Loan risk assessment method, electronic equipment and storage medium

Abstract

The invention discloses a loan risk assessment method, electronic equipment and a storage medium, and the method comprises the steps: obtaining enterprise portrait data of multiple modes, and carrying out the feature extraction of the enterprise portrait data through a preset feature extraction model, and obtaining a training sample set; training a preset twin network model by using the training sample set to obtain a trained twin network model; performing risk scoring on a target customer based on enterprise portrait data of the target customer through the trained twin network to obtain a first score of the target customer; calculating a comprehensive risk assessment score of the target customer in combination with the first score and the second score, and determining a corresponding risk level according to a preset rule; the second score is an artificial score. The lending risk assessment accuracy of the twin network model can be improved, and the bad debt risk is reduced.

Description

Technical Field

[0001] This application relates to the technical field of financial data processing, and particularly relates to a lending risk assessment method, an electronic device, and a storage medium. Background Art

[0002] Credit approval is an important link that comprehensively evaluates the credit status, repayment ability, loan purpose, etc. of borrowers, and can help financial institutions control credit risks and reduce the possibility of bad debts and defaults. In the bill acceptance business, the acceptor party needs to provide credit guarantee for the drawer to the payee, pay the bill amount funds on schedule before the bill payment due date, and recover the amount owed by the drawer. The steps of evaluating the repayment ability of the drawer party and evaluating the overdue and bad debt risks, so as to reasonably calculate the proportion of guarantee margin and discount amount, are an important part of the bill acceptance credit business.

[0003] However, the existing audit methods have many manual participation factors, and the audit results are often questioned. In addition, the fixed parsing and calculation model is difficult to adjust in real time to adapt to market changes, cannot be used in different customer credit scenarios, and requires users to clearly know the functions and usage methods of the model parameters, which has relatively high requirements for users. At the same time, the traditional method is limited by the model and often ignores some actually existing features, so it cannot capture complex non-linear relationships and implicit patterns. If the traditional deep learning classification method is to identify non-linear relationships and implicit patterns, a large amount of training data is required to train the deep learning model. In the credit scenario, the credit data is sensitive and private, and the data of each financial company is often not interoperable. At the same time, the credit lending business has a long development cycle, the credit behavior occurs with a low frequency, and the overall quantity is small, which restricts the solution of directly applying the existing deep learning model for lending assessment. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a lending risk assessment method, an electronic device, and a storage medium, which can improve the accuracy of the twin network model for lending risk assessment and reduce the bad debt risk.

[0005] To achieve the above purpose, the embodiments of this application adopt the following technical solutions: In a first aspect, the embodiments of this application provide a lending risk assessment method, including: Obtain enterprise portrait data of multiple modalities, and use a preset feature extraction model to extract features from the enterprise portrait data to obtain a training sample set; Use the training sample set to train a preset twin network model to obtain a trained twin network model; Based on the enterprise portrait data of the target customer, perform a risk score on the target customer through the trained twin network to obtain a first score of the target customer; Calculate the comprehensive risk assessment score of the target customer by combining the first score and the second score, and determine the corresponding risk level according to the preset rules; the second score is a manual score.

[0006] In one embodiment, after obtaining enterprise portrait data of multiple modalities, the method includes: Divide the enterprise portrait data into quantitative measurement data, word text measurement data, and media measurement data; Combine the quantitative measurement data, word text measurement data, and media measurement data into a total measurement data set, and preprocess the total measurement data set to obtain the available measurement data of the target customer.

[0007] In one embodiment, preprocessing the total measurement data set includes: Eliminate the data with missing required items and abnormal data in the total measurement data set; and / or Verify the consistency of the credit records of the target customer in different periods; and / or Verify the authenticity of the data in the total measurement data set.

[0008] In one embodiment, the enterprise portrait data of multiple modalities includes text data and picture data; using a preset feature extraction model to extract features from the enterprise portrait data includes: Use the BERT model to extract quantitative feature data from the text data, and use the VIT model to extract quantitative feature data from the picture data.

[0009] In one embodiment, using a training sample set to train a preset siamese network model includes: Normalize the quantitative feature data, and screen out redundant information in the quantitative feature data by the chi-square verification method to obtain a training sample set; Perform feature splicing on the data features in the training sample set to obtain a combined feature vector; Input the combined feature vector into the preset siamese network model to train the preset siamese network model.

[0010] In one embodiment, using a training sample set to train a preset siamese network model further includes: Input the first combined feature vector and the second combined feature vector into the preset siamese network model at the same time; Calculate the loss function of the first combined feature vector and the second combined feature vector, and optimize the loss function to improve the recognition ability of the trained siamese network model for data features.

[0011] In one embodiment, calculating the loss function of the first combined feature vector and the second combined feature vector includes: Calculate the loss function using the following formula: ; Wherein, is the loss function, is the first combined feature vector, is the second combined feature vector, and respectively represent the weight ratios of the loss amounts generated by the input samples of the same category and different categories to the total loss amount, represents the distance representation of the output features, and Y is 0 when the input samples are of the same category and 1 when the input samples are of different categories.

[0012] In one embodiment, calculating the comprehensive risk assessment score of the target customer by combining the first score and the second score, including: Calculating the comprehensive risk assessment score of the target customer according to the following formula: ; Wherein, is the comprehensive risk assessment score, A is the weight of the first score in the total score, is the first score, is the second score.

[0013] In one embodiment, calculating the optimal credit amount of multiple customers according to the comprehensive risk assessment score, and the calculation method is as follows: The objective function is defined as:

[0014] ; The constraint conditions are defined as: ; Wherein, Y is the income index of multiple customers, R i is the fixed return rate generated by the credit limit of customer i, a i is the credit amount of customer i as a decision variable; is the comprehensive risk assessment score of customer i determined by any method in claims 1 to 8 ， A is the upper limit of the total amount of the guarantee for the credit account, B is the upper limit of the expected risk of the credit; Under the condition of satisfying the constraint conditions, obtaining the optimal credit amount corresponding to each customer when maximizing the value of the objective function a i .

[0015] In a second aspect, the present application provides an electronic device, including: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the steps of the loan risk assessment method according to any one of the first aspect.

[0016] The loan risk assessment method, electronic device and storage medium provided by the present application, the method includes: obtaining enterprise portrait data of multiple modalities, and using a preset feature extraction model to extract features from the enterprise portrait data to obtain a training sample set; using the training sample set to train a preset siamese network model to obtain a trained siamese network model; based on the enterprise portrait data of the target customer, performing a risk score on the target customer through the trained siamese network to obtain a first score of the target customer; calculating a comprehensive risk assessment score of the target customer by combining the first score and a second score, and determining a corresponding risk level according to a preset rule; the second score is an artificial score. Therefore, the loan risk assessment method, electronic device and storage medium provided by the present application increase the reusability of training samples by using enterprise portrait data of multiple modalities to train a preset siamese network model, thereby enhancing the recognition ability of the model, and further enabling the model to accurately score the target enterprise based on the enterprise portrait data of the target customer. In this way, the score of the target customer can be made closer to the actual situation, so as to accurately determine the risk level of the target customer, thereby improving the accuracy of the model in loan risk assessment and reducing the bad debt risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings: Figure 1 It is a schematic flow chart of a loan risk assessment method provided by an embodiment of the present application; Figure 2 It is a schematic diagram of the process of training a preset siamese network model in the present invention; Figure 3 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0019] The terms "first", "second", etc. in this specification and the claims are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein. In addition, "and / or" in this specification and the claims means at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the associated objects before and after.

[0020] Refer to Figure 1 , which is a schematic flowchart of a lending risk assessment method provided by an embodiment of the present application. The lending risk assessment method can be executed by an electronic device provided by an embodiment of the present application. The electronic device can be implemented in a software and / or hardware manner, such as a server or a computer.

[0021] In this embodiment, taking the execution subject of the lending risk assessment method as the processor of a computer as an example, the lending risk assessment method provided by this embodiment includes: Step S1, obtain enterprise portrait data in multiple modalities, and use a preset feature extraction model to extract features from the enterprise portrait data to obtain a training sample set.

[0022] In this embodiment, the enterprise portrait data of the target customer includes the customer's revenue data, registered capital, business scope, queryable credit data, risk records, etc. A multi-modal data set with the target customer as a unit is constructed through these data.

[0023] Enterprise portrait data can include text modality, numerical modality, image modality, time series modality, etc. Specifically, the text modality can be enterprise descriptions, news reports, industry analysis reports, etc.; the numerical modality can be the enterprise's revenue data, registered capital, and number of employees; the image modality can be the enterprise's logo, product pictures, and office environment photos, etc.; the time series modality can be time series data such as historical stock price trends and annual financial reports.

[0024] Constructing a multi-modal data set with the target customer as a unit and combining data of multiple different modalities to construct a sample set can capture information about the customer in different dimensions, providing a comprehensive and detailed feature representation. Moreover, by combining multiple types of data, the model can learn richer and more complex patterns, thereby improving the generalization ability of the model on unknown data. In addition, multi-modal data provides more information sources, which can enable the model to not only rely on the features of a single type of data, helping to reduce the risk of overfitting of the model on the training data.

[0025] In one embodiment, after obtaining enterprise portrait data of multiple modalities, the method includes: dividing the enterprise portrait data into quantitative measurement data, word text measurement data, and media measurement data; forming a total set of measurement data from the quantitative measurement data, word text measurement data, and media measurement data, and preprocessing the total set of measurement data to obtain available measurement data of target customers.

[0026] Here, dividing the enterprise portrait data means classifying the enterprise portrait data according to its modality.

[0027] In one embodiment, preprocessing the total set of measurement data includes: removing data with missing required fields and abnormal data from the total set of measurement data; and / or performing consistency verification on the credit records of the target customer at different times; and / or verifying the authenticity of the data in the total set of measurement data.

[0028] It can be understood that during the data collection process, some fields are marked as required fields. When a piece of data has no value or is a null value in the required fields, this piece of data contains missing item data. For example, collect quantitative information such as revenue data, registered capital, business scope, and queryable credit data in the credit approval form to form a set Q of quantitative measurement dimension data, denoted as: , where m is the number of required fields for quantitative information; collect text information such as the company's operation analysis form and company risk assessment description uploaded with attachments to form a set T of word text measurement dimension data, denoted as: T , where n is the number of required fields for text fields; collect picture medium information such as business licenses, financial statements, and credit reports uploaded with attachments to form a set P of media measurement dimension data, denoted as: P , where l is the number of required fields for media media. Here, m, n, and l can be used to represent the number of each field respectively. In the data preprocessing stage, knowing the values of m, n, and l helps to understand the number of data items that need to be verified and cleaned, and ensure that each record meets the integrity requirements before being imported into the model.

[0029] Perform data cleaning and data sorting on the above-mentioned total set of measurement data D = Q ∪ T ∪ P data set. Remove the data set for data with missing required fields, and screen and remove abnormal samples and noise data. At the same time, to ensure the authenticity of the data as much as possible, perform consistency verification on the credit records of the same customer at different times, and manually participate in data verification and authenticity identification to ensure the integrity, availability, and authenticity of the data.

[0030] It can be understood that enterprise portrait data of multiple modalities includes text data and picture data; using a preset feature extraction model to extract features from the enterprise portrait data includes: using the BERT model to extract quantitative feature data from the text data, and using the VIT model to extract quantitative feature data from the picture data.

[0031] Specifically, by using different deep learning models (BERT model and VIT model), shallow and deep feature sets are obtained from text information and image information, forming feature extraction in different dimensions. The extracted feature sets are denoted as .

[0032] Step S2: Use the training sample set to train the preset siamese network model to obtain the trained siamese network model.

[0033] In one implementation, using the training sample set to train the preset siamese network model includes: normalizing the quantized feature data, and screening out redundant information in the quantized feature data through the chi-square verification method to obtain the training sample set; performing feature splicing on the data features in the training sample set to obtain a combined feature vector; inputting the combined feature vector into the preset siamese network model to train the preset siamese network model.

[0034] In this embodiment, min-max scale is used to normalize the quantized information. Before feature screening, the features within each modal dimension are uniformly normalized to prevent skewness problems caused by numerical issues.

[0035] Before formal training, feature screening is performed for the phenomenon of partial feature redundancy in the training sample set. Specifically, part of the redundant information is screened out through the chi-square verification method. The relevant formula for chi-square verification is:

[0036] where E is the mean of a certain feature dimension within a certain modality represents i the actual value of the feature dimension. Through the chi-square verification method, the independence between each feature is tested, part of the redundant features are screened out, the model complexity is reduced, and the subsequent model training time is shortened. After determining the dimension of the information to be screened out, the feature data to be screened out is removed, and the remaining features are spliced according to the preset rules to obtain a combined feature vector. The combined feature vector is input into the input layer of the siamese network to train the siamese network model.

[0037] It can be understood that the siamese network model is a special network architecture. It has two or more identical network branches that share weights and can process data in pairs.

[0038] Specifically, using the training sample set to train the preset siamese network model further includes: Input the first merged feature vector and the second merged feature vector into a preset siamese network model simultaneously; calculate the loss function of the first merged feature vector and the second merged feature vector, and optimize the loss function to improve the recognition ability of the trained siamese network model for data features.

[0039] Refer to Figure 2 , which is a schematic diagram of the process of training the siamese network model in the embodiment of the present application. As Figure 2 shown, and represent two input feature samples, which respectively pass through two convolutional neural networks with shared parameters. represents the output features of the two input samples of the metric distance loss function.

[0040] Here, the metric distance of features refers to the quantitative representation of the similarity or difference between two feature vectors in a specific feature space.

[0041] Specifically, calculating the loss function of the first merged feature vector and the second merged feature vector includes: calculating the loss function using the following formula: ; where is the loss function, is the first merged feature vector, is the second merged feature vector, and respectively represent the weight ratios of the loss amounts generated by the same-class input samples and different-class input samples to the total loss amount. represents the distance representation of the output features. Y is 0 when the input samples are of the same class and 1 when the input samples are of different classes.

[0042] Here, the first merged feature vector and the second merged feature vector refer to any two merged feature vectors.

[0043] It can be understood that the training objective is to make the distances between two similar inputs as small as possible and the distances between two different-class inputs as large as possible. During the training process, since the feature data is input in pairs, the diversity of the data set is increased, and the problem of fewer training samples is solved.

[0044] Step S3, based on the enterprise portrait data of the target customer, perform a risk score on the target customer through the trained siamese network to obtain the first score of the target customer.

[0045] As described above, after training the preset siamese network model using the training sample set, the trained siamese network model is obtained. The trained siamese network model has the ability to recognize feature vectors of various modalities. The enterprise portrait data of the target customer is preprocessed to obtain the corresponding feature vectors, and the trained siamese network model performs a risk score on the target customer according to the corresponding feature vectors, which can make the score result more in line with the actual situation, so as to accurately understand the risk status of the target customer.

[0046] Step S4, calculate the comprehensive risk assessment score of the target customer by combining the first score and the second score, and determine the corresponding risk level according to the preset rules; the second score is the manual score.

[0047] Specifically, calculating the comprehensive risk assessment score of the target customer by combining the first score and the second score includes: Calculate the comprehensive risk assessment score of the target customer according to the following formula: ; Wherein, is the comprehensive risk assessment score, A is the weight of the first score in the total score, is the first score, is the second score.

[0048] In an implementation manner, calculate the optimal credit amount of multiple customers according to the comprehensive risk assessment score, and the calculation method is as follows: The objective function is defined as:

[0049] ; The constraint conditions are defined as: ; Wherein, Y is the profit index of multiple customers, R i is the fixed return rate generated by the credit limit of the i-th customer, a i is the credit amount of the i-th customer as a decision variable; is the comprehensive risk assessment score of the i-th customer determined by any method in claims 1 to 8 ， A is the upper limit of the total amount of the guarantee of the credit account, B is the upper limit of the expected risk of the credit; under the condition of satisfying the constraint conditions, obtain the optimal credit amount corresponding to each customer when maximizing the value of the objective function a i .

[0050] Understandably, since the total credit limit amount is a fixed value, the credit limit amount accepted by each customer a i， and the risk assessment score are used as risk weights to calculate the income index Y , in order to ensure risk controllability, it is stipulated that the total margin of each customer is less than the upper limit of the total guarantee amount of the credit account, and the sum of the product of the proportion of each customer's margin in the total margin and the comprehensive risk score is less than the upper limit of the expected credit risk. Calculate the margin amount corresponding to the maximum income index of each customer under the condition of meeting the risk limit conditions, and then determine the credit limit for this customer. Therefore, through the multi-objective linear programming model, the overall credit lending income is controlled while ensuring risk controllability, and the yield is increased while enhancing the safety of bill acceptance.

[0051] The lending risk assessment method provided by the embodiments of the present application uses a scientific and reasonable risk assessment model to ensure the unbiasedness of the calculation of the risk assessment score during the credit granting process. It rates the credit of credit customers from multiple dimensions, and at the same time supports the dynamic adjustment of the weights of evaluation indicators, ensuring that financial risks are within a controllable range and adapting to changes in the market trend. It solves the problem that the deep learning model for credit risk assessment requires a large number of parameters and is difficult to implement in practice, and prevents the overfitting and poor generalization effects that occur with a small amount of data samples during the practical process. By integrating multi-modal technologies, it supports richer data sources and makes the evaluation results more accurate. Through the multi-objective linear programming model, the overall credit lending income is controlled while ensuring risk controllability, and the yield is increased while enhancing the safety of bill acceptance.

[0052] The specific embodiments of this specification have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multi-tasking and parallel processing are also possible or may be advantageous.

[0053] Figure 3 is a schematic structural diagram of an electronic device according to an embodiment of the present application. Please refer to Figure 3 , at the hardware level, this electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. Among them, the memory may include internal memory, such as high-speed random access memory (Random-Access Memory, RAM), and may also include non-volatile memory, such as at least one disk memory, etc. Of course, this electronic device may also include other hardware required for other services.

[0054] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 only a bidirectional arrow is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0055] Memory, used to store programs. Specifically, the program can include program code, and the program code includes computer operation instructions. The memory can include a memory and a non-volatile memory, and provides instructions and data to the processor.

[0056] The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it, forming a data processing device at the logical level. The processor executes the program stored in the memory and is specifically used to execute the loan risk assessment method provided in this embodiment.

[0057] Of course, in addition to the software implementation, the electronic device of the present application does not exclude other implementation manners, such as a logic device or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logical unit, and can also be a hardware or a logic device.

[0058] In summary, the above are only the preferred embodiments of the present application and are not used to limit the protection scope of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0059] The system, device, module, or unit illustrated in the above embodiments can be specifically implemented by a computer chip or an entity, or by a product with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0060] A computer-readable medium includes both permanent and non-permanent, removable and non-removable media and can implement information storage 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, a computer-readable medium does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0061] It should also be noted that the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device 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 device. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0062] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and reference can be made to the relevant parts of the method embodiments for the related content.

Claims

1. A loan risk assessment method, characterized in that: include: Acquire enterprise portrait data of multiple modalities, and use a preset feature extraction model to extract features from the enterprise portrait data to obtain a training sample set; Using the training sample set to train the preset twin network model to obtain a trained twin network model; Performing a risk score on the target customer based on the target customer's corporate profile data through the trained twin network to obtain a first score for the target customer; Calculate the comprehensive risk assessment score of the target customer by combining the first score and the second score, and determine the corresponding risk level according to preset rules; The second rating is a manual rating.

2. The method according to claim 1, characterized in that After obtaining the enterprise portrait data in multiple modes, the method includes: Dividing the enterprise portrait data into quantitative measurement data, word text measurement data and media measurement data; The quantitative metering data, the word text metering data and the media metering data are combined into a metering data set, and the metering data set is preprocessed to obtain usable metering data for the target customer.

3. The method according to claim 2, characterized in that The preprocessing of the total set of metering data includes: Eliminate missing data and abnormal data in the required items of the measurement data set; and / or Verify the consistency of the target customer's credit records at different times; and / or The authenticity of the data in the total set of metering data is verified.

4. The method according to claim 1, characterized in that The enterprise portrait data of multiple modalities includes text data and image data; and the feature extraction of the enterprise portrait data using a preset feature extraction model includes: The BERT model is used to extract the quantitative feature data in the text data, and the VIT model is used to extract the quantitative feature data in the image data.

5. The method according to claim 4, characterized in that The using the training sample set to train the preset twin network model includes: Normalizing the quantitative feature data, and filtering out redundant information in the quantitative feature data by a chi-square validation method to obtain the training sample set; Performing feature concatenation on the data features in the training sample set to obtain a merged feature vector; The merged feature vector is input into the preset twin network model to train the preset twin network model.

6. The method according to claim 5, characterized in that The method of using the training sample set to train the preset twin network model further includes: Inputting the first merged feature vector and the second merged feature vector into the preset twin network model simultaneously; Calculate the loss function of the first merged feature vector and the second merged feature vector, and optimize the loss function to improve the ability of the trained twin network model to recognize data features.

7. The method according to claim 6, characterized in that The calculating the loss function of the first merged feature vector and the second merged feature vector includes: The loss function is calculated using the following formula: ； in, is the loss function, is the first merged feature vector, is the second merged eigenvector, and They represent the weight ratio of the loss amount of the same category input samples and different category input samples to the total loss amount. Represents the distance representation of the output feature. Y is 0 when the input samples are of the same category and 1 when the input samples are of different categories.

8. The method according to claim 1, characterized in that Calculating the comprehensive risk assessment score of the target customer by combining the first score and the second score includes: The comprehensive risk assessment score of the target customer is calculated according to the following formula: ； in, is the comprehensive risk assessment score, A is the weight of the first score in the total score, For the first score, Score this second one.

9. The method according to claim 1, characterized in that: The method further includes: calculating the optimal credit amount for multiple customers according to the comprehensive risk assessment score, and the calculation method is as follows: The objective function is defined as: ； The constraints are defined as: ； in, Y is the revenue indicator of the plurality of customers, R i The fixed rate of return generated by the credit line of customer i, a i is the credit amount of customer i as the decision variable; is the comprehensive risk assessment score of customer i determined by any method in claims 1 to 8 ， A is the upper limit of the total guarantee amount of the credit account, B The expected risk limit for the credit line; Under the condition that the restriction condition is met, the optimal credit amount corresponding to each customer is obtained under the condition that the value of the objective function is maximized. a i .

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the steps of the risk assessment method as described in any one of claims 1 to 8.

Images