit risk assessment model. In the present application, the target domain credit risk assessment model is obtained by performing transfer learning and incremental learning on a source domain credit risk assessment model. In the transfer learning process, the knowledge transfer weight of the source domain and the target domain is dynamically adjusted according to the target domain data volume and the data distribution change, so that the dependence on the source domain knowledge is enhanced when the target domain data is small, and the influence of the source domain is gradually reduced as the target domain data accumulates, thereby preventing the negative migration problem and improving the long-term adaptability and generalization ability of the model, so that a more accurate and efficient risk control solution can be provided for a financial institution. Attached Figure Description

[0032] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0033] Figure 1 A flowchart of a credit risk assessment method based on dynamic transfer and incremental learning provided in this application is shown;

[0034] Figure 2 A schematic diagram illustrating the training process of the target domain credit risk assessment model provided in this application is shown.

[0035] Figure 3 A schematic diagram of a credit risk assessment device based on dynamic transfer and incremental learning provided in this application is shown. Detailed Implementation

[0036] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0037] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0038] Furthermore, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order. Additionally, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to those processes, methods, products, or apparatuses.

[0039] This application provides a credit risk assessment method and apparatus based on dynamic transfer and incremental learning, an electronic device, and a computer-readable storage medium, which will be described below with reference to the accompanying drawings.

[0040] Please refer to Figure 1 The document illustrates a flowchart of a credit risk assessment method based on dynamic transfer and incremental learning, provided by some embodiments of this application.Figure 1 As shown, the method may include the following steps:

[0041] S101. Obtain the borrower's target domain credit data;

[0042] S102. Input the target domain credit data into the trained target domain credit risk assessment model to obtain the borrower's default probability output by the target domain credit risk assessment model;

[0043] The target domain credit risk assessment model is obtained by transfer learning and incremental learning of the source domain credit risk assessment model. During the transfer learning process, the knowledge transfer weights of the source domain and the target domain are dynamically adjusted according to the changes in the amount and distribution of data in the target domain.

[0044] Figure 2 The diagram shown illustrates the training process of the target domain credit risk assessment model. Figure 2 As shown, the training process of the above-mentioned target domain credit risk assessment model includes:

[0045] 1. Source domain modeling steps: Construct a deep neural network (DNN) as the source domain model S on the source domain data (historical credit data). Perform data preprocessing on the source domain data using Z-score standardization to obtain source domain data A. Train the source domain model S on the source domain data A. Combine attention mechanism to identify key features. The optimization objective is to minimize the Binary Cross-Entropy loss. The output is the borrower's default probability.

[0046] In practical applications, the dataset originates from the Home Credit Default Risk competition on the Kaggle platform, containing approximately 300,000 records. Fields related to credit risk assessment include revolving credit utilization, debt ratio, number of delinquencies within 30-59 days, number of delinquencies exceeding 90 days, monthly income, age, and number of dependents. Categorical features include a severe delinquency label (yes / no). The dataset consists of application_train.csv and application_test.csv. application_train.csv, containing 150,000 records, is selected as the source domain data to train the source domain model and learn general credit risk assessment patterns. application_test.csv is used as the target domain data to simulate emerging credit scenarios and construct a low-resource transfer learning task.

[0047] Data preprocessing: Missing values ​​are imputed for numerical feature data. Median imputation is used for features such as income and loan amount. All numerical features are Z-score standardized to ensure that data of different dimensions are within the same scale range.

[0048] A deep neural network is built as the base model in the source domain. After data preprocessing, BatchNormalization is first used to normalize the numerical features, followed by multiple fully connected layers with ReLU as the activation function. A Self-Attention mechanism is also incorporated to enhance attention to key features (such as debt-to-income ratio and historical delinquency count). The model ultimately outputs the default probability of the source domain data. The training objective is to minimize the Binary Cross-Entropy loss function, while L2 regularization is added to prevent overfitting.

[0049] 2. Target Domain Transfer Steps: After training in the source domain, the source domain model S is used to extract features from the target domain data (application_test.csv). Maximum Mean Difference (MMD) loss is used to reduce the feature distribution difference between the source and target domains. Simultaneously, a dynamic transfer weight controller dynamically adjusts the transfer weights of the source and target domains based on changes in data distribution. When the target domain data volume is small, the dynamic transfer weight controller automatically increases the source domain weights to provide more transfer knowledge; as the target domain data gradually accumulates, the dynamic transfer weight controller gradually decreases the source domain weights to reduce the impact of negative transfer. This method ensures that when the target domain data volume is small, the source domain knowledge accounts for a higher proportion, and as data accumulates, the transfer weights automatically decrease, allowing the model to gradually adapt to the target domain data distribution and improve generalization ability.

[0050] In practical applications, the source domain data is denoted as X. s The target domain data is X t The hidden layers of the DNN model are used to extract the feature representations of both, and their mean vectors in the reproducing kernel Hilbert space are calculated respectively:

[0051]

[0052] Where φ(·) is the kernel function mapping to the high-dimensional feature space, and N s and N t These represent the number of samples in the source and target domains, respectively. Calculate the mean distance between the two distributions and minimize this distance to make the features of the source and target domains more similar.

[0053] L MMD =||μ s -μ t || 2 ;

[0054] This loss term will be added to the training process of the source domain model, so that the model can learn from the source domain data and adapt to the data distribution of the target domain.

[0055] To address the potential negative transfer problem, a dynamic transfer weight controller is introduced. The innovation lies in adjusting the transfer learning weights in real time based on the amount of data in the target domain, so that the influence of the source domain gradually decreases as the target domain data grows. Specifically, an exponential decay function is used to dynamically adjust the transfer weights of the source and target domains. The specific adjustment formula is as follows:

[0056]

[0057] Among them, W t N represents the influence weight of the source domain during the current transfer learning process. t Where N is the amount of data in the target domain, and α and β are hyperparameters that control the initial values ​​and decay rate of the transfer weights. When the target domain data is small (N... t When the value of Wt is relatively small, a larger value makes the model rely more on source domain knowledge, enhancing stability during the cold start phase, and vice versa. Combining the MMD loss with dynamically adjusted transfer weights and the source domain loss, the final optimization objective is:

[0058] L = L CE +W t λL MMD +α||θ|| 2 ;

[0059] Among them, L CE It is the binary cross-entropy loss of the source domain, which optimizes the default prediction ability of the source domain. ||θ|| 2 This is the L2 regularization term, used to prevent overfitting. λ and α are hyperparameters. After training, the target domain data is input into the model, and the DNN with aligned features outputs the default probability of the target domain.

[0060] Dynamic Transfer Weight Adjustment Mechanism: A dynamic transfer weight control method based on changes in data distribution is proposed. This method adjusts the transfer weight of source domain knowledge in real time according to the amount of data in the target domain and the differences in feature distribution. An exponential decay function is used for weight adjustment, which enhances the dependence on source domain knowledge when target domain data is scarce, while gradually reducing the influence of the source domain as target domain data accumulates. This prevents negative transfer problems and improves the model's adaptability and generalization ability.

[0061] 3. Incremental learning steps: Incrementally learn the target domain model T1 to obtain the target domain model T2, and use the target domain model T2 as the target domain credit risk assessment model; each time incremental learning is performed, only the target domain data of the most recent first preset number of days is used, updated once a day, and the data of the earliest day is removed, and the entire training set always maintains the data of the most recent first preset number of days; if a data point has not been updated for more than the second preset number of days, then the data point is removed.

[0062] For example, to adapt to the ever-changing credit market, each training session uses only the data from the most recent 30 days (first preset number of days) and removes data that has not been updated for more than 6 months (second preset number of days). To prevent catastrophic forgetting caused by the addition of new data, the Fisher information matrix is ​​calculated to measure the importance of each parameter to historical data, and a constraint term is added to the Elastic Weighted Co-location (EWC) loss to ensure that key parameters do not change drastically, thus ensuring that the model retains its memory of old data while learning new data.

[0063] In practical applications, incremental learning is used to ensure the model continuously adapts to new data and prevents catastrophic forgetting caused by the addition of new data. First, each incremental learning iteration uses only the most recent 30 days of data, updated daily, removing the oldest data day. The entire training set always maintains data from the most recent 30 days. Furthermore, any data point that hasn't been updated for more than 6 months is removed, ensuring the training data reflects current market conditions rather than past credit behavior. To prevent the model from forgetting historical data, a Fisher information matrix is ​​calculated to measure the importance of each model parameter to past data. Given model parameters θ, the Fisher information matrix F is defined as follows:

[0064]

[0065] Where P(y|x;θ) is the output probability of the model, F i Reflecting parameter θ i Importance to historical data. A higher value indicates a greater impact of the parameter on historical data, and therefore, it should be protected. The EWC loss function is:

[0066]

[0067] L CE It is the cross-entropy loss of the current data. These are the optimal parameters obtained in the previous training round, and λ is the balance factor. Constraint important parameter θ i Without significant changes, this loss function ensures that the model retains its memory of old data while learning new data.

[0068] Incremental learning mechanism: A sliding window method (30-day window) is used to dynamically manage data, ensuring that the model is always trained based on the latest market data. Combined with the EWC (Elastic Weights Combined) algorithm, the Fisher information matrix is ​​calculated to identify and protect key model parameters, preventing catastrophic forgetting. This allows the model to retain its memory of historical data while continuously learning new data, thereby improving long-term stability and adaptability.

[0069] 4. Adversarial Training Steps: To improve the robustness of the model, an adversarial training method based on Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) is used. During the training process, generated adversarial examples are mixed with original examples. The parameters of the target domain model T2 are used as initial parameters. The loss is calculated on the mini-batch and backpropagation is performed to update the parameters.

[0070] Specifically, to further enhance the model's robustness, prevent adversarial attacks, and improve its generalization ability, adversarial training is used after incremental learning for further optimization. The target domain data B used during training of the target domain model T2 is employed. During training, FGSM is used to generate adversarial examples x for each sample x. adv =x+δ, using PGD for multi-step gradient updates, training on adversarial examples and original samples simultaneously, loading the target domain model T2 as initial parameters during training, calculating the final loss and performing backpropagation for each mini-batch, updating the model parameters, and repeating the training multiple times until convergence.

[0071] 5. Feedback Optimization Steps: Load the target domain model T3 and the source domain model S. Copy the first n layers of parameters from T3 to the corresponding layers in S. To prevent excessive changes during training, freeze these layers. Then, set a low learning rate on the source domain data A and fine-tune it, ensuring that the feature representations of the source and target domains remain consistent during the transfer process. The learning rate is less than a preset value, which can be 0.005, 0.01, etc.

[0072] In other words, after the target domain model T3 is trained, feedback optimization is performed. The target domain model T3 and the source domain model S are loaded, the first n layers of the target domain model T3 are traversed, its parameters are copied to the corresponding layers of the source domain model S, the transfer layers are frozen, a low learning rate is set to prevent them from changing significantly during training, and fine-tuning is performed on the source domain data A.

[0073] In summary, the training process of the target domain credit risk assessment model can be composed of transfer learning, incremental learning, and adversarial training. First, information is collected and cleaned from historical credit data (source domain data A) to form a preliminary source domain dataset. Source domain data typically originates from existing financial institutions and banks, and is of high quality with complete annotations, used to train the basic credit risk assessment model. The source domain model S is trained using this data to extract effective features and construct the basic credit risk assessment model.

[0074] Next, knowledge learned from the source domain model S will be transferred to the target domain through transfer learning. In the target domain, new business scenarios (such as emerging markets or new financial platforms) provide a limited amount of target domain data B with incomplete labels; therefore, transfer learning is needed to adapt the knowledge from the source domain model to the target domain data. First, a preliminary target domain model T1 is generated by modeling the target domain data through basic training. To optimize the transfer learning process, a "dynamic transfer weight controller" mechanism is adopted. This controller can adjust the transfer learning weights in real time based on changes in the amount and distribution of target domain data. Then, through incremental learning, the model is trained in small batches, allowing it to gradually adapt to new data as the amount of target domain data increases, thereby saving computational resources and quickly adapting to market changes.

[0075] During the transfer learning process, adversarial training further enhances the robustness of the target domain model. Adversarial training generates adversarial examples using the Fast Gradient Sign Method (FGSM), adding these specially modified anomalous data samples to the training dataset. This forces the model to learn how to handle challenging data. These adversarial examples can simulate malicious attacks or anomalous behavior, helping to improve the model's ability to identify financial fraud or market manipulation, thus ensuring its stability.

[0076] Furthermore, to further improve the accuracy and robustness of the target domain model, the model is periodically optimized based on feedback. This entire process, through model updates, fine-tuning, and evaluation, forms a closed loop of continuous feedback and optimization, thereby ensuring the long-term effectiveness of the credit risk assessment model.

[0077] This application provides a credit risk assessment method based on dynamic transfer and incremental learning, which has the following characteristics:

[0078] Beneficial effects:

[0079] 1. Dynamic migration weight adjustment improves the model's generalization ability.

[0080] Traditional transfer learning methods typically use fixed weights for knowledge transfer between the source and target domains, which struggles to adapt to dynamic changes in the target domain data and can easily lead to negative transfer problems. This invention introduces a dynamic transfer weight control mechanism that adaptively adjusts the transfer learning weights based on changes in data distribution. When target domain data is scarce, the utilization of source domain knowledge is enhanced; conversely, as target domain data accumulates, the source domain weights are reduced to mitigate the risk of negative transfer, thereby improving the model's generalization ability and long-term adaptability.

[0081] 2. It adapted to market changes and prevented catastrophic amnesia.

[0082] Traditional models are often trained on static data, making it difficult to adapt to rapid market changes and prone to forgetting historical data during incremental training. This application adopts an incremental learning mechanism, dynamically updating the dataset through a sliding window strategy, and combining it with the EWC (Elastic Weights Combined) algorithm to identify and protect key model parameters, ensuring effective learning of new data without losing historical knowledge, thereby improving the long-term stability and adaptability of the model.

[0083] 3. Enhanced robustness against adversarial attacks and improved ability to identify anomalous data.

[0084] Existing technologies are easily affected by data noise or malicious attacks, leading to a decrease in prediction accuracy. This invention introduces an adversarial training method, generating perturbation samples through adversarial attacks such as FGSM and PGD. This allows the model to adapt to various abnormal data scenarios during training, improving its robustness in complex market environments, reducing risks, and enhancing prediction accuracy and security.

[0085] In summary, this application enhances robustness through dynamic transfer weight adjustment, incremental learning optimization, and adversarial training, thereby improving the adaptability, generalization ability, and security of the credit risk assessment model and providing financial institutions with a more accurate and efficient risk control solution.

[0086] In the above embodiments, a credit risk assessment method based on dynamic transfer and incremental learning is provided. Correspondingly, this application also provides a credit risk assessment device based on dynamic transfer and incremental learning, which can be implemented through software, hardware, or a combination of both. For example, the credit risk assessment device based on dynamic transfer and incremental learning may include integrated or separate functional modules or units to perform the corresponding steps in the above methods. Please refer to... Figure 3 This illustration shows a schematic diagram of a credit risk assessment device based on dynamic transfer and incremental learning, provided by some embodiments of this application. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple; relevant details can be found in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0087] like Figure 3 As shown, the credit risk assessment device 10 based on dynamic transfer and incremental learning includes:

[0088] Module 101 is used to acquire the borrower's target domain credit data;

[0089] Evaluation module 102 is used to input the target domain credit data into the trained target domain credit risk assessment model to obtain the borrower's default probability output by the target domain credit risk assessment model;

[0090] The target domain credit risk assessment model is obtained by transfer learning and incremental learning of the source domain credit risk assessment model. During the transfer learning process, the knowledge transfer weights of the source domain and the target domain are dynamically adjusted according to the changes in the amount and distribution of data in the target domain.

[0091] In one possible implementation, the training process of the target domain credit risk assessment model includes:

[0092] Source domain modeling steps: Construct a deep neural network as the source domain model S on the source domain data; perform data preprocessing on the source domain data using Z-score standardization to obtain source domain data A; train the source domain model S on the source domain data A; combine the attention mechanism to identify key features; the optimization objective is to minimize the Binary Cross-Entropy loss; the output is the borrower's default probability.

[0093] Target domain transfer steps: After the source domain is trained, the source domain model S is used to extract target domain data to obtain the target domain model T1. During the transfer process, the difference in feature distribution between the source and target domains is reduced by using the maximum mean difference loss. At the same time, the transfer weights of the source and target domains are dynamically adjusted according to the amount and distribution of target domain data. The dynamic adjustment process of transfer weights is as follows: when the amount of target domain data is small, the weight of the source domain is increased; as the target domain data gradually accumulates, the weight of the source domain is gradually reduced to reduce the impact of negative transfer.

[0094] Incremental learning steps: Incrementally learn the target domain model T1 to obtain the target domain model T2, and use the target domain model T2 as the target domain credit risk assessment model; During each incremental learning, only the target domain data of the most recent first preset number of days is used, updated once a day, and the data of the earliest day is removed. The entire training set always maintains the data of the most recent first preset number of days; If a data point has not been updated for more than the second preset number of days, then that data point is removed.

[0095] In one possible implementation, the target domain migration step specifically includes:

[0096] Based on the changes in the amount and distribution of data in the target domain, the migration weights of the source and target domains are dynamically adjusted using an exponential decay function.

[0097] In one possible implementation, the training process of the target domain credit risk assessment model further includes:

[0098] Adversarial training steps: The target domain model T2 is trained using an adversarial training method based on fast gradient signification and projective gradient descent. During the training process, the generated adversarial samples are mixed with the original samples. The parameters of the target domain model T2 are used as the initial parameters. The loss is calculated on the mini-batch and backpropagation is performed to update the parameters to obtain the target domain model T3. The target domain model T3 is used as the target domain credit risk assessment model.

[0099] In one possible implementation, the training process of the target domain credit risk assessment model further includes:

[0100] Feedback optimization steps: Load the target domain model T3 and the source domain model S, copy the first n layers of parameters of the target domain model T3 to the corresponding layers of the source domain model S, freeze the transferred layers, and then set a learning rate less than the preset value on the source domain data A to fine-tune the source domain model S.

[0101] The credit risk assessment device based on dynamic transfer and incremental learning provided in this application embodiment and the credit risk assessment method based on dynamic transfer and incremental learning provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0102] This application also provides an electronic device corresponding to the credit risk assessment method based on dynamic transfer and incremental learning provided in the aforementioned embodiments. The electronic device may be a mobile phone, laptop computer, tablet computer, desktop computer, etc., to execute the aforementioned credit risk assessment method based on dynamic transfer and incremental learning.

[0103] The electronic device provided in this application embodiment and the credit risk assessment method based on dynamic transfer and incremental learning provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0104] This application also provides a computer-readable storage medium corresponding to the credit risk assessment method based on dynamic migration and incremental learning provided in the foregoing embodiments, wherein a computer program (i.e., a program product) is stored thereon. When the computer program is run by a processor, it executes the credit risk assessment method based on dynamic migration and incremental learning provided in any of the foregoing embodiments.

[0105] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0106] The computer-readable storage medium provided in the above embodiments of this application and the credit risk assessment method based on dynamic migration and incremental learning provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0107] It should be noted that 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 marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive 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 action, or using a combination of dedicated hardware and computer instructions.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A credit risk assessment method based on dynamic transfer and incremental learning, characterized in that, include: Obtain the borrower's target domain credit data; Input the target domain credit data into the trained target domain credit risk assessment model to obtain the borrower's default probability output by the target domain credit risk assessment model; The target domain credit risk assessment model is obtained by transfer learning and incremental learning of the source domain credit risk assessment model. During the transfer learning process, the knowledge transfer weights of the source domain and the target domain are dynamically adjusted according to the changes in the amount and distribution of data in the target domain.

2. The method according to claim 1, characterized in that, The training process of the target domain credit risk assessment model includes: Source domain modeling steps: Construct a deep neural network as the source domain model S on the source domain data; perform data preprocessing on the source domain data using Z-score standardization to obtain source domain data A; train the source domain model S on the source domain data A; combine the attention mechanism to identify key features; the optimization objective is to minimize the Binary Cross-Entropy loss; the output is the borrower's default probability. Target domain transfer steps: After the source domain is trained, the source domain model S is used to extract target domain data to obtain the target domain model T1. During the transfer process, the feature distribution difference between the source and target domains is reduced by using the maximum mean difference loss. At the same time, the transfer weights of the source and target domains are dynamically adjusted according to the amount and distribution changes of the target domain data. The dynamic adjustment process of the transfer weights is as follows: when the amount of target domain data is small, the weight of the source domain is increased; as the target domain data gradually accumulates, the weight of the source domain is gradually decreased. Incremental learning steps: Incrementally learn the target domain model T1 to obtain the target domain model T2, and use the target domain model T2 as the target domain credit risk assessment model; During each incremental learning, only the target domain data of the most recent first preset number of days is used, updated once a day, and the data of the earliest day is removed. The entire training set always maintains the data of the most recent first preset number of days; If a data point has not been updated for more than the second preset number of days, then that data point is removed.

3. The method according to claim 2, characterized in that, The target domain migration step specifically includes: Based on the changes in the amount and distribution of data in the target domain, the migration weights of the source and target domains are dynamically adjusted using an exponential decay function.

4. The method according to claim 2, characterized in that, The training process of the target domain credit risk assessment model also includes: Adversarial training steps: The target domain model T2 is trained using an adversarial training method based on fast gradient signification and projective gradient descent. During the training process, the generated adversarial samples are mixed with the original samples. The parameters of the target domain model T2 are used as the initial parameters. The loss is calculated on the mini-batch and backpropagation is performed to update the parameters to obtain the target domain model T3. The target domain model T3 is used as the target domain credit risk assessment model.

5. The method according to claim 3, characterized in that, The training process of the target domain credit risk assessment model also includes: Feedback optimization steps: Load the target domain model T3 and the source domain model S, copy the first n layers of parameters of the target domain model T3 to the corresponding layers of the source domain model S, freeze the transferred layers, and then set a learning rate less than the preset value on the source domain data A to fine-tune the source domain model S.

6. A credit risk assessment device based on dynamic transfer and incremental learning, characterized in that, include: The acquisition module is used to acquire the borrower's target domain credit data; The evaluation module is used to input the target domain credit data into the trained target domain credit risk assessment model to obtain the borrower's default probability output by the target domain credit risk assessment model. The target domain credit risk assessment model is obtained by transfer learning and incremental learning of the source domain credit risk assessment model. During the transfer learning process, the knowledge transfer weights of the source domain and the target domain are dynamically adjusted according to the changes in the amount and distribution of data in the target domain.

7. The apparatus according to claim 6, characterized in that, The training process of the target domain credit risk assessment model includes: Source domain modeling steps: Construct a deep neural network as the source domain model S on the source domain data; perform data preprocessing on the source domain data using Z-score standardization to obtain source domain data A; train the source domain model S on the source domain data A; combine the attention mechanism to identify key features; the optimization objective is to minimize the Binary Cross-Entropy loss; the output is the borrower's default probability. Target domain transfer steps: After the source domain is trained, the source domain model S is used to extract target domain data to obtain the target domain model T1. During the transfer process, the feature distribution difference between the source and target domains is reduced by using the maximum mean difference loss. At the same time, the transfer weights of the source and target domains are dynamically adjusted according to the amount and distribution changes of the target domain data. The dynamic adjustment process of the transfer weights is as follows: when the amount of target domain data is small, the weight of the source domain is increased; as the target domain data gradually accumulates, the weight of the source domain is gradually decreased. Incremental learning steps: Incrementally learn the target domain model T1 to obtain the target domain model T2, and use the target domain model T2 as the target domain credit risk assessment model; During each incremental learning, only the target domain data of the most recent first preset number of days is used, updated once a day, and the data of the earliest day is removed. The entire training set always maintains the data of the most recent first preset number of days; If a data point has not been updated for more than the second preset number of days, then that data point is removed.

8. The apparatus according to claim 7, characterized in that, The target domain migration step specifically includes: Based on the changes in the amount and distribution of data in the target domain, the migration weights of the source and target domains are dynamically adjusted using an exponential decay function.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, It stores computer-readable instructions that can be executed by a processor to implement the method as described in any one of claims 1 to 5.

Images