Credit risk identification method and device, storage medium and electronic equipment

Abstract

The invention provides a credit risk identification method and device, a storage medium and electronic equipment, and is applied to the technical field of credit risk assessment. The method comprises the following steps: obtaining a credit feature data set, wherein the credit feature data set comprises multiple customer credit feature data; performing data enhancement on the credit feature data set by using a hybrid variational self-encoding generative adversarial network to obtain a model training data set; training a pre-constructed credit risk classifier by using the model training data set to obtain a credit risk identification model, the credit risk classifier being constructed based on a convolutional neural network; and inputting the target customer credit feature data of the target customer into the credit risk identification model to obtain a credit risk prediction result which is output by the credit risk identification model and corresponds to the target customer. According to the method, data enhancement is performed on the existing credit data set through the hybrid self-encoding generative adversarial network, and the credit risk assessment model is trained by using the enhanced data set, so that the accuracy of credit risk identification is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of credit risk assessment, and in particular to a credit risk identification method, device, storage medium and electronic equipment. Background Art

[0002] In the current banking industry's credit business, risk analysis and assessment mainly rely on experienced manual operations. This approach not only leads to high labor costs, but also significantly reduces work efficiency. In addition, with the advancement of technology, deep learning technology has gradually been applied to the field of credit risk assessment.

[0003] However, in practical applications, credit data sets often have an imbalance between positive and negative samples. This data imbalance seriously affects the recognition ability of the model, resulting in low recognition accuracy, which in turn limits the effectiveness and widespread application of deep learning technology in credit risk assessment.

[0004] Therefore, how to improve the accuracy of credit risk identification has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of the above problems, the present invention provides a credit risk identification method, device, storage medium and electronic device that overcome the above problems or at least partially solve the above problems. The technical solution is as follows:

[0006] A credit risk identification method, comprising:

[0007] Obtaining a credit feature data set, wherein the credit feature data set includes a plurality of customer credit feature data;

[0008] Using a hybrid variational autoencoder generative adversarial network to perform data enhancement on the credit feature dataset to obtain a model training dataset;

[0009] Using the model training data set to train a pre-built credit risk classifier to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network;

[0010] The target customer's credit characteristic data is input into the credit risk identification model to obtain the credit risk prediction result corresponding to the target customer output by the credit risk identification model.

[0011] Optionally, obtaining a credit feature dataset includes:

[0012] Collect credit data of multiple customers;

[0013] Performing data cleaning on each of the customer credit data to obtain a customer credit data set including a plurality of the customer credit data after data cleaning;

[0014] The random forest method is used to screen the features of the credit data set and construct a credit feature data set.

[0015] Optionally, the step of using a hybrid variational autoencoder to generate an adversarial network to perform data enhancement on the credit feature dataset to obtain a model training dataset includes:

[0016] Using the credit feature data set to train a hybrid variational autoencoder generative adversarial network to obtain a minority class data sample generated by the trained hybrid variational autoencoder generative adversarial network;

[0017] The credit feature data set is expanded using the minority class data samples to obtain a model training data set.

[0018] Optionally, the step of using the credit feature dataset to train a hybrid variational autoencoder generative adversarial network to obtain a minority class data sample generated by the trained hybrid variational autoencoder generative adversarial network includes:

[0019] Inputting each of the customer credit feature data in the credit feature data set into a variational autoencoder of a hybrid variational autoencoder generative adversarial network to obtain reconstructed credit feature data output by the variational autoencoder;

[0020] The reconstructed credit feature data is input into the dual discriminator of the hybrid variational autoencoder generative adversarial network to perform data attribute judgment to determine whether the reconstructed credit feature data is good credit data. If so, the reconstructed credit feature data is saved as a minority class data sample. If not, the variational autoencoder and the dual discriminator are optimized based on the reconstructed credit feature data.

[0021] Optionally, the total loss function of the hybrid variational autoencoder generative adversarial network is the sum of the loss function of the variational autoencoder and the loss function of the dual discriminator.

[0022] Optionally, the variational autoencoder adopts a dual latent variable structure, and the variational process of the variational autoencoder is:

[0023] z=Enc(x)~q(z2|x,z1)q(z1|x)

[0024] Among them, z represents the latent variable; Enc(x) represents the encoder function, which is used to convert the input data x into the latent variable z; q(z2|x,z1) represents the probability distribution of the latent variable z2 given the input data x and another latent variable z1; q(z1|x) represents the probability distribution of another latent variable z1 given the input data x.

[0025] The generation process of the variational autoencoder is:

[0026]

[0027] in, represents the reconstructed input data, Dec(z) represents the decoder function, which is used to transform the latent variable z back to the data space to reconstruct the input data x; p(x|z1,z2) represents the probability distribution of the input data x given the latent variables z1 and z2.

[0028] Optionally, the loss function of the variational autoencoder is:

[0029]

[0030] Among them, L VAE represents the loss function of the variational autoencoder, represents the expectation under the joint distribution q(z2|x,z1)q(z1|x); p(x|z1,z2) represents the probability of generating input data x given the latent variables z1 and z2; p(z1,z2) represents the prior probability distribution of latent variables z1 and z2; q(z1,z2|x) represents the approximate posterior probability distribution of latent variables z1 and z2 given the input data x;

[0031] The loss function of the dual discriminator is:

[0032]

[0033] Among them, L GAN represents the loss function of the dual discriminator; Represents the distribution of real data P d In the case of sampling x, the expected value of the output of the discriminator D1 for the real data, D1(x) represents the probability estimate of the discriminator D1 that x is real; Denotes the noise distribution P input from the generator g In the case of sampling z, the expected output value of the discriminator D1 for the generated sample G(z); Denotes the noise distribution P input from the generator g In the case of sampling z, the discriminator D2 outputs the expected value of the probability that the generated sample G(z) is in a normal state.

[0034] A credit risk identification device comprises: a credit feature data set acquisition unit, a model training data set acquisition unit, a credit risk identification model acquisition unit and a credit risk prediction result acquisition unit.

[0035] The credit feature data set obtaining unit is used to obtain a credit feature data set, wherein the credit feature data set includes a plurality of customer credit feature data;

[0036] The model training data set acquisition unit is used to perform data enhancement on the credit feature data set using a hybrid variational autoencoder generative adversarial network to obtain a model training data set;

[0037] The credit risk identification model obtaining unit is used to train a pre-built credit risk classifier using the model training data set to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network;

[0038] The credit risk prediction result obtaining unit is used to input the target customer credit characteristic data of the target customer into the credit risk identification model to obtain the credit risk prediction result corresponding to the target customer output by the credit risk identification model.

[0039] A computer-readable storage medium stores a program, wherein the program, when executed by a processor, implements any of the above-mentioned credit risk identification methods.

[0040] An electronic device comprising at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; and the processor is used to call program instructions in the memory to execute any one of the above-mentioned credit risk identification methods.

[0041] By means of the above technical scheme, the present invention provides a credit risk identification method, device, storage medium and electronic device, obtains a credit feature data set, wherein the credit feature data set includes multiple customer credit feature data; uses a hybrid variational autoencoder to generate an adversarial network to enhance the credit feature data set to obtain a model training data set; uses the model training data set to train a pre-built credit risk classifier to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network; inputs the target customer's credit feature data into the credit risk identification model to obtain a credit risk prediction result corresponding to the target customer output by the credit risk identification model. The present invention enhances the existing credit data set through a hybrid autoencoder generative adversarial network, and uses the enhanced data set to train a credit risk assessment model, thereby improving the accuracy of credit risk identification.

[0042] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0044] Figure 1 A flow chart showing an implementation of a credit risk identification method provided by an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of the structure of a generative adversarial network provided by an embodiment of the present invention is shown;

[0046] Figure 3 A flow chart showing another implementation of the credit risk identification method provided by an embodiment of the present invention;

[0047] Figure 4 A flow chart showing another implementation of the credit risk identification method provided by an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of a network architecture of a hybrid variational autoencoder generative adversarial network provided by an embodiment of the present invention is shown;

[0049] Figure 6A schematic diagram of the generator structure of a hybrid variational autoencoder generative adversarial network provided by an embodiment of the present invention is shown;

[0050] Figure 7 A schematic diagram of a credit risk identification training process provided by an embodiment of the present invention is shown;

[0051] Figure 8 A schematic diagram showing the structure of a credit risk identification device provided by an embodiment of the present invention is shown;

[0052] Fig. 9 A schematic structural diagram of an electronic device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0054] As a highly flexible and convenient financing channel, credit is increasingly favored by a wide range of people. However, credit risk has long been an important factor restricting the development of credit platforms, and the serious adverse effects brought about by high default rates are particularly significant. Good loan default prediction can help financial institutions and credit platforms make wise loan decisions.

[0055] In recent years, machine learning methods have been widely used in the field of credit risk prediction, and many scholars at home and abroad have verified the feasibility of these methods. However, when studying credit risk assessment models, the number of overdue users is far less than that of normal repayment users, resulting in a serious imbalance in the sample ratio. Using an unbalanced data set for model training will affect the classification performance of the model, resulting in poor prediction results.

[0056] Based on this, an embodiment of the present invention provides a credit risk identification method, which expands the original data set by generating high-quality minority class samples, and then uses the expanded data set to train the credit risk identification model to improve its accuracy. In addition, when the variational autoencoder is used as the generator part of the generative adversarial network, the extracted coding features replace the traditional noise to achieve controllable sample generation. At the same time, a dual latent variable structure is designed to enhance the low-dimensional performance and decoupling; a dual attention network is adopted to capture the feature dependencies of the spatial and channel dimensions through the self-attention mechanism, and adaptively integrate local and global features; in response to the gradient instability and insufficient diversity of GAN, the Wasserstein distance is used in the discriminator to measure the sample distance, providing good computational properties and topological structure association. It can be seen that the embodiment of the present invention improves the recognition accuracy by performing data enhancement on the credit data set and using the enhanced data set to train the risk assessment model, so as to effectively evaluate customer information, predict credit risks, and help bank credit business improve efficiency, reduce time costs, and reduce default rates.

[0057] like Figure 1 As shown, a flow chart of an implementation of a credit risk identification method provided by an embodiment of the present invention is provided. The method may include:

[0058] S100: Obtain a credit feature data set, wherein the credit feature data set includes credit feature data of multiple customers.

[0059] Among them, the credit feature dataset is a data set specifically used to analyze and predict credit risk. This dataset contains multiple data records related to credit activities, and each record corresponds to the credit information of a customer.

[0060] Among them, customer credit feature data is the basic unit in the credit feature data set. For each customer, customer credit feature data may include multiple features used to assess the customer's credit risk. For example, customer credit feature data may include the customer's personal information, financial status, credit history, and loan information.

[0061] S110. Use a hybrid variational autoencoder to generate an adversarial network to enhance the credit feature dataset and obtain a model training dataset.

[0062] Among them, the Hybrid Variational Autoencoder Generative Adversarial Network is a model architecture that combines the Variational Autoencoder (VAE) and the Generative Adversarial Network (GAN).

[0063] Among them, the variational autoencoder is a generative model that aims to learn the potential distribution of data and generate new data. The variational autoencoder introduces the concept of latent variables, which enables the model to grasp the distribution characteristics of data. By sampling these latent variables, VAE can generate new data samples.

[0064] Among them, the generative adversarial network is a generative model based on deep learning, which consists of two parts: the generator and the discriminator. Figure 2 As shown in Figure 2, these two components are optimized through adversarial training. The goal of the generator is to randomly sample in the latent space to generate realistic synthetic samples, thereby learning the data distribution of real samples.

[0065] Data augmentation refers to increasing the size and diversity of a dataset by performing various transformations and generation on existing data, thereby improving the performance and generalization of the model. For credit feature datasets, data augmentation can be achieved by generating new, synthetic feature data that can mimic the diversity in the real world to some extent.

[0066] The model training dataset is a data set used to train machine learning models. The credit feature dataset after data enhancement can provide more training samples for the model, helping the model to better learn the relationship between different features and credit risk, thereby improving the prediction accuracy and robustness of the model. The model training dataset contains historical samples of known results, including borrower characteristics and their repayment behavior.

[0067] S120. Train a pre-built credit risk classifier using a model training data set to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network.

[0068] Among them, the credit risk classifier is an algorithm or model used to evaluate and classify the credit risk of borrowers. The credit risk classifier is built based on the Convolutional Neural Network (CNN). The credit risk classifier accepts the customer credit feature data in the input model training set data set, and outputs the borrower's possible default risk or credit rating through model analysis.

[0069] The credit risk identification model is a trained system that can perform risk assessment on new loan applications. This model uses a large amount of historical credit data to learn during the training phase to identify and predict the borrower's future default probability. By using this trained credit risk identification model, financial institutions can quickly identify high-risk borrowers and help with decision-making processes, such as whether to approve loans, set loan interest rates, etc.

[0070] S130: Input the target customer's credit feature data into a credit risk identification model to obtain a credit risk prediction result corresponding to the target customer output by the credit risk identification model.

[0071] Among them, the credit risk prediction result is a key information output by the credit risk identification model, which reflects the credit default risk of the target customer in a certain period of time in the future. This result is usually used to help financial institutions evaluate and decide whether to provide loans to the customer and under what conditions.

[0072] Optionally, the credit risk prediction results may include the probability of default, risk level, recommended measures, and credit score. The main function of the credit risk prediction results is to provide data support for financial decision-making, enabling lending institutions to manage risks more effectively, optimize credit approval processes, and formulate personalized loan terms. By analyzing these prediction results, institutions can maximize profits while reducing bad debt losses.

[0073] The present invention provides a credit risk identification method, which includes: obtaining a credit feature data set, wherein the credit feature data set includes multiple customer credit feature data; using a hybrid variational autoencoder generative adversarial network to perform data enhancement on the credit feature data set to obtain a model training data set; using the model training data set to train a pre-built credit risk classifier to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network; inputting the target customer's credit feature data of the target customer into the credit risk identification model to obtain a credit risk prediction result corresponding to the target customer output by the credit risk identification model. The present invention performs data enhancement on the existing credit data set through a hybrid autoencoder generative adversarial network, and uses the enhanced data set to train a credit risk assessment model, thereby improving the accuracy of credit risk identification.

[0074] Optional, based on Figure 1 The method shown, such as Figure 3 As shown, a flow chart of another implementation of the credit risk identification method provided by an embodiment of the present invention is shown, and step S100 may include:

[0075] S300, collecting credit data of multiple customers.

[0076] Among them, customer credit data refers to information related to the credit activities and credit status of individuals or enterprises. Customer credit data can include the customer's personal information, financial information, credit history and other relevant information.

[0077] S310: Clean the credit data of each customer to obtain a customer credit data set including a plurality of cleaned customer credit data.

[0078] Among them, data cleaning can include processing steps such as removing duplicate data, handling missing values, correcting erroneous data, and unifying formats, with the aim of improving data quality and ensuring the accuracy and reliability of analysis results.

[0079] Among them, the customer credit data set is a collection of multiple customer credit data that has been collected and organized.

[0080] S320. Use the random forest method to perform feature screening on the credit data set and construct a credit feature data set.

[0081] Feature screening refers to the process of selecting important features (variables) for customer credit risk from the original credit data set to improve the performance and efficiency of the model. Through feature screening, the model can focus on variables that have a significant impact on the prediction results and reduce computational complexity. The random forest algorithm can help identify and screen features with high predictive power by evaluating the importance of each feature in the decision tree.

[0082] The embodiment of the present invention collects a large amount of customer credit data to ensure data diversity, so that the model can reflect a wider range of customer behaviors and characteristics. By removing noise, correcting errors and processing missing values, data cleaning can greatly improve the quality of the data set, making model training more effective, and feature screening can significantly reduce the amount of data processed by the model, reduce computational complexity, speed up training and prediction, and help understand which factors most affect credit risk, thereby building a credit risk prediction model with excellent performance and providing more valuable support for financial decision-making.

[0083] Optional, based on Figure 1 The method shown, such as Figure 4 As shown, a flow chart of another implementation of the credit risk identification method provided by an embodiment of the present invention is shown, and step S110 may include:

[0084] S400. Use the credit feature data set to train the hybrid variational autoencoder generative adversarial network to obtain minority class data samples generated by the trained hybrid variational autoencoder generative adversarial network.

[0085] Among them, minority data samples refer to data of categories with fewer samples than other categories in the credit feature dataset.

[0086] S410. Use minority data samples to expand the credit feature data set to obtain a model training data set.

[0087] The embodiment of the present invention generates high-quality "fake" credit feature data by training a hybrid variational autoencoder generative adversarial network to expand the minority class in the credit feature data set, thereby achieving data enhancement. Using the enhanced credit feature data set to train a credit risk identification model can effectively alleviate the majority class bias problem caused by data imbalance. At the same time, increasing the sampling weight of the minority class improves the model's recognition accuracy for the minority class.

[0088] Optionally, step S400 may specifically include: inputting each customer credit feature data in the credit feature data set into the variational autoencoder of the hybrid variational autoencoder generative adversarial network to obtain reconstructed credit feature data output by the variational autoencoder. Input the reconstructed credit feature data into the dual discriminator of the hybrid variational autoencoder generative adversarial network to perform data attribute determination to determine whether the reconstructed credit feature data is good credit data. If so, save the reconstructed credit feature data as a minority class data sample. If not, optimize the variational autoencoder and the dual discriminator based on the reconstructed credit feature data.

[0089] The hybrid variational autoencoder generative adversarial network provided in the embodiment of the present invention is a hybrid model (Mix-VAEGAN) that combines variational autoencoder (VAE) and generative adversarial network (GAN). Figure 5 As shown, a schematic diagram of a network architecture of a hybrid variational autoencoder generative adversarial network provided by an embodiment of the present invention is shown, and the core part of the hybrid variational autoencoder generative adversarial network is composed of a generator (variational autoencoder) and a dual discriminator. The hybrid variational autoencoder generative adversarial network combines the functions of generation and discrimination, using a variational autoencoder as a generator to learn the potential representation of data, and at the same time using two discriminators to improve the quality and authenticity of the generated data.

[0090] The generator in the hybrid variational autoencoder generative adversarial network enhances its encoding ability by adding a variational autoencoder before the original generator. This not only solves the problem that the original generative adversarial network cannot effectively select noise when generating high-quality credit data, but also improves the shortcomings of the variational autoencoder generating unclear data through the adversarial training mechanism. In order to further improve the quality of generated credit data, the generator introduces a dual attention fusion network (Dual Attention Network, DANet) in the encoder. DANet can capture global feature dependencies in spatial and channel dimensions, thereby extracting more effective feature information and improving the quality of generated data.

[0091] The generator structure of the hybrid variational autoencoder generative adversarial network provided by the embodiment of the present invention is as follows: Figure 6As shown in the figure, the input data first passes through the first dense block (DB) of the encoder to generate the hidden variable z1, and then passes through the last dense block to generate the hidden variable z2. Then, the hidden variables z1 and z2 are spliced ​​together through the output of the attention mechanism module. Finally, the spliced ​​result is input to the decoder to generate the final data.

[0092] In the hybrid variational autoencoder generative adversarial network provided in an embodiment of the present invention, the discriminator part is composed of two independent discriminators D1 and D2 to distinguish the generated samples. Discriminator D1 is used to judge whether the input data is real good data, while discriminator D2 is used to judge whether the input data is bad data. Both discriminators are the same binary classification neural network, but they do not share parameters. Due to the large number of good samples, an effective classification model can be easily obtained through training.

[0093] However, in order to maintain data diversity and avoid the generated samples being too similar to the real good samples, the discriminator D2 is not optimized during the training process of the hybrid variational autoencoder generative adversarial network. During the training process, only the discriminator D1 and the generator G are iteratively optimized. The training goal of the generator G is to make the discriminator D1 unable to distinguish between the generated samples and the real bad samples, while making the discriminator D2 able to distinguish between the generated samples and the real good samples. Ideally, the probability calculated by the discriminator D1 is close to 0.5, while the probability calculated by the discriminator D2 is close to 0.

[0094] The embodiment of the present invention passes the reconstructed credit feature data through a variational autoencoder, which can retain the key features of the original data, improve the quality and reliability of the data, and effectively identify and extract good credit data by inputting the reconstructed credit feature data into a dual discriminator for discrimination. Saving these data judged to be good as minority class samples helps to balance the category distribution in the data set and improve the imbalance problem of samples. If the reconstructed data is not judged to be good data, the result is used to optimize the variational autoencoder and the dual discriminator. This feedback mechanism can continuously adjust the model parameters to improve the authenticity of the generated data and the accuracy of the discriminator.

[0095] The original GAN ​​generator faces the problems of unstable gradient and insufficient sample diversity. In order to solve these problems, the embodiment of the present invention uses Wasserstein distance in the discriminator of the hybrid variational autoencoder generative adversarial network to measure the distance between the real sample and the generated sample. Compared with other distance measurement methods, Wasserstein distance has good computational properties and is closely related to the topological structure, which helps to solve the problem of gradient disappearance to a certain extent.

[0096] Optionally, the variational autoencoder adopts a dual latent variable structure, that is, an auxiliary latent variable z1 is added on the basis of the original latent variable z2. The variational process and generation process of the variational autoencoder in training are as follows:

[0097] The variational process of the variational autoencoder is:

[0098] z=Enc(x)~q(z2|x,z1)q(z1|x)

[0099] Among them, z represents the latent variable; Enc(x) represents the encoder function, which is used to convert the input data x into the latent variable z; q(z2|x,z1) represents the probability distribution of the latent variable z2 given the input data x and another latent variable z1; q(z1|x) represents the probability distribution of another latent variable z1 given the input data x.

[0100] The generation process of the variational autoencoder is:

[0101]

[0102] in, represents the reconstructed input data, Dec(z) represents the decoder function, which is used to transform the latent variable z back to the data space to reconstruct the input data x; p(x|z1,z2) represents the probability distribution of the input data x given the latent variables z1 and z2.

[0103] The embodiment of the present invention designs a dual latent variable structure in the variational autoencoder and adds an auxiliary latent variable to enhance low-dimensional expression ability and decoupling. This design solves the problem that only a few latent variables of the variational autoencoder may be activated, resulting in poor performance.

[0104] Optionally, the loss function of the variational autoencoder consists of a reconstruction error and a regularization error. The reconstruction error measures the difference between the generated data and the original data, while the regularization error is used to constrain the distribution of the latent variables. Finally, the loss function of the variational autoencoder with auxiliary latent variables is:

[0105]

[0106] Among them, L VAE represents the loss function of the variational autoencoder, represents the expectation under the joint distribution q(z2|x,z1)q(z1|x); p(x|z1,z2) represents the generation probability of the input data x given the latent variables z1 and z2; p(z1,z2) represents the prior probability distribution of the latent variables z1 and z2; q(z1,z2|x) represents the approximate posterior probability distribution of the latent variables z1 and z2 given the input data x.

[0107] Generative Adversarial Networks (GANs) consist of two different networks: the generator G and the discriminator D. The generator G generates virtual samples that are as similar as possible to real samples by mapping random noise to the data space, and its output is represented as G(z). The discriminator D is responsible for classifying real data and generated data, and its output is represented as D(x).

[0108] In the fault diagnosis of high-reliability systems, the generated samples need to be not only similar to the real fault data, but also distinguishable from the real normal data. Therefore, the hybrid variational autoencoder generative adversarial network provided in the embodiment of the present invention adopts a dual discriminator structure to prevent the generated data from causing class overlap problems. In addition, in order to solve the gradient vanishing problem caused by discriminator saturation during training, the Wasserstein distance is used instead of the JS divergence to describe the distance between the generated sample distribution and the real sample distribution. The loss function of the dual discriminator is:

[0109]

[0110] Among them, L GAN Represents the loss function of the dual discriminator; Represents the distribution of real data P d In the case of sampling x, the expected value of the output of the discriminator D1 for the real data, D1(x) represents the probability estimate of the discriminator D1 that x is real; Denotes the noise distribution P input from the generator g In the case of sampling z, the expected output value of the discriminator D1 for the generated sample G(z); Denotes the noise distribution P input from the generator g In the case of sampling z, the discriminator D2 outputs the expected value of the probability that the generated sample G(z) is in a normal state.

[0111] The embodiment of the present invention introduces a dual discriminator structure. On the basis of the traditional single discriminator to evaluate the authenticity of input data, a second discriminator is added to distinguish whether the input data is good data or poor data, thereby improving the overall performance of the discriminator.

[0112] Optionally, the total loss function of the hybrid variational autoencoder generative adversarial network is the sum of the loss function of the variational autoencoder and the loss function of the dual discriminator, that is:

[0113] L=L GAN +L VAE

[0114] In order to facilitate understanding of the credit risk identification training process provided by the embodiment of the present invention, Figure 7Explanation: First, the credit feature data set is input into the hybrid variational autoencoder generative adversarial network. In the data enhancement stage, real samples are input to train the hybrid variational autoencoder generative adversarial network, and the loss is calculated and the parameters are optimized until the maximum number of training iterations is reached to generate good data. Then, in the risk assessment stage, the balanced credit feature data set expanded with the generated samples is input into the credit risk identification model for training until the maximum number of training iterations is reached. The test data is input into the trained credit risk identification model to obtain the classification results output by the credit risk identification model.

[0115] Although operations are depicted in a particular order, this should not be understood as requiring that the operations be performed in the particular order shown or in a sequential order.Multitasking and parallel processing may be advantageous under certain circumstances.

[0116] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0117] Corresponding to the above method embodiment, the present invention provides a credit risk identification device, the structure of which is as follows: Figure 8 As shown, it may include: a credit feature data set obtaining unit 10, a model training data set obtaining unit 20, a credit risk identification model obtaining unit 30 and a credit risk prediction result 40.

[0118] The credit feature data set obtaining unit 10 is used to obtain a credit feature data set, wherein the credit feature data set includes a plurality of customer credit feature data.

[0119] The model training data set acquisition unit 20 is used to perform data enhancement on the credit feature data set by using a hybrid variational autoencoder to generate an adversarial network to obtain a model training data set.

[0120] The credit risk identification model obtaining unit 30 is used to train a pre-built credit risk classifier using a model training data set to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network.

[0121] The credit risk prediction result 40 obtaining unit is used to input the target customer's credit feature data into the credit risk identification model to obtain the credit risk prediction result 40 corresponding to the target customer output by the credit risk identification model.

[0122] Optionally, the credit feature data set obtaining unit 10 may include: a customer credit data collecting subunit, a customer credit data set obtaining subunit and a data set constructing subunit.

[0123] The customer credit data collection subunit is used to collect multiple customer credit data.

[0124] The customer credit data set acquisition subunit is used to clean the data of each customer credit data to obtain a customer credit data set including a plurality of cleaned customer credit data.

[0125] The data set construction subunit is used to perform feature screening on the credit data set using the random forest method to construct a credit feature data set.

[0126] Optionally, the model training data set obtaining unit 20 may include: a minority class data sample obtaining subunit and a training data set obtaining subunit.

[0127] The minority class data sample acquisition subunit is used to train the hybrid variational autoencoder generative adversarial network using the credit feature data set to obtain the minority class data samples generated by the trained hybrid variational autoencoder generative adversarial network.

[0128] The training data set acquisition subunit is used to expand the credit feature data set using minority class data samples to obtain a model training data set.

[0129] Optionally, a minority data sample acquisition subunit can be specifically used to input the credit feature data of each customer in the credit feature data set into the variational autoencoder of the hybrid variational autoencoder generative adversarial network to obtain the reconstructed credit feature data output by the variational autoencoder; input the reconstructed credit feature data into the dual discriminator of the hybrid variational autoencoder generative adversarial network to perform data attribute judgment to determine whether the reconstructed credit feature data is good credit data. If so, the reconstructed credit feature data is saved as a minority data sample. If not, the variational autoencoder and the dual discriminator are optimized based on the reconstructed credit feature data.

[0130] Optionally, the total loss function of the hybrid variational autoencoder generative adversarial network is the sum of the loss function of the variational autoencoder and the loss function of the dual discriminator.

[0131] Optionally, the variational autoencoder adopts a dual latent variable structure, and the variational process of the variational autoencoder is:

[0132] z=Enc(x)~q(z2|x,z1)q(z1|x)

[0133] Among them, z represents the latent variable; Enc(x) represents the encoder function, which is used to convert the input data x into the latent variable z; q(z2|x,z1) represents the probability distribution of the latent variable z2 given the input data x and another latent variable z1; q(z1|x) represents the probability distribution of another latent variable z1 given the input data x.

[0134] The generation process of the variational autoencoder is:

[0135]

[0136] in, represents the reconstructed input data, Dec(z) represents the decoder function, which is used to transform the latent variable z back to the data space to reconstruct the input data x; p(x|z1,z2) represents the probability distribution of the input data x given the latent variables z1 and z2.

[0137] Optionally, the loss function of the variational autoencoder is:

[0138]

[0139] Among them, L VAE represents the loss function of the variational autoencoder, represents the expectation under the joint distribution q(z2|x,z1)q(z1|x); p(x|z1,z2) represents the generation probability of the input data x given the latent variables z1 and z2; p(z1,z2) represents the prior probability distribution of the latent variables z1 and z2; q(z1,z2|x) represents the approximate posterior probability distribution of the latent variables z1 and z2 given the input data x.

[0140] The loss function of the dual discriminator is:

[0141]

[0142] Among them, L GAN Represents the loss function of the dual discriminator; Represents the distribution of real data P d In the case of sampling x, the expected value of the output of the discriminator D1 for the real data, D1(x) represents the probability estimate of the discriminator D1 that x is real; Denotes the noise distribution P input from the generator g In the case of sampling z, the expected output value of the discriminator D1 for the generated sample G(z); Denotes the noise distribution P input from the generator g In the case of sampling z, the discriminator D2 outputs the expected value of the probability that the generated sample G(z) is in a normal state.

[0143] The present invention provides a credit risk identification device, which is used to: obtain a credit feature data set, wherein the credit feature data set includes multiple customer credit feature data; use a hybrid variational autoencoder generative adversarial network to perform data enhancement on the credit feature data set to obtain a model training data set; use the model training data set to train a pre-built credit risk classifier to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network; input the target customer's credit feature data of the target customer into the credit risk identification model to obtain a credit risk prediction result 40 corresponding to the target customer output by the credit risk identification model. The present invention performs data enhancement on the existing credit data set through a hybrid autoencoder generative adversarial network, and uses the enhanced data set to train a credit risk assessment model, thereby improving the accuracy of credit risk identification.

[0144] Regarding the device in the above embodiment, the specific manner in which each unit performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0145] The credit risk identification device includes a processor and a memory. The credit feature data set obtaining unit 10, the model training data set obtaining unit 20, the credit risk identification model obtaining unit 30 and the credit risk prediction result 40 obtaining unit are all stored in the memory as program units, and the processor executes the program units stored in the memory to realize corresponding functions.

[0146] The processor includes a kernel, which calls the corresponding program unit from the memory. One or more kernels can be set, and the kernel parameters are adjusted to enhance the existing credit data set through the hybrid autoencoder generative adversarial network, and the enhanced data set is used to train the credit risk assessment model, thereby improving the accuracy of credit risk identification.

[0147] An embodiment of the present invention provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the credit risk identification method is implemented.

[0148] An embodiment of the present invention provides a processor, which is used to run a program, wherein the credit risk identification method is executed when the program is run.

[0149] like Fig. 9As shown, an embodiment of the present invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003; wherein the processor 1001 and the memory 1002 communicate with each other through the bus 1003; the processor 1001 is used to call the program instructions in the memory 1002 to execute the above-mentioned credit risk identification method. The electronic device in this article can be a server, a PC, a PAD, a mobile phone, etc.

[0150] The present invention also provides a computer program product, which, when executed on an electronic device, is suitable for executing a program for initializing the steps of the credit risk identification method.

[0151] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable device to generate a machine, so that the instructions executed by the processor of the computer or other programmable device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0152] In a typical configuration, an electronic device includes one or more processors (CPU), a memory, and a bus. The electronic device may also include an input / output interface, a network interface, and the like.

[0153] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip. The memory is an example of a computer-readable medium.

[0154] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media 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 memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0155] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0156] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0157] In the description of the present invention, it should be understood that the terms "up", "down", "front", "back", "left" and "right" etc. indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the positions or elements referred to must have specific directions, be constructed and operate in specific directions. Therefore, they should not be understood as limitations of the present invention.

[0158] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0159] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0160] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A credit risk identification method, characterized in that: include: Obtaining a credit feature data set, wherein the credit feature data set includes a plurality of customer credit feature data; Using a hybrid variational autoencoder generative adversarial network to perform data enhancement on the credit feature dataset to obtain a model training dataset; Using the model training data set to train a pre-built credit risk classifier to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network; The target customer's credit characteristic data is input into the credit risk identification model to obtain the credit risk prediction result corresponding to the target customer output by the credit risk identification model.

2. The method according to claim 1, characterized in that The obtaining of the credit feature dataset includes: Collect credit data of multiple customers; Performing data cleaning on each of the customer credit data to obtain a customer credit data set including a plurality of the customer credit data after data cleaning; The random forest method is used to screen the features of the credit data set and construct a credit feature data set.

3. The method according to claim 1, characterized in that The method of using a hybrid variational autoencoder to generate an adversarial network to perform data enhancement on the credit feature dataset to obtain a model training dataset includes: Using the credit feature data set to train a hybrid variational autoencoder generative adversarial network to obtain a minority class data sample generated by the trained hybrid variational autoencoder generative adversarial network; The credit feature data set is expanded using the minority class data samples to obtain a model training data set.

4. The method according to claim 3, characterized in that The step of using the credit feature data set to train a hybrid variational autoencoder generative adversarial network to obtain a minority class data sample generated by the trained hybrid variational autoencoder generative adversarial network includes: Inputting each of the customer credit feature data in the credit feature data set into a variational autoencoder of a hybrid variational autoencoder generative adversarial network to obtain reconstructed credit feature data output by the variational autoencoder; The reconstructed credit feature data is input into the dual discriminator of the hybrid variational autoencoder generative adversarial network to perform data attribute judgment to determine whether the reconstructed credit feature data is good credit data. If so, the reconstructed credit feature data is saved as a minority class data sample. If not, the variational autoencoder and the dual discriminator are optimized based on the reconstructed credit feature data.

5. The method according to claim 4, characterized in that The total loss function of the hybrid variational autoencoder generative adversarial network is the sum of the loss function of the variational autoencoder and the loss function of the dual discriminator.

6. The method according to claim 5, characterized in that The variational autoencoder adopts a dual latent variable structure, and the variational process of the variational autoencoder is: z=Enc(x)~q(z2|x,z1)q(z1|x) Where z represents a latent variable; Enc(x) represents an encoder function used to convert input data x into a latent variable z; q(z2|x,z1) represents the probability distribution of the latent variable z2 given the input data x and another latent variable z1; q(z1|x) represents the probability distribution of another latent variable z1 given the input data x; The generation process of the variational autoencoder is: in, represents the reconstructed input data, Dec(z) represents the decoder function, which is used to transform the latent variable z back to the data space to reconstruct the input data x; p(x|z1,z2) represents the probability distribution of the input data x given the latent variables z1 and z2.

7. The method according to claim 6, characterized in that The loss function of the variational autoencoder is: Among them, L VAE represents the loss function of the variational autoencoder, represents the expectation under the joint distribution q(z2|x,z1)q(z1|x); p(x|z1,z2) represents the probability of generating input data x given the latent variables z1 and z2; p(z1,z2) represents the prior probability distribution of latent variables z1 and z2; q(z1,z2|x) represents the approximate posterior probability distribution of latent variables z1 and z2 given the input data x; The loss function of the dual discriminator is: Among them, L GAN represents the loss function of the dual discriminator; Represents the distribution of real data P d In the case of sampling x, the expected value of the output of the discriminator D1 for the real data, D1(x) represents the probability estimate of the discriminator D1 that x is real; Denotes the noise distribution P input from the generator g In the case of sampling z, the expected output value of the discriminator D1 for the generated sample G(z); Denotes the noise distribution P input from the generator g In the case of sampling z, the discriminator D2 outputs the expected value of the probability that the generated sample G(z) is in a normal state.

8. A credit risk identification device, characterized in that: include: A credit feature data set acquisition unit, a model training data set acquisition unit, a credit risk identification model acquisition unit, and a credit risk prediction result acquisition unit. The credit feature data set obtaining unit is used to obtain a credit feature data set, wherein the credit feature data set includes a plurality of customer credit feature data; The model training data set acquisition unit is used to perform data enhancement on the credit feature data set using a hybrid variational autoencoder generative adversarial network to obtain a model training data set; The credit risk identification model obtaining unit is used to train a pre-built credit risk classifier using the model training data set to obtain a credit risk identification model, wherein the credit risk classifier is built based on a convolutional neural network; The credit risk prediction result obtaining unit is used to input the target customer credit characteristic data of the target customer into the credit risk identification model to obtain the credit risk prediction result corresponding to the target customer output by the credit risk identification model.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the credit risk identification method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: The electronic device includes at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the credit risk identification method as described in any one of claims 1 to 7.

Images