Loan default prediction method and device, equipment, storage medium and product

Abstract

The embodiment of the invention provides a loan default prediction method, device and equipment, a computer readable storage medium and a computer program product, relates to the technical field of artificial intelligence, and can be applied to the field of financial science and technology. The method comprises the following steps: training a local neural network model based on local loan data to obtain an initial loan default prediction model; processing the target loan data by using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability; uploading the local loan default prediction soft labels to a central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all bank terminals to obtain a global loan default prediction probability; and receiving a global loan default prediction probability, carrying out distillation training on the initial loan default prediction model, and minimizing the difference between the local loan default prediction probability and the global loan default prediction probability to obtain a target loan default prediction model.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a loan default prediction method, apparatus, device, computer-readable storage medium, and computer program product. Background Art

[0002] In recent years, with the rapid development of the financial industry, the financial services provided by financial institutions have also been rapidly developed and widely used. As an important financial service project, the security of loan business has received more and more attention and importance in the industry.

[0003] Therefore, how to accurately predict the customer's loan default risk has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0004] Embodiments of the present application provide a loan default prediction method, apparatus, device, computer-readable storage medium, and computer program product for improving the accuracy of loan default prediction.

[0005] In a first aspect, an embodiment of the present application provides a loan default prediction method, comprising:

[0006] Training a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information;

[0007] Processing the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability;

[0008] Uploading the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability;

[0009] The global loan default prediction probability is received, and the initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

[0010] In a second aspect, an embodiment of the present application provides a loan default prediction device, comprising:

[0011] An initial training module, used to train a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information;

[0012] A data generation module, used to process the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability;

[0013] A data uploading module, used to upload the local loan default prediction soft label to the central server, so that the central server can perform federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability;

[0014] A distillation learning module is used to receive the global loan default prediction probability and use the target loan data to perform distillation training on the initial loan default prediction model to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the loan default prediction method provided in the first aspect of the embodiment of the present application are implemented.

[0016] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the loan default prediction method provided in the first aspect of the embodiment of the present application are implemented.

[0017] In a fifth aspect, an embodiment of the present application provides a computer program product having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the loan default prediction method provided in the first aspect of the embodiment of the present application are implemented.

[0018] The technical solution provided in the embodiment of the present application is to train a local neural network model based on local loan data to obtain an initial loan default prediction model, use the initial loan default prediction model to perform reasoning on the target loan data to obtain a local loan default prediction soft label and a local loan default prediction probability, upload the local loan default prediction soft label to the central server, so that the central server can perform federated learning on the local loan default prediction soft labels uploaded by all bank terminals to obtain a global loan default prediction probability, receive the global loan default prediction probability, and use the target loan data to perform distillation training on the initial loan default prediction model to minimize the difference between the local loan default prediction probability and the global loan default prediction probability to obtain a target loan default prediction model; that is, the technical solution can enable each bank terminal to continuously adjust the local loan default prediction model based on the global loan default prediction knowledge issued by the central server through federated learning without sharing the loan data of each bank terminal, thereby realizing the collaborative training of the loan default prediction model by each bank terminal, improving the training effect of the loan default prediction model, thereby improving the accuracy of the loan default prediction, and also improving the security of the loan data of each bank terminal. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a flow chart of a loan default prediction method provided in an embodiment of the present application;

[0020] Figure 2 Another schematic diagram of a flow chart of a loan default prediction method provided in an embodiment of the present application;

[0021] Figure 3 A schematic diagram of another flow chart of the loan default prediction method provided in the embodiment of the present application;

[0022] Figure 4 A schematic diagram of the structure of a loan default prediction device provided in an embodiment of the present application;

[0023] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application are further described in detail through the following embodiments and in combination with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. Those skilled in the art can make adjustments to them as needed to suit specific application scenarios. It should also be noted that, for ease of description, only the parts related to the present application are shown in the accompanying drawings, rather than all structures.

[0025] In order to achieve effective loan management, machine learning is widely used to assist in loan default risk prediction. However, when training loan default prediction models, due to privacy and compliance reasons, the data between financial institutions is not open, that is, each bank institution uses the loan data within its own institution for training. Due to the relatively limited amount of loan data within a single bank institution, this will result in the loan default prediction model being unable to fully learn the complex feature relationships about loan defaults from multiple angles and dimensions, resulting in insufficient prediction capabilities of the trained loan default prediction model.

[0026] To this end, the technical solution provided in the embodiment of the present application, under the premise that each banking institution does not share its own privacy data, realizes collaborative training of the loan default prediction model with other banking institutions through federated learning, thereby improving the accuracy of loan default prediction.

[0027] It should be noted that the method provided in the embodiment of the present application can be executed by a loan default prediction device, which can be implemented by software and / or hardware and can be integrated into an electronic device. Optionally, the electronic device can be a bank's main server, or a bank branch server, or a bank's local terminal device. The main server data between banks will not be shared or connected. For a more secure restriction, the data of each branch of the same bank may not be shared, and the embodiment of the present application is not limited to this.

[0028] In addition, it should be noted that the information collected in the embodiments of the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data comply with the relevant laws, regulations and standards of relevant countries and regions, take necessary confidentiality measures, do not violate public order and good customs, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0029] Figure 1 A flow chart of a loan default prediction method provided in an embodiment of the present application. Figure 1 As shown, the method may include:

[0030] S101. Train a local neural network model based on local loan data to obtain an initial loan default prediction model.

[0031] Among them, local loan data includes user loan information, user feature information and user loan default information. User loan information may include loan term, loan amount, loan grade, loan type and loan interest rate, etc. User feature information may include user basic information and user personal asset information. User basic information includes customer number, age, gender, education level, marital status and industry information, etc. User personal asset information includes movable property information and immovable property information, which can be defined as feature information. User loan default information can be considered as hard label data, including two types of labels: default and non-default. The above-mentioned local neural network model can adopt the existing neural network model architecture, such as convolutional neural network, recurrent neural network / bidirectional recursive neural network, long short-term memory network / bidirectional long short-term memory network, attention model, etc. Each bank can select an adapted local neural network model architecture based on the software and hardware constraints and requirements of its own local device. For example, when the computing power of the local device is low, a lightweight model architecture can be selected. For example, when the content capacity of the local device is limited, a model architecture that occupies less memory can be selected to improve the model training efficiency.

[0032] Each bank can obtain local loan data within a historical period from the database, and perform data cleaning, feature encoding, and data alignment on the local loan data to meet the input requirements of the model. Furthermore, the local neural network model is trained using the pre-processed local loan data, and the training goal is to minimize the gap between the predicted probability of loan default and the actual loan default probability based on the local loan data. Specifically, each bank can use user loan information and user feature information as the input of the local neural network model, and user loan default information (i.e., the above-mentioned actual loan default probability) as the expected output of the local neural network model. The local neural network model is trained using a preset loss function, and the loss value of the loss function is minimized by the stochastic gradient descent algorithm in the iteration, thereby obtaining an initial loan default prediction model.

[0033] Optionally, the above loss function L i It can be expressed by the following formula:

[0034]

[0035] Among them, l ( x , y ) (w) represents the loss value generated by the initial loan default prediction model on a sample (x, y) in the local loan data, D i Represents all local loan data involved in training.

[0036] Optionally, when the tth round of local training is completed, the model parameters of the local neural network model can be updated based on the following formula:

[0037]

[0038] in, is the model parameter before updating, α i represents the learning rate, is the gradient value, is the updated model parameter, t is the current iteration number of the local neural network model, and the iteration number increases by 1 after each update and iteration.

[0039] S102: Process the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability.

[0040] After obtaining the initial loan default prediction model, the initial loan default prediction model can be used to infer the target loan data to obtain the local loan default prediction soft label and local loan default prediction probability on the target loan data. Among them, the local loan default prediction soft label is the output of the last layer before the softmax layer in the initial loan default prediction model, and the local loan default prediction probability is the result of the local loan default prediction soft label after the softmax function is processed. Among them, the target loan data here can be considered as a data set with hidden privacy information, and does not involve relevant privacy information.

[0041] S103: Upload the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability.

[0042] Each bank can upload its local loan default prediction soft labels to the central server, which will conduct federated learning on the local loan default prediction soft labels uploaded by all banks to generate and distribute the global loan default prediction probability.

[0043] Optionally, the above-mentioned global loan default prediction probability can be obtained through the following process: the central server aggregates the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction soft label; the central server processes the global loan default prediction soft label through a softmax function to obtain a global loan default prediction probability.

[0044] Specifically, the central server can aggregate the local loan default prediction soft labels uploaded by all banks through the following formula:

[0045]

[0046] in, represents the global loan default prediction soft label, N represents the number of banks participating in federated learning, and z i Represents the local loan default prediction soft label reported by the i-th bank.

[0047] Furthermore, after obtaining the global loan default prediction soft label, the central server processes the global loan default prediction soft label using the softmax function to obtain the global loan default prediction probability, and sends the global loan default prediction probability to each bank.

[0048] For the security of data transmission, each bank can encrypt the local loan default prediction soft label before uploading it to the central server. The central server can also encrypt the generated global loan default prediction probability before sending it to each bank.

[0049] S104, receiving a global loan default prediction probability, and performing distillation training on an initial loan default prediction model using target loan data to minimize a difference between a local loan default prediction probability and a global loan default prediction probability, thereby obtaining a target loan default prediction model.

[0050] Each bank receives the global loan default prediction probability sent by the central server, and each bank performs distillation training on the initial loan default prediction model again based on the target loan data. The training goal is to minimize the difference between the local loan default prediction probability and the global loan default prediction probability until the difference between the local loan default prediction probability and the global loan default prediction probability reaches the convergence condition, thereby obtaining the target loan default prediction model. Among them, the target loan default prediction model is used to predict the user's loan default information, that is, the loan information and user feature information of the user to be predicted are input into the target loan default prediction model, and the target loan default prediction model outputs the default probability of the user to be predicted.

[0051] The loan default prediction method provided in the embodiment of the present application trains a local neural network model based on local loan data to obtain an initial loan default prediction model, uses the initial loan default prediction model to perform reasoning on target loan data to obtain local loan default prediction soft labels and local loan default prediction probabilities, uploads the local loan default prediction soft labels to a central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all bank terminals to obtain a global loan default prediction probability, receives the global loan default prediction probability, and uses the target loan data to perform distillation training on the initial loan default prediction model to minimize the difference between the local loan default prediction probability and the global loan default prediction probability to obtain a target loan default prediction model; that is, the technical solution can enable each bank terminal to continuously adjust the local loan default prediction model based on the global loan default prediction knowledge issued by the central server through federated learning without sharing the loan data of each bank terminal, thereby realizing the collaborative training of the loan default prediction model by each bank terminal, improving the training effect of the loan default prediction model, thereby improving the accuracy of the loan default prediction, and also improving the security of the loan data of each bank terminal.

[0052] In one embodiment, optionally, Figure 2 As shown, the above S102 may include:

[0053] S201. Perform gradient pruning on the initial loan default prediction model.

[0054] Optionally, the initial loan default prediction model is gradient clipped using the following formula:

[0055]

[0056] Among them, w i is the model parameter of the initial loan default prediction model, S is the maximum value of the Euclidean norm of the initial loan default prediction model, and w i ′ is the model parameter after gradient clipping.

[0057] By clipping the gradient of model parameters and limiting the maximum value of the gradient, we can prevent gradient explosion and ensure the stability of model parameter updates.

[0058] S202. Add Gaussian noise that satisfies differential privacy protection to the initial loan default prediction model after gradient clipping to obtain the protected initial loan default prediction model.

[0059] Specifically, the query function is used to calculate the sensitivity of differential privacy, and based on the sensitivity, the preset privacy budget and the relaxation, a mean of 0 and a standard deviation of σ is generated. 2The generated Gaussian noise matrix is ​​added to the model parameters of the initial loan default prediction model to achieve differential privacy protection of the model parameters.

[0060] This method can protect the user's privacy data and improve the security of user data.

[0061] S203. Process the target loan data using the protected initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability.

[0062] After obtaining the protected initial loan default prediction model, the protected initial loan default prediction model can be used to perform inference on the target loan data to obtain a local loan default prediction soft label and a local loan default prediction probability on the target loan data.

[0063] In this embodiment, by performing differential privacy protection on the model parameters of the initial loan default prediction model, the model can resist attacks on data privacy and improve data security. At the same time, during the training process, adding differential privacy protection can also reduce the uncertainty caused by overfitting and underfitting and improve the stability and robustness of the model.

[0064] In one embodiment, optionally, Figure 3 As shown, in the above S104, the initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability. The process of obtaining the target loan default prediction model may include:

[0065] S301. Determine an initial loss value of a loss function based on a local predicted probability of loan default and an actual probability of loan default in target loan data.

[0066] The loss function may be any binary classification loss function such as a cross entropy loss function, a logarithmic loss function, a logical loss function or an exponential loss function. The local loan default prediction probability inferred by the initial loan default prediction model and the corresponding actual loan default probability are substituted into the loss function to obtain the initial loss value of the loss function.

[0067] S302: Determine the information divergence between the local loan default prediction probability and the global loan default prediction probability.

[0068] Among them, information divergence is used to indicate the degree of difference between the local loan default prediction probability and the global loan default prediction probability.

[0069] S303: Determine a target loss value of the loss function based on the initial loss value and the information divergence.

[0070] In some examples, the initial loss value and the information divergence may be summed, and the summed result may be determined as the target loss value of the loss function.

[0071] In some examples, a distillation weighting factor may also be obtained, the information divergence may be weighted using the distillation weighting factor, and the target loss value of the loss function may be determined based on the weighted information divergence and the initial loss value. For example, the weighted information divergence and the initial loss value may be summed, and the sum may be determined as the target loss value of the loss function. The above-mentioned distillation weighting factor is used to balance the difference between local training and global training, and the specific value may be set according to actual needs.

[0072] Optionally, the loss function L of the above distillation learning is i ′ can be expressed by the following formula:

[0073]

[0074] in, represents the loss between the local loan default prediction probability and the actual loan default probability inferred by the initial loan default prediction model on the target loan data, represents the information divergence between the local loan default prediction probability obtained by the initial loan default prediction model on the target loan data and the global loan default prediction probability issued by the central server, β i Represents the distillation weighting factor.

[0075] S304. Perform distillation training on the initial loan default prediction model based on the target loss value until the target loss value reaches a convergence condition, thereby obtaining a target loan default prediction model.

[0076] Specifically, if the target loss value of the loss function does not meet the convergence conditions, the gradient descent algorithm is used to adjust the model parameters of the initial loan default prediction model, and the inference is re-performed on the target loan data based on the adjusted model parameters, and the target loss value of the loss function is recalculated in the above manner until the target loss value meets the convergence conditions, thereby obtaining the target loan default prediction model.

[0077] During the distillation training process, the model parameters of the initial loan default prediction model can be updated based on the following formula:

[0078]

[0079] Among them, α i ′ represents the learning rate of knowledge distillation on the bank side, L i ′ is the target loss value of the loss function of distillation learning.

[0080] In this embodiment, by introducing the information divergence between the local loan default prediction probability and the global loan default prediction probability into the loss function, during the training process of the target loan default prediction model, by continuously narrowing the difference between the local loan default prediction probability and the global loan default prediction probability, each bank can learn from each other's default prediction experience, thereby improving the accuracy of the target loan default prediction model.

[0081] Figure 4 A schematic diagram of the structure of a loan default prediction device provided in an embodiment of the present application. Figure 4 As shown, the device may include: an initial training module 401, a data generation module 402, a data upload module 403 and a distillation learning module 404.

[0082] Specifically, the initial training module 401 is used to train the local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information;

[0083] The data generation module 402 is used to process the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability;

[0084] The data uploading module 403 is used to upload the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability;

[0085] The distillation learning module 404 is used to receive the global loan default prediction probability, and use the target loan data to perform distillation training on the initial loan default prediction model to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

[0086] Based on the above embodiment, optionally, the data generation module 402 is also used to perform gradient clipping on the initial loan default prediction model; add Gaussian noise that satisfies differential privacy protection to the initial loan default prediction model after gradient clipping to obtain the protected initial loan default prediction model; use the protected initial loan default prediction model to process the target loan data to obtain a local loan default prediction soft label and a local loan default prediction probability.

[0087] Based on the above embodiment, optionally, the data generation module 402 is further configured to perform gradient pruning on the initial loan default prediction model using the following formula:

[0088]

[0089] Among them, w i is the model parameter of the initial loan default prediction model, S is the maximum value of the Euclidean norm of the initial loan default prediction model, and w i ′ is the model parameter after gradient clipping.

[0090] Based on the above embodiment, optionally, the global loan default prediction probability is obtained through the following process: the central server aggregates the local loan default prediction soft labels uploaded by all banks to obtain the global loan default prediction soft labels; the central server processes the global loan default prediction soft labels through a softmax function to obtain the global loan default prediction probability.

[0091] Based on the above embodiment, optionally, the distillation learning module 404 is also used to determine an initial loss value of a loss function based on the local loan default prediction probability and the actual loan default probability in the target loan data; determine the information divergence between the local loan default prediction probability and the global loan default prediction probability; determine a target loss value of the loss function based on the initial loss value and the information divergence; and perform distillation training on the initial loan default prediction model based on the target loss value until the target loss value reaches a convergence condition to obtain a target loan default prediction model.

[0092] Based on the above embodiment, optionally, the distillation learning module 404 is also used to obtain a distillation weighting factor; use the distillation weighting factor to weight the information divergence, and determine the target loss value of the loss function based on the weighted information divergence and the initial loss value.

[0093] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application, such as Figure 5 As shown, the device includes a processor 50, a memory 51, an input device 52 and an output device 53; the number of the processor 50 in the device can be one or more. Figure 5 A processor 50 is taken as an example; the processor 50, memory 51, input device 52 and output device 53 in the device can be connected by a bus or other means. Figure 5 The example of connecting through bus is taken in the following.

[0094] The memory 51, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the loan default prediction method in the embodiments of the present application (for example, the initial training module 401, the data generation module 402, the data upload module 403 and the distillation learning module 404 used in the loan default prediction device). The processor 50 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 51, that is, implements the loan default prediction method provided in any of the above embodiments.

[0095] The memory 51 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created in the loan default prediction process, etc. In addition, the memory 51 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 51 may further include a memory remotely arranged relative to the processor 50, and these remote memories may be connected to the device / terminal / server via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0096] The input device 52 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the electronic device. The output device 53 may include a display device such as a display screen.

[0097] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0098] Training a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information;

[0099] Processing the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability;

[0100] Uploading the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability;

[0101] The global loan default prediction probability is received, and the initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

[0102] In one embodiment, a computer program product is further provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0103] Training a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information;

[0104] Processing the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability;

[0105] Uploading the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability;

[0106] The global loan default prediction probability is received, and the initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

[0107] The loan default prediction device, electronic device, computer-readable storage medium, and computer program product provided in the above embodiments can execute the loan default prediction method provided in any embodiment of the present application, and have the corresponding functional modules and beneficial effects of executing the method. For technical details not described in detail in the above embodiments, please refer to the loan default prediction method provided in any embodiment of the present application.

[0108] Through the above description of the implementation method, the technicians in the relevant field can clearly understand that the present application can be implemented with the help of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of each embodiment of the present application.

[0109] It is worth noting that the various units and modules included in the above embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application.

[0110] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A loan default prediction method, characterized in that: include: Training a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information; Processing the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability; Uploading the local loan default prediction soft label to the central server, so that the central server performs federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability; The global loan default prediction probability is received, and the initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

2. The method according to claim 1, characterized in that The target loan data is processed using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability, including: Performing gradient pruning on the initial loan default prediction model; Adding Gaussian noise that satisfies differential privacy protection to the initial loan default prediction model after gradient clipping to obtain the protected initial loan default prediction model; The protected initial loan default prediction model is used to process the target loan data to obtain a local loan default prediction soft label and a local loan default prediction probability.

3. The method according to claim 2, characterized in that The initial loan default prediction model is gradient pruned using the following formula: Among them, w i is the model parameter of the initial loan default prediction model, S is the maximum value of the Euclidean norm of the initial loan default prediction model, and w i ′ is the model parameter after gradient clipping.

4. The method according to any one of claims 1 to 3, characterized in that The global loan default prediction probability is obtained through the following process: The central server aggregates the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction soft label; The central server processes the global loan default prediction soft label through a softmax function to obtain the global loan default prediction probability.

5. The method according to any one of claims 1 to 3, characterized in that The initial loan default prediction model is distilled and trained using the target loan data to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model, including: Determining an initial loss value of a loss function based on the local loan default prediction probability and the actual loan default probability in the target loan data; Determining an information divergence between the local loan default prediction probability and the global loan default prediction probability; Determining a target loss value of the loss function based on the initial loss value and the information divergence; The initial loan default prediction model is distilled and trained based on the target loss value until the target loss value reaches a convergence condition, thereby obtaining a target loan default prediction model.

6. The method according to claim 5, characterized in that Determining a target loss value of the loss function based on the initial loss value and the information divergence includes: Get the distillation weighting factor; The information divergence is weighted by using the distillation weighting factor, and a target loss value of the loss function is determined based on the weighted information divergence and the initial loss value.

7. A loan default prediction device, characterized in that: include: An initial training module, used to train a local neural network model based on local loan data to obtain an initial loan default prediction model; wherein the local loan data includes user loan information, user feature information, and user loan default information; A data generation module, used to process the target loan data using the initial loan default prediction model to obtain a local loan default prediction soft label and a local loan default prediction probability; A data uploading module, used to upload the local loan default prediction soft label to the central server, so that the central server can perform federated learning on the local loan default prediction soft labels uploaded by all banks to obtain a global loan default prediction probability; A distillation learning module is used to receive the global loan default prediction probability and use the target loan data to perform distillation training on the initial loan default prediction model to minimize the difference between the local loan default prediction probability and the global loan default prediction probability, thereby obtaining a target loan default prediction model; wherein the target loan default prediction model is used to predict the user's loan default information.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images