Credit risk management method, device and computer program product

Abstract

The invention provides a credit risk management method and device and a computer program product, and the method comprises the steps: calculating the data quality weights of a plurality of model training participants through employing a quality federated average algorithm, eliminating the model training participants with the data quality weights smaller than a set threshold value, obtaining at least one remaining model training participant, and obtaining the data quality weights of the remaining model training participants; the model training participant is a credit issuing mechanism, and the data quality weight is the weight of the credit-related information of the model training participant influencing the accuracy of the model; training a neural network model by adopting credit related information of at least one residual model training participant to obtain a credit risk analysis model; inputting credit related information of the credit applicant into the credit risk analysis model to obtain a risk credit score of the credit applicant; and when the risk credit score of the credit applicant is less than the predetermined score, a credit risk early warning is sent to remind the credit applicant to apply for a credit issuing mechanism of credit, so that the problem of inaccurate credit risk assessment is solved.

Description

Technical Field

[0001] The present invention relates to the field of credit risk management, and more particularly, to a credit risk management method, apparatus, computer-readable storage medium, and computer program product. Background Art

[0002] In existing risk credit score prediction models, the quality and quantity of data together constitute the cornerstone of model accuracy. However, traditional risk credit score prediction models often rely too much on the quantity of data while relatively neglecting the importance of data quality. This tendency may not only lead to the model being affected by malicious participants but also significantly reduce the accuracy and reliability of the evaluation results due to the existence of low-quality data sources.

[0003] Currently, credit risk assessment mainly includes three types: credit scoring models, machine learning-based models, and hybrid models of credit scoring and machine learning. Although the above three solutions have improved the accuracy and efficiency of credit risk assessment to a certain extent, there are still research deficiencies.

[0004] Credit scoring models usually rely on historical data and financial information, which may not be applicable to emerging industries or new borrowers, resulting in certain data limitations. Moreover, the setting of their scoring criteria and weights often depends on the experience and judgment of experts, which may be subjective and biased.

[0005] Machine learning-based models can usually obtain more accurate prediction results. However, the correct selection and adjustment of model parameters may require professional knowledge, and the process is complex and time-consuming. Summary of the Invention

[0006] The main objective of this application is to provide a credit risk management method, apparatus, computer-readable storage medium, and computer program product to at least solve the problem of inaccurate credit risk assessment in the prior art.

[0007] To achieve the above object, according to one aspect of the present application, a credit risk management method is provided, including: calculating the data quality weights of multiple model training participants using the Federated Averaging algorithm, and eliminating the model training participants with data quality weights less than a set threshold to obtain at least one remaining model training participant, where the model training participant is a credit granting institution, and the data quality weight is the weight of the credit-related information of the model training participant affecting the model accuracy; training a neural network model using the credit-related information of at least one of the remaining model training participants to obtain a credit risk analysis model; inputting the credit-related information of a credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant; and issuing a credit risk warning when the risk credit score of the credit applicant is less than a predetermined score to remind the credit granting institution that the credit applicant applies for credit.

[0008] Optionally, the neural network model includes an encoder-decoder structure for extracting features of the credit-related information. Calculating the data quality weights of multiple model training participants using the Federated Averaging algorithm includes: training the neural network model using the historical credit-related information of each model training participant; when one training of the neural network model of each model training participant is completed, sending the updated model parameters of each model training participant to a central parameter server, and performing weighted averaging on the updated model parameters of each model training participant to obtain the model update parameters of the central parameter server, and updating the model parameters of the neural network model of the central parameter server using the model update parameters of the central parameter server until the number of updates of the neural network model of the central parameter server is equal to a first predetermined number of iterations or the loss function of the neural network model of the central parameter server converges, to obtain the pre-trained sub-models of each model training participant and the pre-trained total model of the central parameter server; inputting the historical credit-related information of each model training participant into the encoder-decoder structure of the corresponding pre-trained sub-model and the encoder-decoder structure of the pre-trained total model to obtain multiple shared sub-features and final features; and calculating the similarity of each shared sub-feature and the final feature to obtain the data quality weight of each model training participant.

[0009] Optionally, performing weighted averaging on the updated model parameters of each of the model training participants to obtain the model update parameters of the central parameter server, including: determining the ratio of the data volume of each of the model training participants to the total data volume of all the model training participants as the weight of each of the updated model parameters, where the data volume is the quantity of the historical credit-related information; and performing weighted averaging on the updated models using the weights of the updated model parameters to obtain the model update parameters of the central parameter server.

[0010] Optionally, calculating the similarity between each of the shared sub-features and the final feature to obtain the data quality weight of each of the model training participants, including: calculating the Euclidean distance between each of the shared sub-features and the final feature to obtain a plurality of Euclidean distances; determining the similarity corresponding to each of the Euclidean distances according to the Euclidean distance-similarity mapping relationship; and determining the similarity corresponding to each of the Euclidean distances as the data quality weight of the corresponding model training participant.

[0011] Optionally, training the neural network model using the historical credit-related information of at least one of the model training participants, including: a partitioning step of randomly selecting a predetermined quantity of the historical credit-related information each time from the historical credit-related information of a target model training participant to obtain a plurality of data sets, where the target model training participant is any one of the remaining model training participants; a training step of sequentially training the neural network model of the target model training participant using the data sets to obtain the model parameters updated multiple times; and sequentially repeating the partitioning step and the training step at least once until the number of updates of the neural network model of the target model training participant is equal to a second predetermined number of iterations or the loss function of the neural network model of the target model training participant converges, thereby completing one training.

[0012] Optionally, the neural network model includes an encoder-decoder structure and a softmax layer. Inputting the credit-related information of a credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant, including: in the case of inputting the credit-related information of the credit applicant into the credit risk analysis model, the encoder-decoder structure of the credit risk analysis model encodes the credit-related information of the credit applicant into a fixed-length vector to obtain an extracted feature, and decodes the extracted feature to generate a prediction sequence, and the softmax layer of the credit risk analysis model normalizes the prediction sequence to obtain the risk credit score of the credit applicant.

[0013] Optionally, the credit - related information of each of the remaining model training parties is used to train a neural network model to obtain a credit risk analysis model, including: using the historical credit - related information of each of the remaining model training parties to train the neural network model respectively; when the neural network models of each of the remaining model training parties complete one training, sending the updated model parameters of each of the remaining model training parties to the central parameter server, and performing weighted averaging on the updated model parameters of each of the remaining model training parties to obtain the model update parameters of the central parameter server, using the model update parameters of the central parameter server to update the model parameters of the neural network model of the central parameter server until the number of updates of the neural network model of the central parameter server is equal to the first predetermined number of iterations or the loss function of the neural network model of the central parameter server converges, and determining the neural network model of the central parameter server as the credit risk analysis model.

[0014] According to another aspect of the present application, a credit risk management device is provided, including: a calculation unit, configured to calculate the data quality weights of a plurality of model training parties by using the quality - federated averaging algorithm, and eliminate the model training parties with data quality weights less than a set threshold to obtain at least one remaining model training party, where the model training party is an institution for credit granting, and the data quality weight is the weight of the credit - related information of the model training party affecting the model accuracy; a training unit, configured to use the credit - related information of at least one remaining model training party to train a neural network model to obtain a credit risk analysis model; an analysis unit, configured to input the credit - related information of a credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant; and an early - warning unit, configured to issue a credit risk early - warning when the risk credit score of the credit applicant is less than a predetermined score to remind the credit - granting institution that the credit applicant applies for credit.

[0015] According to still another aspect of the present application, a computer - readable storage medium is provided, where the computer - readable storage medium includes a stored program, and when the program runs, it controls the device where the computer - readable storage medium is located to execute any one of the above - mentioned methods.

[0016] According to yet another aspect of the present application, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, it implements any one of the above - mentioned methods.

[0017] Applying the technical solution of the present application, in the above credit risk management method, the data quality weights of multiple model training participants are calculated through the quality federated averaging algorithm to eliminate the model training participants with poor data quality. The remaining model training participants jointly train to obtain a credit risk analysis model for credit risk early warning. This method greatly improves the data quality of model training, thereby improving the accuracy of credit risk early warning of the credit risk analysis model and solving the problem of inaccurate credit risk assessment in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 FIG. shows a hardware structure block diagram of a mobile terminal for executing a credit risk management method provided in an embodiment of the present application;

[0019] Figure 2 FIG. shows a schematic flowchart of a credit risk management method provided in an embodiment of the present application;

[0020] Figure 3 FIG. shows a schematic diagram of a horizontal federated learning framework provided in an embodiment of the present application;

[0021] Figure 4 FIG. shows a schematic diagram of an ED_LSTM model provided in an embodiment of the present application;

[0022] Figure 5 FIG. shows a schematic diagram of the module structure of a credit risk management system provided in an embodiment of the present application;

[0023] Figure 6 FIG. shows a block diagram of the structure of a credit risk management device provided in an embodiment of the present application.

[0024] Among them, the above-mentioned drawings include the following reference numerals:

[0025] 102. Processor; 104. Memory; 106. Transmission device; 108. Input / output device. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0027] To enable those skilled in the art to better understand the solution of this application, the following will clearly and completely describe the technical solution in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the scope of protection of this application.

[0028] It should be noted that the terms "first", "second", etc. in the description and claims of this application and the above-mentioned accompanying drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so as to describe the embodiments of this application here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0029] For the convenience of description, some nouns or terms involved in the embodiments of this application are described below:

[0030] Multi-task learning: A deep learning method that allows a model to learn multiple related tasks simultaneously. Its basic idea is that through sharing certain underlying parameters or feature representations, multiple tasks can assist each other, thereby improving the performance of each task.

[0031] Encoder-decoder structure: A neural network architecture commonly used to handle sequence-to-sequence problems, such as machine translation, speech recognition, etc. This structure consists of two parts: an encoder and a decoder. The encoder is responsible for encoding the input sequence into a fixed-length vector, and the decoder is responsible for decoding this vector into an output sequence.

[0032] Long Short-Term Memory Model (LSTM): A special recurrent neural network (RNN) architecture used to process sequence data. Compared with traditional RNNs, LSTM can better capture long-term dependencies in sequences, thereby avoiding the problems of gradient vanishing or gradient explosion. LSTM controls the flow of information by introducing "gate" mechanisms (such as input gates, forget gates, and output gates), enabling the model to remember important historical information and ignore irrelevant information.

[0033] Horizontal Federated Learning (HFL): A machine learning framework. Multiple participants (e.g., different organizations or devices) each have their own local datasets, train a part of the model on their local datasets, and share model parameters or updates through a secure communication protocol. These shared parameters are then securely aggregated by a central parameter server to update the global model. In this way, horizontal federated learning can improve the performance of the model using distributed data while protecting privacy.

[0034] As introduced in the background art, the credit risk assessment in the prior art is inaccurate. To solve this technical problem, embodiments of the present application provide a credit risk management method, apparatus, computer-readable storage medium, and computer program product.

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] The method embodiments provided in the embodiments of the present application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking the operation on a mobile terminal as an example, Figure 1 is a hardware structure block diagram of a mobile terminal of a credit risk management method according to an embodiment of the present invention. As Figure 1 shown, the mobile terminal may include one or more ( Figure 1 only one is shown in Figure 1 ) processors 102 (the processors 102 may include, but are not limited to, processing devices such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Among them, the above mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown in Figure 1 is only schematic and does not limit the structure of the above mobile terminal. For example, the mobile terminal may further include more or fewer components than those shown in

[0037] The memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the credit risk management method in the embodiments of the present invention. The processor 102 executes various functional applications and data processing by running the computer programs stored in the memory 104, that is, the above-mentioned method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories may be connected to the mobile terminal through 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. The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the mobile terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus can communicate with the Internet. In one instance, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0038] In this embodiment, a credit risk management method running on a mobile terminal, a computer terminal, or a similar computing device is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0039] Figure 2 It is a flowchart of the credit risk management method according to the embodiments of the present application. As Figure 2 shown, the method includes the following steps:

[0040] Step S201, calculating the data quality weights of multiple model training participants by using the Federated Averaging algorithm for Quality, and eliminating the above-mentioned model training participants whose data quality weights are less than a set threshold, to obtain at least one remaining model training participant, where the above-mentioned model training participants are credit granting institutions, and the above-mentioned data quality weights are the weights of the credit-related information of the above-mentioned model training participants affecting the model accuracy;

[0041] Step S202, training a neural network model by using the above-mentioned credit-related information of at least one of the above-mentioned remaining model training participants to obtain a credit risk analysis model;

[0042] Step S203: Input the credit-related information of the credit applicant into the above-mentioned credit risk analysis model to obtain the risk credit score of the credit applicant.

[0043] Step S204: When the risk credit score of the credit applicant is less than the predetermined score, issue a credit risk warning to remind the credit issuing institution that the credit applicant applies for credit.

[0044] In the above-mentioned credit risk management method, the data quality weights of multiple model training participants are calculated through the quality federated averaging algorithm to eliminate the model training participants with poor data quality. The remaining model training participants jointly train to obtain a credit risk analysis model for credit risk warning. This method greatly improves the data quality of model training, thereby improving the accuracy of credit risk warning of the credit risk analysis model, and solves the problem of inaccurate credit risk assessment in the prior art. This method can effectively improve the accuracy of credit risk assessment, reduce the credit losses of financial institutions, and is applicable to the credit approval processes of various credit institutions such as banks and microloan companies.

[0045] It should be noted that the horizontal federated learning framework is adopted to realize the joint utilization of data under the premise of protecting privacy. As Figure 3 shown, in the horizontal federated learning framework, each participant (such as the head office of a commercial bank, its subordinate first-level branch institution A, second-level branch institution B, etc., and other commercial banks X) shares the updated information of the model through secure encrypted communication, rather than directly sharing the original data. This mechanism ensures the security of data privacy while allowing data collaboration among different institutions to jointly improve the model performance.

[0046] The above-mentioned credit-related information is shown in Table 1. Merge the above various types of data according to the enterprise ID or other unique identifier to form a complete enterprise credit risk assessment data set; then, data standardization and normalization: perform standardization or normalization processing on numerical data to eliminate the dimension difference and improve the model stability. Finally, data partitioning: divide the processed data set into a training set, a validation set, and a test set for model training and evaluation.

[0047] Table 1

[0048]

[0049]

[0050] To improve the accuracy of the model's judgment on credit risk, in an optional implementation manner, the above-mentioned neural network model includes an encoding-decoding structure, and the encoding-decoding structure is used to extract the features of the above-mentioned credit-related information. The above step S201 includes:

[0051] Step S2011: Use the historical credit - related information of each of the above - mentioned model training participants to train the above - mentioned neural network model respectively;

[0052] Step S2012: When the above - mentioned neural network models of each of the above - mentioned model training participants complete one training, send the updated model parameters of each of the above - mentioned model training participants to the central parameter server, and perform weighted averaging on the above - mentioned updated model parameters of each of the above - mentioned model training participants to obtain the model update parameters of the above - mentioned central parameter server. Use the above - mentioned model update parameters of the above - mentioned central parameter server to update the model parameters of the above - mentioned neural network model of the above - mentioned central parameter server until the number of updates of the above - mentioned neural network model of the above - mentioned central parameter server is equal to the first predetermined number of iterations or the loss function of the above - mentioned neural network model of the above - mentioned central parameter server converges, so as to obtain the pre - trained sub - models of each of the above - mentioned model training participants and the pre - trained total model of the above - mentioned central parameter server;

[0053] Step S2013: Input the above - mentioned historical credit - related information of each of the above - mentioned model training participants into the above - mentioned encoding - decoding structure of the corresponding pre - trained sub - model and the above - mentioned encoding - decoding structure of the above - mentioned pre - trained total model to obtain multiple shared sub - features and final features;

[0054] Step S2014: Calculate the similarity of each of the above - mentioned shared sub - features and the above - mentioned final features to obtain the above - mentioned data quality weights of each of the above - mentioned model training participants.

[0055] In the above - mentioned implementation, since in credit risk assessment, the time - series data of borrowing customers, such as historical credit records and repayment behaviors, is crucial for accurately assessing risks. These data have long - term dependence characteristics, that is, the current decision may be affected by past events. Therefore, it is crucial to capture and process these long - term dependence relationships. The LSTM model is used to process the long - term dependence relationships in time - series data. Through its gating mechanism and internal memory units, it can accurately predict the future behaviors of borrowing customers. At the same time, the LSTM model has strong feature - learning ability and can extract key information from complex data. In the horizontal federated learning architecture, an ED_LSTM model with an encoding - decoding structure is designed, as Figure 4 shown. The encoder compresses the input time - series data into a fixed - length vector, retaining key information; the decoder generates a prediction sequence based on this vector. The ED_LSTM model consists of multiple neural network layers. Multiple LSTM layers can be stacked to improve the prediction accuracy, flexibly process time - series data of different lengths, and adapt to different business scenarios.

[0056] In addition, before the start of each iterative training, the central server calculates the similarity between the shared features of each participating party and the feature representation of the total model. The calculation result of this similarity is recorded in the quality weight matrix Q. The calculation method of the similarity in the present invention uses the Euclidean distance to measure. Taking the first-level branch institution A as an example, its local model is represented by the ED_LSTM-A model, and the generated shared sub-features are represented by X A denoted, the total model generated by the central parameter server is represented by ED_LSTM-final, and the generated final features are represented by X final denoted. According to the similarity between each of the above-mentioned shared sub-features X A and the above-mentioned final features X final , the data quality weight of the above-mentioned model training participants is measured. The same applies to other participants. The lower the similarity, the worse the data quality and the lower the weight.

[0057] In order to consider the influence of data volume on model accuracy, in an alternative embodiment, the above step S2012 includes:

[0058] Step S20121, determining the ratio of the data volume of each of the above model training participants to the total data volume of all the above model training participants as the weight of each of the above updated model parameters, where the data volume is the quantity of the above historical credit-related information;

[0059] Step S20122, performing weighted averaging on the above updated model using the weights of each of the above updated model parameters to obtain the model update parameters of the central parameter server.

[0060] In the above embodiment, generally speaking, the larger the data volume used for training, the better the performance of the trained model. Therefore, the ratio of the data volume of the model training participant to the total data volume is used to measure the contribution of the parameters of the model training participant to the model update parameters of the central parameter server, that is, as the influence of the weight of the above updated model parameters, so that the model update parameters of the central parameter server obtained by weighted averaging are more reasonable and more representative.

[0061] In order to accurately measure the similarity between the shared sub-features and the final features, in an alternative embodiment, the above step S2014 includes:

[0062] Step S20141, calculating the Euclidean distance between each of the above shared sub-features and the above final features to obtain a plurality of Euclidean distances;

[0063] Step S20142, determining the similarity corresponding to each of the above Euclidean distances according to the Euclidean distance-similarity mapping relationship;

[0064] Step S20143: Determine the similarities corresponding to the above Euclidean distances as the data quality weights of the above model training participants.

[0065] In the above embodiment, taking the first-level branch institution A as an example, calculate the shared sub-feature X A and the final feature X final The Euclidean distance is as follows: D(X A , X final ) = (X A - X final ). 2 For other participants, it is the same. The greater the distance, the worse the data quality, and the smaller the contribution to the final model, and even it may cause certain damage to the model. Map the distance to the corresponding percentage as the similarity, so as to determine the similarities corresponding to the above Euclidean distances as the data quality weights of the above model training participants, and then the contribution to the final model can be measured by the weights.

[0066] In order to implement the model training of the remaining model training participants, in an optional embodiment, the above step S202 includes:

[0067] Step S2021, the partitioning step: Randomly select a predetermined number of the above historical credit-related information from the above historical credit-related information of the target model training participant each time to obtain multiple data sets, where the target model training participant is any one of the above remaining model training participants;

[0068] Step S2022, the training step: Sequentially use the above data sets to train the above neural network model of the target model training participant to obtain the model parameters after multiple updates;

[0069] Step S2023: Sequentially repeat the above partitioning step and the above training step at least once until the update times of the above neural network model of the target model training participant are equal to the second predetermined iteration times or the loss function of the above neural network model of the target model training participant converges, and one training is completed.

[0070] In the above embodiment, obtain the latest model parameters from the central parameter server, w 1,1 (k) = w t . Randomly divide the data set into batches with a batch size of B for training. For the batch serial number b from 1 to the number of batches B, calculate the batch gradient g k (b), and update the local model parameters: w b+1,i (k) = w b,i (k) - μg k (b), where μ is the learning rate, and one training is completed.

[0071] In an alternative embodiment for accurately measuring the magnitude of risk, the above neural network model includes an encoder-decoder structure and a softmax layer, and the above step S203 includes:

[0072] Step S2031, when the credit-related information of the above credit applicant is input into the above credit risk analysis model, the encoder-decoder structure of the above credit risk analysis model encodes the above credit-related information of the above credit applicant into a fixed-length vector to obtain an extracted feature, and decodes the above extracted feature to generate a prediction sequence. The softmax layer of the above credit risk analysis model normalizes the above prediction sequence to obtain the above risk credit score of the above credit applicant.

[0073] In the above embodiment, as Figure 4 shown, the data participating in the training is first input into the encoder-decoder part. This part is responsible for mining low-redundancy shared feature representations to achieve effective information extraction and dimensionality reduction between data. Subsequently, these shared features are passed to the softmax layer for performing the risk credit score regression task, thereby obtaining a quantified risk assessment result. The entire training process adopts a multi-task learning strategy, that is, the two tasks of shared feature mining and risk score regression guide and promote each other to achieve the optimal training effect.

[0074] In an alternative embodiment for implementing federated learning, the above step S202 further includes:

[0075] Step S2024, training the above neural network model respectively with the historical credit-related information of each of the above remaining model training participants;

[0076] Step S2025, when the above neural network models of each of the above remaining model training participants complete one training, sending the updated model parameters of each of the above remaining model training participants to the central parameter server, and performing weighted averaging on the above updated model parameters of each of the above remaining model training participants to obtain the model update parameters of the above central parameter server. Using the above model update parameters of the above central parameter server to update the model parameters of the above neural network model of the above central parameter server until the number of updates of the above neural network model of the above central parameter server is equal to the first predetermined number of iterations or the loss function of the above neural network model of the above central parameter server converges, determining the above neural network model of the above central parameter server as the above credit risk analysis model.

[0077] In the above embodiments, when the number of times the model parameters of the central parameter server are updated ≥ M, the central parameter server eliminates malicious participants or participants with poor data quality according to the quality weight matrix Q and the threshold q. The central parameter server uses the weighted average method to perform secure aggregation on the model parameters of the remaining participants: where n k represents the size of the dataset of the k-th client participating in the federated training, and n represents the sum of the sample numbers of all client datasets. The central parameter server determines whether the loss function has converged. If it has converged, a signal is sent to each client to stop the model training. The central parameter server broadcasts the aggregated model parameters W t+1 to all clients.

[0078] In order to enable those skilled in the art to more clearly understand the technical solution of the present application, the implementation process of the credit risk management method of the present application will be described in detail below in combination with specific embodiments.

[0079] This embodiment relates to a specific credit risk management system, as Figure 5 shown, including:

[0080] The credit risk management system of this embodiment consists of a data preprocessing module, a horizontal federated learning module, a multi-task learning module, a malicious participant identification module, and a risk assessment module.

[0081] 1. Data preprocessing module:

[0082] Merge the above various types of data according to the enterprise ID or other unique identifier to form a complete enterprise credit risk assessment dataset; then, perform data standardization and normalization: perform standardization or normalization processing on numerical data to eliminate the dimensional difference and improve the model stability. Finally, perform data partitioning: partition the processed dataset into a training set, a validation set, and a test set for model training and evaluation.

[0083] 2. Horizontal federated learning module:

[0084] As Figure 3 shown, in this framework, each participant (such as the head office of a commercial bank, the first-level branch institution A under it, the second-level branch institution B, etc., and other commercial banks X) shares the update information of the model through secure encrypted communication, rather than directly sharing the original data. This mechanism ensures the security of data privacy and at the same time allows data collaboration among different institutions to jointly improve the model performance.

[0085] In credit risk assessment, the time-series data of borrowing customers, such as historical credit records and repayment behaviors, is crucial for accurately assessing risks. These data have the characteristic of long-term dependence, that is, the current decision may be affected by past events. Therefore, it is crucial to capture and process these long-term dependencies.

[0086] The LSTM model is used to handle the long-term dependencies in time-series data. Through its gating mechanism and internal memory units, it can achieve accurate prediction of the future behaviors of borrowing customers. At the same time, the LSTM model has strong feature learning ability and can extract key information from complex data.

[0087] In the horizontal federated learning architecture, the ED_LSTM model with an encoder-decoder structure is designed, as Figure 4 shown. The encoder compresses the input time-series data into a fixed-length vector, retaining the key information; the decoder generates the prediction sequence based on this vector. The ED_LSTM model consists of multiple neural network layers. Multiple LSTM layers can be stacked to improve the prediction accuracy, flexibly handle time-series data of different lengths, and adapt to different business scenarios.

[0088] 3. Multi-task learning module:

[0089] As Figure 4 shown, the data participating in the training is first input into the encoder-decoder part. This part is responsible for mining low-redundancy shared feature representations to achieve effective information extraction and dimensionality reduction between data. Subsequently, these shared features are passed to the softmax layer to perform the risk credit score regression task, thereby obtaining a quantitative risk assessment result. The entire training process adopts a multi-task learning strategy, that is, the two tasks of shared feature mining and risk score regression guide and promote each other to achieve the optimal training effect.

[0090] 4. Malicious participant identification module:

[0091] Before each iteration of training, the central server calculates the similarity between the shared features of each participant and the feature representation of the total model. The calculation result of this similarity is recorded in the quality weight matrix Q. The calculation method of the similarity in the present invention uses the Euclidean distance to measure. Taking the first-level branch institution A as an example, calculate the Euclidean distance between the shared sub-feature X A and the final feature X final as follows: D(X A , X final ) = (X A - X final ) 2 . The same applies to other participants. The greater the distance, the worse the data quality and the smaller the contribution to the final model, and even cause certain damage to the model. Finally, the quality weight matrix Q is calculated.

[0092] 5. Risk assessment module:

[0093] When the number of iterations ≥ k, according to the updated quality weight matrix Q, for the parties with similarity lower than the set threshold q, the method will automatically eliminate them, leaving the group of high-quality parties for subsequent training to ensure the overall effect of the federated model.

[0094] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0095] The embodiment of the present application also provides a credit risk management device. It should be noted that the credit risk management device of the embodiment of the present application can be used to execute the credit risk management method provided by the embodiment of the present application. This device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0096] The following introduces the credit risk management device provided by the embodiment of the present application.

[0097] Figure 6 It is a structural block diagram of the credit risk management device according to the embodiment of the present application. As Figure 6 shown, the device includes:

[0098] Calculation unit 10, which is used to calculate the data quality weights of multiple model training parties by using the quality federated average algorithm, and eliminate the above model training parties with data quality weights less than the set threshold to obtain at least one remaining model training party. The above model training parties are credit-granting institutions, and the above data quality weights are the weights of the credit-related information of the above model training parties affecting the model accuracy;

[0099] Training unit 20, which is used to train a neural network model by using the above credit-related information of at least one of the above remaining model training parties to obtain a credit risk analysis model;

[0100] Analysis unit 30, which is used to input the credit-related information of a credit applicant into the above credit risk analysis model to obtain the risk credit score of the above credit applicant;

[0101] Warning unit 40, which is used to issue a credit risk warning when the above risk credit score of the above credit applicant is less than a predetermined score to remind the above credit-granting institution of the above credit applicant applying for credit.

[0102] In the above credit risk management device, the data quality weights of multiple model training participants are calculated through the Federated Averaging algorithm to eliminate the model training participants with poor data quality. The remaining model training participants jointly train to obtain a credit risk analysis model for credit risk early warning. This method greatly improves the data quality of model training, thereby improving the accuracy of credit risk early warning of the credit risk analysis model and solving the problem of inaccurate credit risk assessment in the prior art. This method can effectively improve the accuracy of credit risk assessment, reduce the credit losses of financial institutions, and is applicable to the credit approval processes of various credit institutions such as banks and microloan companies.

[0103] It should be noted that the horizontal federated learning framework is adopted to realize the joint utilization of data under the premise of protecting privacy. As Figure 3 shown, in the horizontal federated learning framework, each participant (such as the head office of a commercial bank, subordinate first-level branch institution A, second-level branch institution B, etc., and other commercial banks X) shares the updated information of the model through secure encrypted communication, rather than directly sharing the original data. This mechanism ensures the security of data privacy and at the same time allows data collaboration between different institutions to jointly improve the model performance.

[0104] The above credit-related information is shown in Table 1. The above various types of data are merged according to the enterprise ID or other unique identifier to form a complete enterprise credit risk assessment data set; then, data standardization and normalization: the numerical data is standardized or normalized to eliminate the dimension difference and improve the model stability. Finally, data partitioning: the processed data set is partitioned into a training set, a validation set, and a test set for model training and evaluation.

[0105] To improve the accuracy of the model's judgment on credit risk, in an optional implementation manner, the above neural network model includes an encoder-decoder structure, and the encoder-decoder structure is used to extract the features of the above credit-related information. The above computing unit includes:

[0106] A first training module that trains the above neural network model respectively using the historical credit-related information of each of the above model training participants;

[0107] The first calculation module is used to, when the neural network models of the above-mentioned model training participants complete one training, send the updated model parameters of the above-mentioned model training participants to the central parameter server, and perform weighted averaging on the above-mentioned updated model parameters of the above-mentioned model training participants to obtain the model update parameters of the above-mentioned central parameter server. The model parameters of the neural network model of the above-mentioned central parameter server are updated using the above-mentioned model update parameters of the above-mentioned central parameter server until the number of updates of the neural network model of the above-mentioned central parameter server is equal to the first predetermined number of iterations or the loss function of the neural network model of the above-mentioned central parameter server converges, so as to obtain the pre-trained sub-models of the above-mentioned model training participants and the pre-trained total model of the above-mentioned central parameter server;

[0108] The processing module is used to input the above-mentioned historical credit-related information of the above-mentioned model training participants into the above-mentioned encoding-decoding structure of the corresponding above-mentioned pre-trained sub-model and the above-mentioned encoding-decoding structure of the above-mentioned pre-trained total model to obtain a plurality of shared sub-features and final features;

[0109] The second calculation module is used to calculate the similarity of the above-mentioned shared sub-features and the above-mentioned final features to obtain the above-mentioned data quality weights of the above-mentioned model training participants.

[0110] In the above implementation, since in credit risk assessment, the time-series data of borrowing customers, such as historical credit records and repayment behaviors, is crucial for accurately assessing risks. These data have the characteristic of long-term dependence, that is, the current decision may be affected by past events. Therefore, it is crucial to capture and process these long-term dependence relationships. The LSTM model is used to process the long-term dependence relationships in time-series data. Through its gating mechanism and internal memory unit, it can accurately predict the future behaviors of borrowing customers. At the same time, the LSTM model has strong feature learning ability and can extract key information from complex data. In the horizontal federated learning architecture, an ED_LSTM model with an encoding-decoding structure is designed, as Figure 4 shown. The encoder compresses the input time-series data into a fixed-length vector, retaining the key information; the decoder generates a prediction sequence based on this vector. The ED_LSTM model consists of multiple neural network layers, and multiple LSTM layers can be stacked to improve the prediction accuracy, flexibly process time-series data of different lengths, and adapt to different business scenarios.

[0111] In addition, before the start of each iterative training, the central server will calculate the similarity between the shared features of each participant and the feature representation of the total model. The calculation result of this similarity is recorded in the quality weight matrix Q. The calculation method of the similarity in the present invention uses the Euclidean distance to measure. Taking the first-level branch institution A as an example, its local model is represented by the ED_LSTM-A model, and the generated shared sub-feature is represented by X AIt is shown that the overall model generated by the central parameter server is denoted as ED_LSTM-final, and the generated final feature is denoted as X final It is shown that according to each of the above-mentioned shared sub-features X A and the above-mentioned final feature X final the similarity, the data quality weights of the above-mentioned model training participants are measured. The same applies to other participants. The lower the similarity, the worse the data quality and the lower the weight.

[0112] In order to consider the influence of data volume on model accuracy, in an alternative implementation, the above-mentioned first calculation module includes:

[0113] A first determination sub-module for determining the ratio of the data volume of each of the above-mentioned model training participants to the total data volume of all the above-mentioned model training participants as the weight of each of the above-mentioned updated model parameters, where the data volume is the quantity of the above-mentioned historical credit-related information;

[0114] A first calculation sub-module for performing a weighted average on the above-mentioned updated model using the weights of each of the above-mentioned updated model parameters to obtain the model update parameters of the above-mentioned central parameter server.

[0115] In the above implementation, generally speaking, the larger the data volume used for training, the better the performance of the trained model. Therefore, the ratio of the data volume of the model training participant to the total data volume is used to measure the contribution of the parameters of the model training participant to the model update parameters of the central parameter server, that is, as the influence of the weight of the above-mentioned updated model parameters, so that the model update parameters of the central parameter server obtained by weighted average are more reasonable and more representative.

[0116] In order to accurately measure the similarity between the shared sub-feature and the final feature, in an alternative embodiment, the above-mentioned second calculation module includes:

[0117] A second calculation sub-module for calculating the Euclidean distance between each of the above-mentioned shared sub-features and the above-mentioned final feature to obtain a plurality of Euclidean distances;

[0118] A second determination sub-module for determining the similarity corresponding to each of the above-mentioned Euclidean distances according to the Euclidean distance-similarity mapping relationship;

[0119] A third determination sub-module for determining the similarity corresponding to each of the above-mentioned Euclidean distances as the data quality weight of the corresponding each of the above-mentioned model training participants.

[0120] In the above implementation, taking the first-level branch institution A as an example, the Euclidean distance between the shared sub-feature X A and the final feature X final is calculated as follows: D(X A , X final ) = (X A-X final ) 2 , and the same applies to other participants. The greater the distance, the worse the data quality, and the smaller the contribution to the final model, or even causing certain damage to the model. Map the distance to the corresponding percentage as the similarity, so as to determine the data quality weights of the above-mentioned model training participants corresponding to the above-mentioned Euclidean distances, and then the contribution to the final model can be measured by the weights.

[0121] In order to implement the model training of the remaining model training participants, in an optional implementation manner, the above-mentioned training unit includes:

[0122] A partitioning module, configured to execute the partitioning step, randomly select a predetermined number of the above-mentioned historical credit-related information each time from the above-mentioned historical credit-related information of the target model training participant to obtain multiple data sets, and the target model training participant is any one of the above-mentioned remaining model training participants;

[0123] A second training module, configured to execute the training step, and sequentially use the above-mentioned data sets to train the above-mentioned neural network model of the target model training participant to obtain the model parameters updated multiple times;

[0124] A repeating module, configured to sequentially repeat the above-mentioned partitioning step and the above-mentioned training step at least once until the update times of the above-mentioned neural network model of the target model training participant are equal to the second predetermined iteration times or the loss function of the above-mentioned neural network model of the target model training participant converges, and one training is completed.

[0125] In the above implementation manner, obtain the latest model parameters from the central parameter server, w 1,1 (k) = w t . Randomly divide the data set into batches with a batch size of B for training. From the batch number b from 1 to the number of batches B, calculate the batch gradient g k (b), and update the local model parameters: w b+1,i (k) = w b,i (k) - μg k (b), where μ is the learning rate, and one training is completed.

[0126] In order to accurately measure the risk magnitude, in an optional embodiment, the above-mentioned neural network model includes an encoding-decoding structure and a softmax layer, and the above-mentioned analysis unit includes:

[0127] An analysis module, which, when inputting the credit-related information of the above-mentioned credit applicant into the above-mentioned credit risk analysis model, the above-mentioned encoding-decoding structure of the above-mentioned credit risk analysis model encodes the above-mentioned credit-related information of the above-mentioned credit applicant into a fixed-length vector to obtain extracted features, and decodes the above-mentioned extracted features to generate a prediction sequence. The above-mentioned softmax layer of the above-mentioned credit risk analysis model normalizes the above-mentioned prediction sequence to obtain the above-mentioned risk credit score of the above-mentioned credit applicant.

[0128] In the above embodiment, as Figure 4 shown, the data participating in the training is first input into the encoder-decoder part. This part is responsible for mining low-redundancy shared feature representations to achieve effective information extraction and dimensionality reduction between data. Subsequently, these shared features are passed to the softmax layer for performing the risk credit score regression task, thereby obtaining a quantified risk assessment result. The entire training process adopts a multi-task learning strategy, that is, the two tasks of shared feature mining and risk score regression guide and promote each other to achieve the optimal training effect.

[0129] To implement federated learning, in an optional embodiment, the above training unit further includes:

[0130] A third training module, which is used to train the above neural network model respectively with the historical credit-related information of each of the above remaining model training parties;

[0131] An update module, which, when the above neural network models of each of the above remaining model training parties complete one training, sends the updated model parameters of each of the above remaining model training parties to the central parameter server, and performs a weighted average on the updated model parameters of each of the above remaining model training parties to obtain the model update parameters of the above central parameter server. The model parameters of the above neural network model of the above central parameter server are updated using the above model update parameters of the above central parameter server until the number of updates of the above neural network model of the above central parameter server is equal to the first predetermined number of iterations or the loss function of the above neural network model of the above central parameter server converges, and the above neural network model of the above central parameter server is determined as the above credit risk analysis model.

[0132] In the above embodiment, when the number of updates of the model parameters of the central parameter server ≥ M, the central parameter server eliminates malicious parties or parties with poor data quality according to the quality weight matrix Q and the threshold q. The central parameter server performs secure aggregation on the model parameters of the remaining parties using the weighted average method: where, n kdenotes the dataset size of the k-th client participating in the federated training, and n denotes the sum of the sample numbers of all client datasets. The central parameter server determines the loss function whether it has converged. If it has converged, it sends signals to each client to stop the model training. The central parameter server broadcasts the aggregated model parameters W t+1 to all clients.

[0133] The above-mentioned credit risk management device includes a processor and a memory. The above-mentioned computing unit, training unit, analysis unit, warning unit, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to implement corresponding functions. The above-mentioned modules are all located in the same processor; or, the above-mentioned each module is located in different processors in any combined form.

[0134] The processor contains a kernel, and the kernel retrieves the corresponding program units from the memory. One or more kernels can be set, and by adjusting the kernel parameters, the problem of inaccurate credit risk assessment in the prior art can be solved.

[0135] 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, for example, read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one storage chip.

[0136] An embodiment of the present invention provides a computer-readable storage medium. The above-mentioned computer-readable storage medium includes a stored program, wherein when the above-mentioned program runs, it controls the device where the above-mentioned computer-readable storage medium is located to execute the above-mentioned credit risk management method.

[0137] Specifically, the credit risk management method includes:

[0138] Step S201, calculate the data quality weights of multiple model training participants by using the quality federated average algorithm, and eliminate the above-mentioned model training participants whose data quality weights are less than the set threshold to obtain at least one remaining model training participant. The above-mentioned model training participants are credit granting institutions, and the above-mentioned data quality weights are the weights of the credit-related information of the above-mentioned model training participants affecting the model accuracy;

[0139] Step S202, train a neural network model by using the above-mentioned credit-related information of at least one of the above-mentioned remaining model training participants to obtain a credit risk analysis model;

[0140] Step S203, input the credit-related information of a credit applicant into the above-mentioned credit risk analysis model to obtain the risk credit score of the above-mentioned credit applicant;

[0141] Step S204, when the risk credit score of the credit applicant is less than a predetermined score, issue a credit risk warning to remind the credit issuing institution to which the credit applicant applies for credit.

[0142] An embodiment of the present invention provides a processor for running a program, wherein when the program runs, it executes the above credit risk management method.

[0143] Specifically, the credit risk management method includes:

[0144] Step S201, calculate the data quality weights of multiple model training participants using the quality federated averaging algorithm, and eliminate the model training participants whose data quality weights are less than a set threshold to obtain at least one remaining model training participant. The model training participants are credit issuing institutions, and the data quality weights are the weights of the credit-related information of the model training participants affecting the model accuracy;

[0145] Step S202, train a neural network model using the credit-related information of at least one of the remaining model training participants to obtain a credit risk analysis model;

[0146] Step S203, input the credit-related information of the credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant;

[0147] Step S204, when the risk credit score of the credit applicant is less than a predetermined score, issue a credit risk warning to remind the credit issuing institution to which the credit applicant applies for credit.

[0148] An embodiment of the present invention provides a device, which includes a processor, a memory, and a program stored on the memory and executable on the processor. When the processor executes the program, it implements at least the following steps:

[0149] Step S201, calculate the data quality weights of multiple model training participants using the quality federated averaging algorithm, and eliminate the model training participants whose data quality weights are less than a set threshold to obtain at least one remaining model training participant. The model training participants are credit issuing institutions, and the data quality weights are the weights of the credit-related information of the model training participants affecting the model accuracy;

[0150] Step S202, train a neural network model using the credit-related information of at least one of the remaining model training participants to obtain a credit risk analysis model;

[0151] Step S203, input the credit-related information of the credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant;

[0152] Step S204, when the risk credit score of the credit applicant is less than a predetermined score, issue a credit risk warning to remind the credit issuing institution to which the credit applicant applies for credit.

[0153] This application also provides a computer program product, which when executed on a data processing device, is adapted to execute a program initialized with at least the following method steps:

[0154] Step S201, calculate the data quality weights of multiple model training participants using the Federated Averaging algorithm for Quality, and eliminate the model training participants whose data quality weights are less than a set threshold, to obtain at least one remaining model training participant. The model training participants are credit issuing institutions, and the data quality weights are the weights of the credit-related information of the model training participants affecting the accuracy of the model;

[0155] Step S202, train a neural network model using the credit-related information of at least one of the remaining model training participants to obtain a credit risk analysis model;

[0156] Step S203, input the credit-related information of the credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant;

[0157] Step S204, when the risk credit score of the credit applicant is less than a predetermined score, issue a credit risk warning to remind the credit issuing institution to which the credit applicant applies for credit.

[0158] Obviously, those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. They can be implemented by program codes executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately made into individual integrated circuit modules, or multiple modules or steps among them can be made into a single integrated circuit module to implement. In this way, the present invention is not limited to any specific combination of hardware and software.

[0159] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0160] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows 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 the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0161] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implement the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0162] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0163] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0164] The memory may include non-permanent memory in the computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

[0165] A computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The 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 technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media do not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0166] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.

[0167] From the above description, it can be seen that the above embodiments of the present application achieve the following technical effects:

[0168] 1) In the credit risk management method of the present application, the data quality weights of multiple model training participants are calculated by the quality federated averaging algorithm to eliminate model training participants with poor data quality, and the remaining model training participants jointly train to obtain a credit risk analysis model for credit risk early warning. This method greatly improves the data quality of model training, thereby improving the accuracy of credit risk early warning of the credit risk analysis model and solving the problem of inaccurate credit risk assessment in the prior art. This method can effectively improve the accuracy of credit risk assessment, reduce the credit losses of financial institutions, and is applicable to the credit approval processes of various credit institutions such as banks and microloan companies.

[0169] 2) In the credit risk management device of the present application, the data quality weights of multiple model training participants are calculated through the Federated Averaging algorithm for Quality (FedAvg-Q) to eliminate the model training participants with poor data quality. The remaining model training participants jointly train to obtain a credit risk analysis model for credit risk early warning. This method greatly improves the data quality of model training, thereby improving the accuracy of credit risk early warning of the credit risk analysis model, and solving the problem of inaccurate credit risk assessment in the prior art. This method can effectively improve the accuracy of credit risk assessment, reduce the credit losses of financial institutions, and is applicable to the credit approval processes of various credit institutions such as banks and microloan companies.

[0170] The above are only the preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A credit risk management method, characterized in that: include: The data quality weights of multiple model training participants are calculated using a quality federation average algorithm, and the model training participants whose data quality weights are less than a set threshold are eliminated to obtain at least one remaining model training participant, wherein the model training participant is a credit issuing institution, and the data quality weight is the weight of the credit-related information of the model training participant affecting the accuracy of the model; Using the credit-related information of at least one of the remaining models to train a neural network model to obtain a credit risk analysis model; Inputting the credit-related information of the credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant; In the case that the risk credit score of the credit applicant is less than a predetermined score, a credit risk warning is issued to remind the credit applicant to apply for credit from the credit issuing institution.

2. The method according to claim 1, characterized in that The neural network model includes an encoding-decoding structure, which is used to extract features of the credit-related information and calculate the data quality weights of multiple model training participants using a quality federation average algorithm, including: Use the historical credit-related information of each model training participant to train the neural network model respectively; When the neural network model of each model training participant completes one training, the updated model parameters of each model training participant are sent to the central parameter server, and the updated model parameters of each model training participant are weighted averaged to obtain the model update parameters of the central parameter server, and the model update parameters of the central parameter server are used to update the model parameters of the neural network model of the central parameter server until the number of updates of the neural network model of the central parameter server is equal to the first predetermined number of iterations or the loss function of the neural network model of the central parameter server converges, thereby obtaining the pre-trained sub-models of each model training participant and the pre-trained total model of the central parameter server; Inputting the historical credit-related information of each model training participant into the encoding-decoding structure of the corresponding pre-trained sub-model and the encoding-decoding structure of the pre-trained overall model to obtain a plurality of shared sub-features and a final feature; The similarity between each of the shared sub-features and the final feature is calculated to obtain the data quality weight of each of the model training participants.

3. The method according to claim 2, characterized in that The updated model parameters of each of the model training participants are weighted averaged to obtain the model update parameters of the central parameter server, including: Determine the weight of each updated model parameter as the ratio of the data volume of each model training participant to the total data volume of all model training participants, the data volume being the amount of the historical credit-related information; The updated model is weighted averaged using the weights of the updated model parameters to obtain the model update parameters of the central parameter server.

4. The method according to claim 2, characterized in that: Calculating the similarity between each of the shared sub-features and the final feature to obtain the data quality weight of each of the model training participants includes: Calculating the Euclidean distance between each of the shared sub-features and the final feature to obtain a plurality of Euclidean distances; Determine the similarity corresponding to each Euclidean distance according to the Euclidean distance-similarity mapping relationship; The similarity corresponding to each of the Euclidean distances is determined as the data quality weight of each corresponding model training participant.

5. The method according to claim 1, characterized in that Using at least one of the model training participants' historical credit-related information to train the neural network model includes: a dividing step of randomly selecting a predetermined number of the historical credit related information from the historical credit related information of the target model training participant each time to obtain multiple data sets, wherein the target model training participant is any one of the remaining model training participants; A training step, sequentially using the data set to train the neural network model of the target model training participant to obtain multiple updated model parameters; Repeat the division step and the training step at least once in sequence until the update number of the neural network model of the target model training participant is equal to the second predetermined number of iterations or the loss function of the neural network model of the target model training participant converges, thereby completing one training.

6. The method according to claim 1, characterized in that The neural network model includes an encoding-decoding structure and a softmax layer, and inputs the credit-related information of the credit applicant into the credit risk analysis model to obtain the risk credit score of the credit applicant, including: When the credit-related information of the credit applicant is input into the credit risk analysis model, the encoding-decoding structure of the credit risk analysis model encodes the credit-related information of the credit applicant into a fixed-length vector to obtain extracted features, and decodes the extracted features to generate a prediction sequence. The softmax layer of the credit risk analysis model normalizes the prediction sequence to obtain the risk credit score of the credit applicant.

7. The method according to any one of claims 1 to 6, characterized in that Using at least one of the remaining models to train the credit-related information of the participant to train a neural network model to obtain a credit risk analysis model, including: Use the historical credit-related information of the participants in each of the remaining model training participants to train the neural network model respectively; When the neural network model of each of the remaining model training participants completes one training, the updated model parameters of each of the remaining model training participants are sent to the central parameter server, and the updated model parameters of each of the remaining model training participants are weighted averaged to obtain the model update parameters of the central parameter server, and the model update parameters of the central parameter server are used to update the model parameters of the neural network model of the central parameter server until the number of updates of the neural network model of the central parameter server is equal to the first predetermined number of iterations or the loss function of the neural network model of the central parameter server converges, and the neural network model of the central parameter server is determined as the credit risk analysis model.

8. A credit risk management device, characterized in that: include: A calculation unit, configured to calculate data quality weights of multiple model training participants by using a quality federation average algorithm, and eliminate the model training participants whose data quality weights are less than a set threshold, to obtain at least one remaining model training participant, wherein the model training participant is a credit issuing institution, and the data quality weight is a weight of the credit-related information of the model training participant affecting the accuracy of the model; A training unit, configured to train a neural network model using the credit-related information of at least one of the remaining models to train the participant, to obtain a credit risk analysis model; An analysis unit, configured to input the credit-related information of the credit applicant into the credit risk analysis model to obtain a risk credit score of the credit applicant; The early warning unit is used to issue a credit risk early warning when the risk credit score of the credit applicant is less than a predetermined score, so as to remind the credit applicant to apply for credit from the credit issuing institution.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images