Credit scoring model training method, credit scoring method and related device

Abstract

The invention discloses a training method of a credit scoring model, a credit scoring method and a related device. When the method provided by the embodiment of the invention is executed, a sample training set containing a plurality of credit score labels is obtained, each sample not only has the credit score label, but also comprises a plurality of sample data marked with risk modes, the risk mode labels indicate the monotonic relationship between the sample data and credit scores, and the credit score labels indicate the monotonic relationship between the sample data and the credit scores. Comprising monotone increasing, monotone decreasing or non-monotone relation, and the sample data covers credit investigation information of the sample user. On this basis, a back propagation algorithm is used to train a full-connection neural network (i.e., an initial scoring model) constructed based on a combined activation function, thereby obtaining an optimized credit scoring model. According to the method, the sample risk mode label is introduced into the training set, and the credit scoring model is constructed by using the combined activation function, so that the definite definition and capture of the monotonicity relation are realized, and the model is ensured to follow the monotonicity constraint during the training period.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a credit scoring model training method, a credit scoring method and related devices. Background Art

[0002] Existing credit risk assessment models, such as logistic regression, decision trees, and traditional neural networks, can predict corporate credit risk based on historical data to a certain extent. However, since the prediction process of these models is usually based on the statistical laws of data rather than clear economic theories, there is a lack of control over monotonicity constraints, and therefore prediction results that are inconsistent with economic logic will appear. For example, in some cases, although the company's financial situation has improved, the model may still predict that its credit risk will increase, reflecting that the model's capture of the relationship between variables does not fully conform to actual economic laws, and cannot guarantee the theoretical consistency and monotonicity of the prediction results.

[0003] Therefore, how to introduce monotonicity constraints into the credit risk assessment model to ensure that the prediction results are consistent with the actual economic logic is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0004] Based on the above problems, the present application provides a credit scoring model training method, a credit scoring method and related devices, which can improve the accuracy and efficiency of identifying customer purchasing power levels.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] A method for training a credit scoring model, the method comprising:

[0007] Acquire a training set; the training set includes a plurality of samples with credit score labels, and one of the samples includes a plurality of sample data with sample risk pattern labels; the risk pattern label indicates whether the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; the sample data includes the credit information of the sample user;

[0008] Based on the training set and the credit score label, the initial scoring model is trained using a back propagation algorithm to obtain a credit scoring model; the initial scoring model is a fully connected network neural network constructed based on a combined activation function.

[0009] In a possible implementation, the construction process of the combined activation function includes:

[0010] Select the original activation function ρ;

[0011] The original activation function ρ is subjected to point reflection based on the origin to obtain a first activation function which is monotonically opposite to the original activation function ρ. x represents the input data after data preprocessing;

[0012] Based on the first activation function And the original activation function ρ constructs the second activation function

[0013] The original activation function ρ, the first activation function And the second activation function Combine to obtain the combined activation function ρ s ; s = (s1, s2, s3); s is a triple, s1 represents the first activation function The number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, s2 represents the number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, and s3 represents the second activation function The layer structure is the number of neurons in the monotonicity constrained fully connected layer.

[0014] In a possible implementation, the fully connected network neural network includes an input layer, multiple hidden layers connected in series or in parallel, and an output layer;

[0015] The construction process of the monotonicity constrained fully connected network includes:

[0016] Redefine the output function of the output layer of the fully connected network neural network based on the combined activation function to obtain the monotonicity constrained fully connected network;

[0017] Wherein, the output function of the output layer of the monotonicity-constrained fully connected network is y=ρ s (|W T | t ·X+b), W represents the weight matrix of the fully connected network neural network; t represents the monotonicity indicator vector, when the monotonic relationship between the input data x and the credit score is a monotonic increasing relationship, the value of t is 1, when the monotonic relationship between the input data x and the credit score is a monotonic decreasing relationship, the value of t is -1, when the monotonic relationship between the input data x and the credit score is a non-monotonic relationship, the value of t is 0; b represents the bias term of the fully connected network neural network; |W T | t It represents the absolute value operation of the transpose of the weight matrix W according to the monotonicity indicator vector t.

[0018] In a possible implementation, the data preprocessing includes:

[0019] The input data is standardized or normalized.

[0020] In a possible implementation, the step of training the initial scoring model using a back propagation algorithm based on the training set and the credit scoring label to obtain a credit scoring model includes:

[0021] Inputting the training set into the initial scoring model to obtain an output score;

[0022] Calculating the difference between the output score and the credit score label to obtain a model error;

[0023] Based on the model error, the back propagation algorithm and the loss function are used to iteratively adjust the parameters of the binary classifier to obtain the credit scoring model.

[0024] A credit scoring method, the method comprising:

[0025] Acquire a scoring data set of a target customer; the scoring data set includes the credit information of the target customer; the credit information of the target customer is annotated with a data risk pattern label; the data risk pattern label indicates whether the monotonic relationship between the credit information of the target customer and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship;

[0026] The scoring data set of the target customer is input into a credit scoring model for scoring to obtain the credit score of the target customer; the credit scoring model is trained according to the training method of the credit scoring model as described above.

[0027] A training device for a credit scoring model, the device comprising:

[0028] A training set acquisition unit, used to acquire a training set; the training set includes a plurality of samples with credit score labels, one of the samples includes a plurality of sample data with sample risk pattern labels; the risk pattern label indicates whether the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; the sample data includes the credit information of the sample user;

[0029] A model training unit is used to perform model training on the initial scoring model based on the training set and the credit scoring label using a back propagation algorithm to obtain a credit scoring model; the initial scoring model is a fully connected network neural network constructed based on a combined activation function.

[0030] A credit scoring device, comprising:

[0031] A scoring data set acquisition unit is used to acquire a scoring data set of a target customer; the scoring data set includes the credit information of the target customer; the credit information of the target customer is annotated with a data risk pattern label; the data risk pattern label indicates whether the monotonic relationship between the credit information of the target customer and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship;

[0032] An input unit is used to input the scoring data set of the target customer into a credit scoring model for scoring, so as to obtain the credit score of the target customer; the credit scoring model is trained according to the training method of the credit scoring model as described above.

[0033] A training device for a credit scoring model comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the training method for the credit scoring model as described above is implemented.

[0034] A credit scoring device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the credit scoring method as described above is implemented.

[0035] Compared with the prior art, this application has the following beneficial effects:

[0036] The present application provides a training method for a credit scoring model, a credit scoring method and a related device. Specifically, when executing the training of the credit scoring model provided by the embodiment of the present application, a training set including multiple samples with credit scoring labels can be first obtained, and a sample includes multiple sample data with sample risk pattern labels, and the risk pattern label indicates that the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship or a non-monotonic relationship. The sample data includes the credit information of the sample user. Based on the training set and the credit scoring label, the initial scoring model is trained using the back propagation algorithm to obtain a credit scoring model, wherein the initial scoring model is a fully connected network neural network constructed based on a combined activation function. In the training set of the present application, each sample not only contains a credit scoring label, but also contains a sample risk pattern label, which clearly indicates the monotonic relationship between the sample data and the credit score (monotonic increasing, monotonic decreasing or non-monotonic). This helps to ensure that the model can capture the monotonic constraints between variables during the training process. At the same time, the initial scoring model is constructed using a combined activation function so that the model can better follow the monotonic constraints during the training process. The design of this combined activation function can ensure that the model output changes in a direction that is consistent with economic logic. For example, when the financial situation of a company improves, the credit risk prediction value decreases. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 A method flow chart of a credit scoring model training method provided in an embodiment of the present application;

[0039] Figure 2 A method flow chart of a method for constructing a combined activation function provided in an embodiment of the present application;

[0040] Figure 3 A method flow chart of a credit scoring method provided in an embodiment of the present application;

[0041] Figure 4 A schematic diagram of the structure of a credit scoring model training device provided in an embodiment of the present application;

[0042] Figure 5 A schematic diagram of the structure of a credit scoring device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0044] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.

[0045] Existing credit risk assessment models, such as logistic regression, decision trees, and traditional neural networks, can effectively predict corporate credit risk to a certain extent. These models generate probabilistic risk assessment results by analyzing the company's financial data, historical credit records, and other relevant features. For example, the logistic regression model can calculate the probability of credit default by linearly combining input features and applying the sigmoid function; the decision tree model divides different credit risk levels by constructing a series of conditional judgments; and the traditional neural network can capture more complex feature relationships through complex nonlinear transformations of multiple layers of neurons.

[0046] However, these models have an important flaw: they cannot guarantee the monotonicity of the prediction results. Specifically, this means that the model may generate prediction results that are contrary to economic logic in some cases. For example, when the financial situation of a company improves, according to the basic principles of economics, its credit risk should decrease accordingly. However, in the above model, due to the lack of monotonicity constraints, it may appear that the improvement of the company's financial situation is accompanied by an increase in credit risk.

[0047] This phenomenon that goes against economic intuition not only affects the interpretability and credibility of the model, but may also cause decision makers to make decisions based on incorrect predictions, thereby bringing potential risks and losses. For example, if the financial health of a company has improved significantly, but the model still predicts that its credit risk has increased, this will reduce the confidence of financial institutions or investors in the company, and may misjudge the credit status of the company, affecting the accuracy of decision-making.

[0048] In order to solve this problem, a training method for a credit scoring model, a credit scoring method and related devices are provided in an embodiment of the present application. First, a training set is constructed, which contains multiple samples with credit scoring labels, and each sample contains not only a credit scoring label, but also multiple sample data marked with risk patterns. The risk pattern label indicates the monotonic relationship between the sample data and the credit score, specifically a monotonically increasing, monotonically decreasing or non-monotonic relationship, and the sample data refers to the user's credit information. Next, based on the training set and the credit scoring label, the initial scoring model is trained by the back propagation algorithm. The initial model is a fully connected neural network based on a combined activation function. After training, the final credit scoring model is obtained. The present application introduces sample risk pattern labels in the training set and uses a combined activation function to ensure that the model can capture and follow the monotonic constraints between variables during the training process, so that the model output is more in line with economic logic. For example, when the financial situation of the enterprise improves, the credit risk prediction value decreases.

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] See also Figure 1 , which is a flow chart of a method for training a credit scoring model provided in an embodiment of the present application, such as Figure 1 As shown, the training of the credit scoring model may include steps S101-S102:

[0051] S101: Obtain a training set.

[0052] The process of obtaining a training set is one of the key steps in building a credit scoring model. In this process, the training set consists of multiple samples, each of which has a credit score label, representing the credit score level of the user corresponding to the sample. In order to improve the predictive ability of the model, the sample data not only contains the traditional credit score label, but also includes the risk pattern label of the sample. The risk pattern label indicates the monotonic relationship between each feature involved in the sample data and the credit score, that is, the direction of the impact of data changes on the credit score. These monotonic relationships can be divided into three types: monotonic increasing relationship, monotonic decreasing relationship and non-monotonic relationship.

[0053] Specifically, a monotonic increasing relationship means that when certain sample data (such as debt repayment ability, etc.) becomes better, the credit score should show a corresponding improvement; a monotonic decreasing relationship means that when certain sample data deteriorates, the credit score should decrease; a non-monotonic relationship indicates that there is no clear single trend between the data and the credit score, and there may be a complex nonlinear change relationship. These risk pattern labels can help the model better understand and capture the laws in the data, ensuring that the model can follow the monotonicity principle consistent with the actual economic logic during the training process.

[0054] In addition, the sample data also includes the user's credit information. If the sample user is an individual, the sample data includes the individual credit information of the sample user; if the sample user is an enterprise, the sample data includes the enterprise credit information of the sample user.

[0055] In a possible implementation, if the sample user is an individual, the sample user's personal credit information includes but is not limited to personal information, financial status information, repayment ability information, loan history information, guarantee and mortgage information, employment and career stability information, and social behavior and psychological assessment information.

[0056] (1) Personal information includes:

[0057] Identity information: including basic personal information such as name, gender, age, and contact information;

[0058] Living situation: The user’s current address, whether he or she has lived there for a long time, etc., can reflect his or her stability.

[0059] (2) Financial status information includes:

[0060] Income level: the borrower’s income source, income amount, and income stability;

[0061] Expenditure: monthly or annual expenditure of the borrower;

[0062] Debt situation: the total amount of debt the borrower currently has, including various loans, credit card balances, etc.;

[0063] Asset status: the borrower's fixed assets (such as real estate, cars, etc.) and current assets (such as cash, savings, etc.);

[0064] Bank deposits and working capital: reflect the borrower's current liquidity status and whether it has sufficient cash reserves to deal with emergencies.

[0065] (3) Repayment ability information includes:

[0066] Income-to-Debt Ratio: The ratio of a borrower’s debt to their income, used to assess their ability to repay debts;

[0067] Financial stability: assess whether the borrower has a long-term stable source of income and the volatility of its income;

[0068] Repayment history: whether the borrower has fulfilled past loan contracts on time and whether there is a record of overdue payments or defaults.

[0069] (4) Loan history information includes:

[0070] Loan types and amounts: types of loans (such as personal loans, housing loans, car loans, etc.) and amounts that the borrower has applied for;

[0071] Repayment record: the repayment status of each loan, whether there is overdue payment, early repayment, partial repayment, etc.;

[0072] Credit card usage: whether the borrower has multiple credit cards, how they are used, and whether there is an over-reliance on credit cards.

[0073] (5) Guarantee and mortgage information includes:

[0074] Guarantor or co-borrower: whether there is a guarantor or co-borrower, and their credit status;

[0075] Value and liquidity of collateral: whether collateral is provided for the loan, the valuation, market liquidity and risk of the collateral.

[0076] (6) Employment and career stability information including:

[0077] Workplace: industry, scale, stability, etc. of the borrower's company or unit;

[0078] Position and career history: whether the borrower has a stable career, whether he changes jobs frequently, the level of the position, etc.

[0079] (7) Social behavior and psychological assessment information includes:

[0080] Consumption behavior: the borrower’s past consumption patterns, such as whether there is a history of impulsive consumption or irrational spending;

[0081] Willingness to repay: Whether the borrower has the social and psychological willingness to repay. Some high-risk groups may have a lower willingness to repay.

[0082] In one possible implementation, if the sample user is an enterprise, the sample user's enterprise credit information includes but is not limited to the enterprise's basic information, financial statement information, credit record information, market performance information, industry background information, guarantee and mortgage information, corporate governance and management information, and social responsibility and risk management information and other related information.

[0083] (1) Basic information of the enterprise includes:

[0084] Company name and registration information: including company name, registered capital, legal representative, company establishment date, place of registration and other basic information;

[0085] Enterprise type and size: such as whether the enterprise is state-owned, private, joint venture or foreign-invested, and the size of the enterprise (small, medium, large), etc.;

[0086] Registered address and place of operation: The registered place and main place of operation of an enterprise reflect the geographical distribution and operational stability of the enterprise.

[0087] (2) Financial statement information includes:

[0088] Balance sheet: The assets, liabilities and net assets of a company, reflecting its debt-paying ability, financial stability and capital structure;

[0089] Profit and loss statement: the company's revenue, costs, profits, etc., to assess the company's profitability and operating efficiency;

[0090] Cash flow statement: the company's operating cash flow, investment cash flow and financing cash flow, to assess the company's cash liquidity and capital management capabilities;

[0091] Financial indicators: Key financial indicators such as debt-to-asset ratio, gross profit margin, net profit margin, current ratio, etc. are used to measure the financial health of the company.

[0092] (3) Credit record information includes:

[0093] Loan and debt history: the company's loan record, credit card usage and debt repayment history to assess the company's credit history and repayment ability;

[0094] Overdue and default records: whether the enterprise has any overdue or default behavior, and whether it has any other bad credit records;

[0095] Credit rating: The credit rating of an enterprise, including the credit rating given by a third-party credit rating agency (such as Dagong International, United Credit, etc.).

[0096] (4) Market performance information includes:

[0097] Sales performance and market share: the company's annual sales, market share, market performance of its main products, etc., to assess the company's market competitiveness;

[0098] Customer groups and brand influence: the company's main customer types, customer distribution, brand awareness, brand loyalty, etc., reflecting the company's position in the market;

[0099] Stock price performance (if listed company): If the enterprise is a listed company, its stock price fluctuations, shareholder structure, capital market performance, etc. should also be included.

[0100] (5) Industry background information includes:

[0101] Industry development trends: the development trends, competition landscape, industry cycles, etc. of the industry in which the enterprise operates, and the assessment of the external environment faced by the enterprise;

[0102] Industry policies and regulations: the impact of relevant policies and regulations on business operations, as well as the compliance and regulatory requirements of enterprises in the industry;

[0103] Industry competitiveness analysis: Comparative analysis of the company and its competitors in the same industry to evaluate the company's relative position in the industry.

[0104] (6) Guarantee and mortgage information includes:

[0105] Guarantor or co-borrower: whether the enterprise has a guarantor or co-borrower, and the credit status of the guarantor or co-borrower;

[0106] Value and liquidity of collateral: whether the enterprise has provided collateral or collateral, the valuation, market liquidity and risk of the collateral.

[0107] (7) Corporate governance and management information includes:

[0108] Shareholder structure and management team: the company's shareholder structure, the background of major shareholders, and the experience and stability of the management team;

[0109] Corporate governance structure: the composition of the company's board of directors, supervisory board and senior management team, and the assessment of the company's decision-making mechanism and governance level.

[0110] (8) Social responsibility and risk management information includes:

[0111] Corporate social responsibility: whether the company actively fulfills its social responsibilities and participates in public welfare and environmental protection projects;

[0112] Enterprise risk management capabilities: Whether the enterprise has a complete risk management system that can cope with market fluctuations, operational risks and emergencies.

[0113] This credit information is the basis for the training and prediction of the credit scoring model. By analyzing this information, the model can identify the key factors that affect the credit score and make effective predictions based on the changes in these factors. The introduction of risk pattern labels enables the model to clearly know the relationship between different types of features and credit scores during training, so as to more accurately learn the impact of these factors on credit scores.

[0114] In a possible implementation, the method further includes: cleaning and preprocessing the training set data, so that the credit scoring model is subsequently trained using the cleaned and preprocessed training set. Data cleaning includes processing missing values ​​and identifying and removing outliers. Data preprocessing includes, but is not limited to, standardization (Z-score) or normalization.

[0115] Standardization (Z-score): The income data can be transformed into a standard normal distribution with positive and negative values.

[0116] Normalization: compress the income data into a specific range and divide it into categories greater than 0 or less than 0 based on the threshold.

[0117] S102: Based on the training set and the credit score labels, the initial scoring model is trained using a back propagation algorithm to obtain a credit score model.

[0118] Based on the training set and credit score labels, the back propagation algorithm is used to train the initial scoring model to obtain a more accurate credit scoring model. The initial scoring model is a fully connected neural network built based on a combined activation function. Through this process, the model parameters can be optimized, the accuracy and generalization ability of the model can be improved, and the individual's credit risk can be better predicted.

[0119] Specifically, first, the training set is input into the initial scoring model to obtain an output score. Then, the difference between the output score and the credit score label is calculated to obtain a model error. Based on the model error, the back propagation algorithm and the loss function are used to iteratively adjust the parameters of the binary classifier to obtain the credit score model.

[0120] By inputting the training set and calculating the difference between the output score and the credit score label, the binary classifier is iteratively optimized using the back-propagation algorithm and loss function, which effectively improves the accuracy and generalization ability of the model and enables more accurate prediction of credit scores.

[0121] In a possible implementation, the loss function is a cross entropy loss function.

[0122] See also Figure 2 , Figure 2 A method flow chart of a method for constructing a combined activation function provided in an embodiment of the present application can be implemented through steps A1-A4:

[0123] A1: Select the original activation function.

[0124] The original activation function ρ is selected to be monotonically increasing, such as the activation function of the Rectified Linear Unit (ReLU) or its variants (such as the Exponential Linear Unit (ELU), the Scaled Exponential Linear Unit (SELU), etc.). This function will be used in the part of the network that needs to be monotonically increasing.

[0125] A2: Perform point reflection on the original activation function based on the origin to obtain the first activation function with opposite monotonicity to the original activation function.

[0126] For the original activation function ρ, a new activation function is obtained by a point reflection operation based on the origin Monotonicity is the opposite of the original activation function. Specifically, point reflection is to flip the original activation function (i.e., ρ(x)) symmetrically at the point x=0 to generate a new activation function The reflected function can be expressed mathematically as: Here, x represents the input data after data preprocessing. This means that for the original activation function ρ, when the input data is x, the output value given by ρ(x) will be processed in reverse: the input x is first negated (i.e. (-x)), then ρ is applied to this negative value, and then the final result is negated. This transformation makes the new activation function The monotonicity of is opposite to that of the original activation function ρ.

[0127] For example, assuming that the original activation function is ReLU (i.e., ρ(x) = max(0, x)), then after point reflection based on the origin, the resulting function is That is, when the input is positive, the output is negative, and vice versa.

[0128] A3: Construct a second activation function based on the first activation function and the original activation function.

[0129] When the input data x is less than 0, the second activation function The calculation method is ρ(x-1)-ρ(1); when the input data x is greater than 0, then This means that depending on the condition (x is positive or negative), by applying the first activation function By transforming the original activation function ρ, we can obtain a second activation function with different properties and behaviors. To better adapt to the model's requirements for different input ranges.

[0130] The second activation function The expression is:

[0131] A4: The original activation function, the first activation function and the second activation function are combined to obtain a combined activation function.

[0132] The original activation function, the first activation function, and the second activation function are combined into a new combined activation function ρ s . Wherein, j=1, 2 or 3, j=1 means that the risk pattern label of the input data x indicates that the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, j=2 means that the risk pattern label of the input data x indicates that the monotonic relationship between the sample data and the credit score is a monotonic decreasing relationship, and j=3 means that the risk pattern label of the input data x indicates that the monotonic relationship between the sample data and the credit score is a non-monotonic relationship; h j =|W T | t ·x+b, W represents the weight matrix of the fully connected network neural network; t represents the monotonicity indicator vector, when the monotonicity relationship between the input data x and the credit score is a monotonic increasing relationship, the value of t is 1 (that is, when j is 1, x is 1), when the monotonicity relationship between the input data x and the credit score is a monotonic decreasing relationship, the value of t is -1 (that is, when j is 2, x is -1), when the monotonicity relationship between the input data x and the credit score is a non-monotonic relationship, the value of t is 0 (that is, when j is 3, x is 0); b represents the bias term of the fully connected network neural network; represents the absolute value operation of the transpose of the weight matrix W according to the monotonicity indicator vector t.

[0133] In this process, s is a triple, s = (s1, s2, s3), which represents the number of neurons corresponding to different activation functions in the monotonicity-constrained fully connected layer of the neural network. The specific explanation is as follows:

[0134] s1 represents the first activation function The number of neurons in the layer structure where the original activation function ρ is located; s2 represents the number of neurons in the layer structure where the original activation function ρ is located; s3 represents the second activation function The number of neurons in the layer structure in which it is applied. These parameters s1, s2 and s3 reflect the number of neurons in each activation function at different levels, and by combining these activation functions, a new combined activation function ρ can be constructed. s , so that the activation function can flexibly adjust the output of the network according to the number of neurons in different layers in a multi-layer neural network. Combined activation function ρ s The design is to enhance the nonlinear ability of the network and improve the performance of the model by combining different types of activation functions to meet the monotonicity constraint.

[0135] In a possible implementation, the fully connected network neural network includes an input layer, multiple hidden layers connected in series or in parallel, and an output layer;

[0136] The construction process of the monotonicity constrained fully connected network includes:

[0137] Redefine the output function of the output layer of the fully connected network neural network based on the combined activation function to obtain the monotonicity constrained fully connected network;

[0138] Wherein, the output function of the output layer of the monotonicity-constrained fully connected network is y=ρ s (|W T | t ·X+b).

[0139] It can be seen that the output function y of the output layer of the monotonicity-constrained fully connected network is the combined activation function ρ s (h j ).

[0140] In a possible implementation, the method further includes:

[0141] The initial parameters of the initial credit scoring model are selected using a Bayesian optimization algorithm, including the number of neurons, network depth, learning rate, etc.

[0142] Based on the contents of S101-S102, it can be known that a sample training set containing multiple samples with credit score labels is first obtained, wherein each sample contains multiple sample data with risk pattern labels, and these risk pattern labels indicate the monotonic relationship between the sample data and the credit score (monotonic increasing, monotonically decreasing or non-monotonic), and the sample data is the credit information of the sample user. Then, based on the training set and the credit score label, the initial scoring model is trained using the back propagation algorithm, wherein the initial model is a fully connected neural network constructed based on a combined activation function to ensure that the model can better follow the monotonicity constraint and improve the prediction accuracy. This application introduces sample risk pattern labels and combined activation functions to ensure that the model follows the monotonic relationship between the credit score and the variable during the training process, thereby ensuring that the model output conforms to economic logic, such as a decrease in the credit risk prediction value when the finances improve.

[0143] Based on the above-mentioned credit scoring model training method embodiment, the present application embodiment also provides a credit scoring method. Figure 3 , Figure 3 A method flow chart of a credit scoring method provided in an embodiment of the present application. Figure 3 As shown, the method includes S301-S302:

[0144] S301: Obtain a rating dataset of target customers.

[0145] In order to score the target customer's credit, we first need to obtain the target customer's scoring dataset, which contains the target customer's credit information. The target customer's credit information is labeled with data risk pattern labels, which indicate the monotonic relationship between the target customer's credit information and the credit score, which can be monotonically increasing, monotonically decreasing, or non-monotonic. This process aims to establish a dataset about the target customer's credit status and annotate the data risk pattern to further analyze, predict, and manage the customer's credit risk.

[0146] It should be noted that the target customers, like the sample customers, can be either individuals or enterprises. Therefore, the credit information of the target customers is of the same type as the sample customers, and can be either personal credit information or enterprise credit information, which will not be elaborated here.

[0147] In a possible implementation manner, the method further includes: performing data preprocessing on the scoring data set.

[0148] Specifically, the scoring dataset may be standardized or normalized.

[0149] S302: Input the target customer's scoring data set into the credit scoring model for scoring, and obtain the target customer's credit score.

[0150] The target customer's score data set is input into the credit scoring model for scoring, thereby obtaining the target customer's credit score. This credit scoring model is trained using the training method of the credit scoring model described above.

[0151] Based on the contents of S301-S302, it can be seen that by using the credit scoring model to perform credit scoring on target customers based on the scoring data set, a clear monotonic relationship can be established between the output score and the credit information in the scoring data set, thereby making the scoring result more explainable and accurate.

[0152] See also Figure 4 , Figure 4 A structural diagram of a credit scoring model training device provided in an embodiment of the present application. Figure 4 As shown, the training device of the credit scoring model includes:

[0153] The training set acquisition unit 401 is used to acquire a training set; the training set includes a plurality of samples with credit score labels, and one of the samples includes a plurality of sample data with sample risk pattern labels; the risk pattern label indicates whether the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; the sample data includes the credit information of the sample user;

[0154] The model training unit 402 is used to perform model training on the initial scoring model based on the training set and the credit scoring label using a back propagation algorithm to obtain a credit scoring model; the initial scoring model is a fully connected network neural network constructed based on a combined activation function.

[0155] In a possible implementation manner, the device further includes:

[0156] The original activation function unit is used to select the original activation function ρ;

[0157] A point reflection unit is used to perform point reflection on the original activation function ρ based on the origin to obtain a first activation function that is monotonically opposite to the original activation function ρ. x represents the input data after data preprocessing;

[0158] An activation function construction unit, configured to construct an activation function based on the first activation function And the original activation function ρ constructs the second activation function

[0159] A combined activation function acquisition unit is used to combine the original activation function ρ, the first activation function And the second activation function Combine to obtain the combined activation function ρ s ; s = (s1, s2, s3); s is a triple, s1 represents the first activation function The number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, s2 represents the number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, and s3 represents the second activation function The layer structure is the number of neurons in the monotonicity constrained fully connected layer.

[0160] In a possible implementation, the fully connected neural network includes an input layer, multiple layers of hidden layers connected in series or in parallel, and an output layer.

[0161] In a possible implementation manner, the device further includes:

[0162] A monotonicity-constrained fully connected network construction unit, used to redefine the output function of the output layer of the fully connected network neural network based on the combined activation function to obtain the monotonicity-constrained fully connected network;

[0163] Wherein, the output function of the output layer of the monotonicity-constrained fully connected network is y=ρ s (|W T | t ·X+b, W represents the weight matrix of the fully connected network neural network; t represents the monotonicity indicator vector, when the monotonic relationship between the input data x and the credit score is a monotonic increasing relationship, the value of t is 1, when the monotonic relationship between the input data x and the credit score is a monotonic decreasing relationship, the value of t is -1, when the monotonic relationship between the input data x and the credit score is a non-monotonic relationship, the value of t is 0; b represents the bias term of the fully connected network neural network; |W T | t It represents the absolute value operation of the transpose of the weight matrix W according to the monotonicity indicator vector t.

[0164] In a possible implementation manner, the device further includes:

[0165] The first data preprocessing unit is used to perform standardization or normalization processing on the input data.

[0166] In a possible implementation, the model training unit 402 specifically includes:

[0167] An output score acquisition unit, used for inputting the training set into the initial scoring model to obtain an output score;

[0168] A model error acquisition unit, used for calculating the difference between the output score and the credit score label to obtain a model error;

[0169] An iterative adjustment unit is used to iteratively adjust the parameters of the binary classifier based on the model error by using the back propagation algorithm and the loss function to obtain the credit scoring model.

[0170] See also Figure 5 , Figure 5 This is a schematic diagram of the structure of a credit scoring device provided in an embodiment of the present application. Figure 5 As shown, the credit scoring device includes:

[0171] The scoring data set acquisition unit 501 is used to acquire a scoring data set of a target customer; the scoring data set includes the credit information of the target customer; the credit information of the target customer is annotated with a data risk model label; the data risk model label indicates whether the monotonic relationship between the credit information of the target customer and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship;

[0172] The input unit 502 is used to input the scoring data set of the target customer into the credit scoring model for scoring, so as to obtain the credit score of the target customer; the credit scoring model is trained according to the training method of the credit scoring model as described above.

[0173] In a possible implementation manner, the device further includes:

[0174] The second data preprocessing unit is used to standardize or normalize the scoring data set.

[0175] In addition, an embodiment of the present application also provides a training device for a credit scoring model, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the training method for the credit scoring model as described above is implemented.

[0176] In addition, an embodiment of the present application further provides a credit scoring device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the credit scoring method as described above is implemented.

[0177] The embodiment of the present application provides a training method for a credit scoring model, a credit scoring method and related devices. The construction process of the credit scoring model includes: collecting a training set containing samples with credit scoring labels, each sample containing multiple sample data with risk pattern labels as the credit information of the sample user, and these risk pattern labels clearly indicate the monotonic relationship between the sample data and the credit score (monotonic increasing, monotonically decreasing or non-monotonic). Then, using these training sets and credit scoring labels, the initial credit scoring model is trained by the back propagation algorithm, wherein the initial model is a fully connected neural network constructed based on a combined activation function. The present application achieves a clear definition and capture of the monotonic relationship by introducing sample risk pattern labels in the training set and using a combined activation function to construct a credit scoring model, ensuring that the model follows the monotonicity constraints during training and ensuring that the model output is consistent with the economic logic, such as when the financial situation of the enterprise improves, the credit risk prediction value will decrease.

[0178] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. Some or all of them can be selected according to actual needs to achieve the purpose of the embodiment scheme. Ordinary technicians in this field can understand and implement it without paying creative work.

[0179] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for training a credit scoring model, characterized in that: The method comprises: Acquire a training set; the training set includes a plurality of samples with credit score labels, and one of the samples includes a plurality of sample data with sample risk pattern labels; the risk pattern label indicates whether the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; the sample data includes the credit information of the sample user; Based on the training set and the credit score label, the initial scoring model is trained using a back propagation algorithm to obtain a credit scoring model; the initial scoring model is a fully connected network neural network constructed based on a combined activation function.

2. The method according to claim 1, characterized in that: The construction process of the combined activation function includes: Select the original activation function ρ; The original activation function ρ is subjected to point reflection based on the origin to obtain a first activation function which is monotonically opposite to the original activation function ρ. x represents the input data after data preprocessing; Based on the first activation function And the original activation function ρ constructs the second activation function The original activation function ρ, the first activation function And the second activation function Combine to obtain the combined activation function ρ s ; s = (s1, s2, s3); s is a triple, s1 represents the first activation function The number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, s2 represents the number of neurons in the layer structure where the original activation function ρ is located in the monotonicity constrained fully connected layer, and s3 represents the second activation function The layer structure is the number of neurons in the monotonicity constrained fully connected layer.

3. The method according to claim 2, characterized in that The fully connected network neural network includes an input layer, multiple hidden layers connected in series or in parallel, and an output layer; The construction process of the monotonicity constrained fully connected network includes: Redefine the output function of the output layer of the fully connected network neural network based on the combined activation function to obtain the monotonicity constrained fully connected network; Wherein, the output function of the output layer of the monotonicity-constrained fully connected network is y=ρ s (|W T | t ·X+b), W represents the weight matrix of the fully connected network neural network; t represents the monotonicity indicator vector, when the monotonic relationship between the input data x and the credit score is a monotonic increasing relationship, the value of t is 1, when the monotonic relationship between the input data x and the credit score is a monotonic decreasing relationship, the value of t is -1, when the monotonic relationship between the input data x and the credit score is a non-monotonic relationship, the value of t is 0; b represents the bias term of the fully connected network neural network; |W T | t It represents the absolute value operation of the transpose of the weight matrix W according to the monotonicity indicator vector t.

4. The method according to claim 2, characterized in that: The data preprocessing includes: The input data is standardized or normalized.

5. The method according to claim 1, characterized in that The step of training the initial scoring model based on the training set and the credit scoring label using a back propagation algorithm to obtain a credit scoring model includes: Inputting the training set into the initial scoring model to obtain an output score; Calculating the difference between the output score and the credit score label to obtain a model error; Based on the model error, the back propagation algorithm and the loss function are used to iteratively adjust the parameters of the binary classifier to obtain the credit scoring model.

6. A credit scoring method, characterized in that: The method comprises: Acquire a scoring data set of a target customer; the scoring data set includes the credit information of the target customer; the credit information of the target customer is annotated with a data risk pattern label; the data risk pattern label indicates whether the monotonic relationship between the credit information of the target customer and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; The scoring data set of the target customer is input into a credit scoring model for scoring to obtain the credit score of the target customer; the credit scoring model is trained according to the credit scoring model training method according to any one of claims 1 to 5.

7. A training device for a credit scoring model, characterized in that: The device comprises: A training set acquisition unit, used to acquire a training set; the training set includes a plurality of samples with credit score labels, one of the samples includes a plurality of sample data with sample risk pattern labels; the risk pattern label indicates whether the monotonic relationship between the sample data and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; the sample data includes the credit information of the sample user; A model training unit is used to perform model training on the initial scoring model based on the training set and the credit scoring label using a back propagation algorithm to obtain a credit scoring model; the initial scoring model is a fully connected network neural network constructed based on a combined activation function.

8. A credit scoring device, characterized in that: The device comprises: A scoring data set acquisition unit is used to acquire a scoring data set of a target customer; the scoring data set includes the credit information of the target customer; the credit information of the target customer is annotated with a data risk pattern label; the data risk pattern label indicates whether the monotonic relationship between the credit information of the target customer and the credit score is a monotonic increasing relationship, a monotonic decreasing relationship, or a non-monotonic relationship; An input unit is used to input the scoring data set of the target customer into a credit scoring model for scoring, so as to obtain the credit score of the target customer; the credit scoring model is trained according to the training method of the credit scoring model according to any one of claims 1 to 5.

9. A training device for a credit scoring model, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for training a credit scoring model as described in any one of claims 1 to 5 is implemented.

10. A credit scoring device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the credit scoring method according to claim 6 is implemented.

Images