Weight updating method and device, equipment and medium
The risk scoring model is used to infer counter-factually infer the data to be evaluated for bank loan renewal applications, and the weight of the scoring factor of the expert scoring model is identified and updated, which solves the problem of scoring inconsistency under the influence of human factors and improves the accuracy of the approval of loan renewal applications.
Patent Information
- Application Number
- CN202411645020.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-03-11
AI Technical Summary
In the approval of existing bank loan renewal applications, the scoring factors and weight settings of the expert experience model are easily affected by human factors, resulting in inconsistency and uncertainty in scoring and affecting the accuracy of approval.
By obtaining customer data that did not meet the conditions for renewal through the expert scoring model and actually met the conditions, input the risk scoring model for counterfactual inference calculation, determine the evaluation factors that lead to abnormal scoring, and update the scoring factor weights of the expert scoring model.
Significantly reduce the misjudgment rate of expert scoring models, improve the accuracy of scoring models, and thus improve the accuracy of loan renewal application approval.
Smart Images
Figure CN119598144A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of machine learning technology, and in particular to a weight updating method, device, equipment and medium. Background Art
[0002] In the bank's loan business, for small and micro enterprises that still have financing needs after the expiration of their working capital turnover loans but are temporarily facing financial difficulties, banks allow them to actively apply for loan renewal before the original loan expires. This policy is intended to avoid the financial pressure on small and micro enterprises due to loan expiration and ensure their normal operation. With the surge in the volume of loan renewal business, how to accurately approve loan renewal applications has become an important issue facing banks.
[0003] At present, banks mainly rely on expert experience models to evaluate loan renewal applications. The expert experience model is an evaluation method based on expert experience and subjective judgment of historical loan renewal data. It sets a series of scoring factors and corresponding weights, and evaluates customers' loan renewal qualifications by scoring loan renewal data such as customers' loan records, corporate financial data, and corporate operating indicators.
[0004] However, the scoring factors and corresponding weight settings of the expert experience model often rely on personal experience and subjective judgment, and are easily affected by human factors, resulting in inconsistency and uncertainty in the scoring, which in turn affects the accuracy of the approval of the renewal loan application. Summary of the invention
[0005] Based on the above problems, the present application provides a weight updating method, device, equipment and medium, which can improve the accuracy of approving renewal loan applications.
[0006] The embodiments of the present application disclose the following technical solutions:
[0007] In a first aspect, the present application discloses a weight updating method, the method comprising:
[0008] Acquire data to be evaluated, wherein the data to be evaluated is customer data that is determined by an expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions;
[0009] Inputting the data to be evaluated into a risk scoring model, performing counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determining the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally;
[0010] According to the evaluation factors, the scoring factor weights of the expert scoring model are updated.
[0011] Optionally, the risk scoring model is obtained by:
[0012] Get a sample customer dataset of users who apply for loan renewal;
[0013] According to the expert scoring model, a positive sample data set is determined from the sample customer data set, wherein the positive sample data set is a customer data set determined by the expert scoring model to meet the loan renewal conditions;
[0014] The positive sample data set is input into the counterfactual model, and the convolutional neural network model and the counterfactual attention module of the counterfactual model are trained by the positive sample data set to obtain a risk scoring model.
[0015] Optionally, the training of the convolutional neural network model and the counterfactual attention module of the counterfactual model by using the positive sample data set includes:
[0016] Establishing a feature channel for each evaluation factor in the positive sample data set through the convolutional neural network model of the counterfactual model;
[0017] Assigning an attention weight to each of the feature channels by a counterfactual attention module of the counterfactual model, wherein the attention weight is positively correlated with the evaluation factor weight of the evaluation factor;
[0018] Performing an element-by-element multiplication operation on the feature channel and the attention weight to obtain a weighted feature channel;
[0019] The counterfactual model is trained based on the weighted feature channels.
[0020] Optionally, inputting the positive sample data set into the counterfactual model includes:
[0021] Generate a positive sample image set according to the positive sample data set, wherein the positive sample image set represents each evaluation factor in the positive sample data set;
[0022] The positive sample image set is input into the counterfactual model.
[0023] Optionally, inputting the positive sample data set into the counterfactual model includes:
[0024] When all positive sample data in the positive sample data set are numerical values, obtaining the median and quartile range of multiple positive sample data;
[0025] Determining a normalized positive sample data set according to the positive sample data set, the median, and the interquartile range;
[0026] The normalized positive sample data set is input into the counterfactual model.
[0027] In a second aspect, the present application discloses a weight updating device, the device comprising: an acquisition module, a determination module, and an update module;
[0028] The acquisition module is used to acquire the data to be evaluated, wherein the data to be evaluated is customer data that is determined by the expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions;
[0029] The determination module is used to input the data to be evaluated into a risk scoring model, perform counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determine the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally;
[0030] The updating module is used to update the scoring factor weights of the expert scoring model according to the evaluation factors.
[0031] Optionally, the risk scoring model is obtained by the following units:
[0032] A data acquisition unit, used to acquire sample customer data of users applying for loan renewal;
[0033] A data determination unit, configured to determine positive sample data from the sample customer data according to the expert scoring model, wherein the positive sample data is customer data that is determined by the expert scoring model to meet the loan renewal conditions;
[0034] A model training unit is used to input the positive sample data into the counterfactual model, and train the convolutional neural network model and the counterfactual attention module of the counterfactual model through the positive sample data to obtain a risk scoring model.
[0035] Optionally, the model training unit is specifically used to: establish a feature channel for each evaluation factor in the positive sample data set through the convolutional neural network model of the counterfactual model; assign an attention weight to each of the feature channels through the counterfactual attention module of the counterfactual model, wherein the attention weight is positively correlated with the evaluation factor weight of the evaluation factor; perform element-by-element multiplication operation on the feature channel and the attention weight to obtain a weighted feature channel; and train the counterfactual model based on the weighted feature channel.
[0036] Optionally, the model training unit is specifically used to: generate a positive sample image set based on the positive sample data set, wherein the positive sample image set represents each evaluation factor in the positive sample data set; and input the positive sample image set into a counterfactual model.
[0037] Optionally, the model training unit is specifically used to: when the positive sample data in the positive sample data set are all numerical values, obtain the median and interquartile range of multiple positive sample data; determine the normalized positive sample data set based on the positive sample data set, the median and the interquartile range; and input the normalized positive sample data set into the counterfactual model.
[0038] In a third aspect, the present application discloses a computer device, the device comprising: a memory and a processor;
[0039] The memory is used to store programs;
[0040] The processor is used to execute the program to implement the various steps of the weight updating method as described in the first aspect.
[0041] In a fourth aspect, the present application discloses a computer-readable medium having a computer program stored thereon, and when the computer program is executed by a processor, the various steps of the weight updating method described in the first aspect are implemented.
[0042] Compared with the prior art, this application has the following beneficial effects:
[0043] The embodiments of the present application provide a weight updating method, device, equipment and medium. The method includes: obtaining data to be evaluated, the data to be evaluated is customer data that is determined by an expert scoring model to be not eligible for loan renewal conditions but actually meets the loan renewal conditions; inputting the data to be evaluated into a risk scoring model, performing counterfactual inference calculations on the data to be evaluated through the risk scoring model, and determining the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally; and updating the scoring factor weights of the expert scoring model based on the evaluation factors. Thus, the present application performs counterfactual inference calculations on the data to be evaluated through a risk scoring model, identifies the evaluation factors that cause misjudgment in the expert scoring model, and by updating the scoring factor weights of these evaluation factors, it can significantly reduce the misjudgment rate of the expert scoring model, thereby improving the accuracy of the expert scoring model, and further improving the accuracy of approving loan renewal applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in 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 labor.
[0045] Figure 1 A flowchart of a weight updating method provided in an embodiment of the present application;
[0046] Figure 2A A flowchart of another weight updating method provided in an embodiment of the present application;
[0047] Figure 2B A schematic diagram of a graphical processing provided in an embodiment of the present application;
[0048] Figure 2C A schematic diagram of a model training provided in an embodiment of the present application;
[0049] Figure 3 A schematic diagram of a weight updating device provided in an embodiment of the present application;
[0050] Figure 4 A schematic diagram of a computer-readable medium provided for an embodiment of the present application. DETAILED DESCRIPTION
[0051] As described above, in the process of approving loan renewal applications, banks mainly rely on expert experience models for evaluation. The expert experience model is an evaluation method based on expert experience and subjective judgment of historical loan renewal data. It sets a series of scoring factors and corresponding weights, and evaluates customers' loan renewal qualifications by scoring loan renewal data such as customers' loan records, corporate financial data, and corporate operating indicators.
[0052] However, the scoring factors and corresponding weight settings of the expert experience model often rely on personal experience and subjective judgment, and are easily affected by human factors, resulting in inconsistency and uncertainty in the scoring, which in turn affects the accuracy of the approval of the renewal loan application.
[0053] After research, the inventors proposed a weight updating method, device, equipment and medium. This method uses a risk scoring model to perform counterfactual inference calculations on the data to be evaluated, identifies the evaluation factors that lead to misjudgment in the expert scoring model, and updates the scoring factor weights of these evaluation factors. This can significantly reduce the misjudgment rate of the expert scoring model, thereby improving the accuracy of the expert scoring model, and further improving the accuracy of approving renewal loan applications.
[0054] 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.
[0055] See also Figure 1 , which is a flow chart of a model training method provided in an embodiment of the present application. The method includes:
[0056] S101: Acquire data to be evaluated, wherein the data to be evaluated is customer data that is determined by an expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions.
[0057] Data to be evaluated refers to customer data that is judged as not meeting the conditions for loan renewal after evaluation by the expert scoring model, but is confirmed to actually meet the conditions for loan renewal after manual review by the account manager. Exemplarily, customer data may include pre-loan data, such as financial statements, credit scores, historical loan records, etc., or post-loan data, such as repayment records, overdue conditions, etc., or difference data before and after the loan. This application does not limit the specific data to be evaluated.
[0058] For example, there is a customer A whose pre-loan data is as follows: "Financial statement: stable income, reasonable debt-to-asset ratio; credit score: good; historical loan record: on-time repayment, no overdue payment". After evaluation by the expert scoring model, customer A is judged to be ineligible for loan renewal. However, during the manual review process, the account manager found that customer A actually met the loan renewal conditions. In this case, customer A's pre-loan data will be recorded as data to be evaluated.
[0059] S102: Input the data to be evaluated into the risk scoring model, perform counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determine the evaluation factors in the data to be evaluated that cause abnormal scores of the expert scoring model.
[0060] Specifically, first, the data to be evaluated is input into the risk scoring model, and the Convolutional Neural Networks (CNN) of the risk scoring model is used to extract the feature map corresponding to the data to be evaluated. The feature map is obtained after the data is processed by the convolution layer, pooling layer, etc. It contains the core information of the data to be evaluated and is the basis for subsequent analysis.
[0061] Secondly, the counterfactual attention module of the risk scoring model is used to assign attention weights to the feature channels of each feature map. The purpose of assigning attention weights is to emphasize those features that have a greater impact on the risk scoring model and ignore those features that have a smaller impact, thereby improving the efficiency and accuracy of the analysis.
[0062] Subsequently, the feature map is element-wise multiplied with the attention weight to generate a weighted feature map.
[0063] Next, counterfactual inference is performed on the weighted feature graph. Counterfactual inference is an inference method based on known facts, which is used to infer the impact of changes in certain assessment factors (such as financial status, credit score, etc.) on the model output results (scores) of the risk scoring model, thereby finding out the key factors that affect the score.
[0064] Finally, by calculating the changes in the scores output by the risk scoring model before and after the counterfactual inference, we can determine which evaluation factors have the greatest impact on the risk scoring model scores. These evaluation factors are considered to be the factors that lead to abnormal scores of the expert scoring model, so that the expert scoring model can be optimized in a targeted manner.
[0065] S103: Update the scoring factor weights of the expert scoring model according to the evaluation factors.
[0066] Specifically, for those assessment factors that cause the risk scoring model to score too low or too high, that is, the actual score should be higher (or lower) but the score output by the risk scoring model is too low (or too high), the corresponding scoring factor weights should be adjusted in the expert scoring model.
[0067] It should be noted that the updating of the scoring factor weights in the expert scoring model may be performed multiple times in iterations to ensure that the updated expert scoring model can achieve the expected performance improvement.
[0068] To summarize, the embodiment of the present application provides a weight updating method, which performs counterfactual inference calculations on the evaluation data through a risk scoring model, identifies the evaluation factors that lead to misjudgment in the expert scoring model, and updates the scoring factor weights of these evaluation factors. This can significantly reduce the misjudgment rate of the expert scoring model, thereby improving the accuracy of the expert scoring model, and further improving the accuracy of the approval of renewal loan applications.
[0069] See Figure 2, which is a flow chart of a risk scoring method provided in an embodiment of the present application. The method includes:
[0070] S201: Obtain a sample customer data set of all customers who apply for loan renewal.
[0071] Customer data includes the customer's pre-loan data, post-loan data, and pre-loan and post-loan difference data. Customer data covers the customer's financial data, operating indicators, credit score, historical loan records, and other relevant data. For example, the customer data of all customers who apply for loan renewal can be obtained directly from the internal system of a bank or financial institution, or from an external data source such as a third-party data provider, or from the customer. This application does not limit the source of customer data.
[0072] It should be noted that in the process of collecting and processing customer data, relevant laws and regulations must be strictly observed to protect the security of customer personal information.
[0073] S202: Input the sample customer data set into the expert experience model, and determine the positive sample data set and the negative sample data set from the sample customer data set.
[0074] The positive sample data set includes multiple positive sample data. Positive sample data refers to customer data that has been reviewed by the expert experience model, and the loan applications of these customers have been judged to meet the conditions for loan renewal. Specifically, the collected customer data is first input into the expert experience model, and then the expert experience model reviews the customer data of each customer according to the preset scoring rules. Finally, the expert experience model outputs customer data that meets the conditions for loan renewal, and these data are positive sample data. Exemplarily, the pre-loan data of customer B is as follows: financial statement: stable income, reasonable debt-to-asset ratio; credit score: good; historical loan record: repayment on time, no overdue. The expert experience model scores these customer data and concludes that customer B's score is 85 points (assuming the scoring threshold is 80 points). Therefore, customer B's loan application is judged to meet the conditions for loan renewal, and its data is classified as positive sample data.
[0075] The counter-example sample data set includes multiple counter-example sample data. Counter-example sample data refers to customer data that did not pass the expert experience model review, but after manual intervention and review by the account manager, it was determined that the customer met the loan renewal conditions. In other words, the loan applications of these customers were initially rejected by the expert experience model, but after manual intervention and review by the account manager, it was determined that they met the loan renewal conditions. Specifically, for customer data that did not pass the expert experience model review, the account manager judged whether these customers met the loan renewal conditions based on his own experience and further investigation. If the account manager judged that a customer met the loan renewal conditions, these data are counter-example sample data.
[0076] For example, assume that the pre-loan data of customer C is as follows: financial statements: income fluctuates greatly, and the assets and liabilities are relatively high; credit score: general; historical loan record: there is one overdue repayment. The expert experience model scores based on these customer data and concludes that customer C's score is 75 points (assuming the scoring threshold is 80 points). Therefore, customer C's loan application was initially judged to be ineligible for renewal. However, during the manual review process, the account manager discovered that customer C's overdue repayment was due to force majeure, and customer C had taken remedial measures and his credit status was improving. Therefore, the account manager judged that customer C met the conditions for renewal, and his data was classified as counterexample sample data.
[0077] It is understandable that the counter-example sample data set is an important basis for adjusting the weights of the scoring factors of the expert experience model. By analyzing the counter-example sample data set, we can find out where the expert experience model has deviations, and then adjust the weights of the scoring factors corresponding to the deviations, thereby improving the accuracy of the expert experience model and better serving the customer's loan renewal needs.
[0078] S203: Obtain the loan renewal data set corresponding to the positive sample data set.
[0079] For the positive sample customers in the positive sample data set, we further collect the loan renewal data set of the positive sample customers (which is more detailed than the positive sample data set). The loan renewal data set includes pre-loan data and post-loan data. Among them, pre-loan data refers to the various operating and financial data of the enterprise before applying for loan renewal. Post-loan data refers to the various operating and financial data of the enterprise when the loan expires (or when the loan expires in 2 months), reflecting the operating status of the enterprise during the loan period.
[0080] Exemplarily, the loan renewal data set corresponding to small and micro enterprise A may include pre-loan data: the balance sheet shows accounts receivable of 1 million yuan, short-term debt of 500,000 yuan, and long-term loans of 2 million yuan; the income statement shows operating income of 5 million yuan, operating costs of 3 million yuan, and operating profit of 1 million yuan. Post-loan data: when the loan expires, accounts receivable become 1.2 million yuan, short-term debt becomes 450,000 yuan, and long-term loans become 1.9 million yuan; operating income becomes 5.2 million yuan, operating costs become 3.1 million yuan, and operating profit becomes 1.1 million yuan. See Table 1, which is a schematic table of loan renewal data provided in an embodiment of the present application.
[0081] Table 1
[0082]
[0083] It should be noted that if the customer data in the positive sample data set is detailed enough, step S203 may not be performed, and this application does not limit this.
[0084] S204: Preprocess the loan renewal data set to obtain a processed loan renewal data set.
[0085] In a specific implementation, preprocessing may refer to robust scaling normalization to ensure the consistency and comparability of the renewal loan dataset. Robust scaling normalization is mainly used to reduce the impact of outliers in the renewal loan dataset on the risk scoring model. In many practical application scenarios, especially in financial data, outliers are very common, such as extremely high loan amounts or abnormally high interest rates. Outliers may distort the distribution of the renewal loan dataset, causing the risk scoring model to pay too much attention to these extreme values during training, thereby affecting the generalization ability and prediction accuracy of the risk scoring model. By reducing the impact of outliers, the risk scoring model can be made more robust and better adapted to various data distributions. Specifically, first, when the data in the renewal loan dataset are all numerical values, their median is calculated. The median is the middle value of a set of data after sorting, and is not affected by extreme values. Secondly, their interquartile range (IQR) is calculated. The interquartile range (IQR) is the difference between the upper quartile (Q3) and the lower quartile (Q1), that is, IQR = Q3-Q1I. The interquartile range is also not affected by extreme values. Finally, based on the loan renewal data set, the median and the interquartile range, the normalized loan renewal data set is determined, which is the processed loan renewal data set. The specific determination method is to subtract the median from the data in the loan renewal data set, and then divide it by the interquartile range to obtain the processed loan renewal data set.
[0086] In another specific implementation, preprocessing may also refer to graphical processing. Graphic processing aims to convert multi-dimensional loan renewal information into a format suitable for processing by a convolutional neural network (CNN) model. Specifically, multiple data tables are first constructed based on the loan renewal data set, each of which represents an evaluation factor. For example, there may be a financial ratio table, a credit score table, a repayment record table, etc. Assuming that there are C types of evaluation factors in total, C data tables will be constructed. It should be noted that the maximum width of each data table is set to M, which is to ensure that the width of all data tables is consistent and convenient for subsequent image processing. If the actual width of a data table is less than M, it is filled with 0 to ensure that the width of the data table reaches M. In addition, the data of each customer is expanded into three parts: pre-loan data, post-loan data, and difference data between pre-loan and post-loan. This means that the actual amount of data for each customer is three times the original amount. N represents the total number of customer data entries, that is, N = 3 × the number of customers. Subsequently, for each data table of evaluation factors, it is converted into an image. The rows in each data table can be regarded as columns in the image, and the columns in each data table can be regarded as rows in the image. The width of the image is M and the height is N. In this way, each evaluation factor generates an M×N image. Finally, the images generated by each evaluation factor are combined into an image set, and each customer corresponds to a set of images. This image set is the processed loan renewal dataset.
[0087] It should be noted that if step S203 is not performed, then the median and interquartile range of multiple positive sample data can be directly obtained when the positive sample data in the positive sample data set are all numerical values; according to the positive sample data set, the median and interquartile range, the normalized positive sample data set, that is, the processed positive sample data set, is determined. Alternatively, a positive sample image set is generated directly based on the positive sample data set, and the positive sample image set represents each evaluation factor in the positive sample data set, and the generated positive sample image set is the processed positive sample data set.
[0088] S205: Input the processed loan renewal data set into the counterfactual model, and train the convolutional neural network model and counterfactual attention module of the counterfactual model through the processed loan renewal data set to obtain a risk scoring model.
[0089] See also Figure 2C , which is a schematic diagram of a model training provided in an embodiment of the present application. First, through the convolutional neural network model of the counterfactual model (i.e. Figure 2C The convolutional layer in ) establishes a feature channel for each evaluation factor in the processed loan renewal dataset, capturing the complex patterns in the processed loan renewal dataset.
[0090] Secondly, an attention weight is assigned to each feature channel through the counterfactual attention module of the counterfactual model. Among them, the attention weight is positively correlated with the evaluation factor weight of the evaluation factor, which means that the more important the evaluation factor, the higher the attention weight. Assume that the output tensor of a certain layer is represented as X, and its shape can be expressed as [N, C, H, W], where N is the batch size, C is the number of feature channels, and H and W represent the spatial height and width of the feature map, respectively. Specifically, a global average pooling operation is required for each feature channel to compress each H×W feature map into a scalar. The result of this step is a vector of shape [N, C], in which each element represents the average value of the corresponding channel. Subsequently, the vector obtained in the previous step is passed through a fully connected layer, and the ReLU activation function of the activation layer is applied. The purpose is to introduce nonlinear transformation and increase the expressive power. Subsequently, after passing through another fully connected layer, the Sigmoid activation function (i.e. Figure 2C The S-shaped function in ) converts it into an attention weight ranging between (0,1). The output of Sigmoid is also a vector of shape [N,C], where each element represents the importance of the corresponding channel.
[0091] Subsequently, the feature channel is element-wise multiplied with the attention weight to obtain the weighted feature channel, thereby strengthening the influence of key evaluation factors and weakening the role of unimportant evaluation factors. For each layer in the CNN, the attention module can be introduced in this way to scale the output of each layer.
[0092] Finally, the counterfactual model is trained based on the weighted feature channels to obtain the risk scoring model.
[0093] In some specific implementations, the loss function L of the risk scoring model may include a cross entropy loss function L CE and the counterfactual loss function L CF Among them, the cross entropy loss function L CE It is used to measure the difference between the model prediction result and the true label. In the weight updating method provided in the embodiment of the present application, L CE Used to measure the difference between the renewal label y predicted by the model and the actual label.
[0094] Counterfactual loss function L CF It is used to measure the improvement effect of original attention and counterfactual attention on the final prediction result. For example, the loss function of the risk scoring model can be shown as follows:
[0095] L=L CE (Y effect ,y)+λL CF (1)
[0096] Among them, λ is a hyperparameter used to balance the cross entropy loss function L CE and the counterfactual loss function L CF The weight of , y represents the loan renewal label.
[0097] Cross entropy loss function L CE Y effect is the overall effect of the dependent variable under the intervention of the attention mechanism, calculated by comparing the difference between the expected values when the attention mechanism is turned on and turned off. This helps to evaluate the net effect of the attention mechanism on the model's prediction results, and its calculation formula can be shown as follows:
[0098] Y effect =E[Y(do(A=A),X=X)-Y(do(A=A-),X=X)](2)
[0099] Among them, E[] represents the expected value, that is, the operation of taking the average under all possible conditions. In causal inference, the "do" operator represents external intervention on the variable A to make it equal to a specific value. do(A=A) means keeping the attention mechanism unchanged, that is, maintaining the attention distribution in the current state, and do(A=A-) means implementing the opposite intervention on the attention mechanism, that is, setting the attention mechanism to its opposite or default state. A- is the case of no attention mechanism, or the state where the attention mechanism is completely suppressed. Y() represents the dependent variable (result) under given conditions. X=X means that all other factors (covariates) except the attention mechanism remain unchanged. In summary, Y effect The meaning of is to calculate the difference between the expected value of the dependent variable (outcome) when the attention mechanism is turned on and when it is turned off, all other things being equal. In short, it is a way to assess the net effect of the attention mechanism on the outcome variable.
[0100] Counterfactual loss function L CF The calculation formula can be shown as follows:
[0101] L CF =||yy CF || 2 (3)
[0102] Among them, y is the reasoning result of the original attention, y CF is the inference result of counterfactual attention. This loss term measures the difference in the impact of original attention and counterfactual attention on the final prediction result. By minimizing this loss, the model can be encouraged to generate more influential attention maps.
[0103] Specifically, during the model training process, the loss function L of the risk scoring model is used to optimize the model parameters. effect and LCF , we can evaluate the contribution of the attention mechanism to the model performance. Finally, based on the performance on the validation set, we adjust λ to find the best balance.
[0104] S206: Acquire data to be evaluated, wherein the data to be evaluated is customer data that is determined by an expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions.
[0105] It should be noted that steps S206-S208 are similar to steps S101-S103 and will not be described in detail here.
[0106] S207: Input the data to be evaluated into the risk scoring model, perform counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determine the evaluation factors in the data to be evaluated that cause abnormal scores of the expert scoring model.
[0107] S208: Update the scoring factor weights of the expert scoring model according to the evaluation factors.
[0108] To summarize, the embodiment of the present application provides a weight updating method, which performs counterfactual inference calculations on the evaluation data through a risk scoring model, identifies the evaluation factors that lead to misjudgment in the expert scoring model, and updates the scoring factor weights of these evaluation factors. This can significantly reduce the misjudgment rate of the expert scoring model, thereby improving the accuracy of the expert scoring model, and further improving the accuracy of the approval of renewal loan applications.
[0109] See also Figure 3 , which is a schematic diagram of a weight updating device provided in an embodiment of the present application. The weight updating device 300 includes: an acquisition module 301, a determination module 302, and an update module 303.
[0110] The acquisition module 301 is used to acquire the data to be evaluated, wherein the data to be evaluated is the customer data that is determined by the expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions;
[0111] Determination module 302, for inputting the data to be evaluated into the risk scoring model, performing counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determining the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally;
[0112] The updating module 303 is used to update the scoring factor weights of the expert scoring model according to the evaluation factors.
[0113] In some specific implementations, the risk scoring model is obtained through the following units:
[0114] A data acquisition unit, used to acquire sample customer data of users applying for loan renewal;
[0115] A data determination unit, used to determine positive sample data from sample customer data according to the expert scoring model, wherein the positive sample data is customer data determined by the expert scoring model to meet the loan renewal conditions;
[0116] The model training unit is used to input positive sample data into the counterfactual model, train the convolutional neural network model and the counterfactual attention module of the counterfactual model through the positive sample data, and obtain the risk scoring model.
[0117] In some specific implementations, the model training unit is specifically used to: establish a feature channel for each evaluation factor in the positive sample data set through a convolutional neural network model of a counterfactual model; assign an attention weight to each feature channel through a counterfactual attention module of the counterfactual model, wherein the attention weight is positively correlated with the evaluation factor weight of the evaluation factor; perform element-by-element multiplication operation on the feature channel and the attention weight to obtain a weighted feature channel; and train the counterfactual model based on the weighted feature channel.
[0118] In some specific implementations, the model training unit is specifically used to: generate a positive sample image set based on the positive sample data set, the positive sample image set representing each evaluation factor in the positive sample data set; and input the positive sample image set into the counterfactual model.
[0119] In some specific implementations, the model training unit is specifically used to: obtain the median and interquartile range of multiple positive sample data when the positive sample data in the positive sample data set are all numerical values; determine the normalized positive sample data set based on the positive sample data set, the median and the interquartile range; and input the normalized positive sample data set into the counterfactual model.
[0120] To summarize, the embodiment of the present application provides a weight updating device, which performs counterfactual inference calculations on the evaluation data through a risk scoring model, identifies the evaluation factors that lead to misjudgment in the expert scoring model, and updates the scoring factor weights of these evaluation factors. This can significantly reduce the misjudgment rate of the expert scoring model, thereby improving the accuracy of the expert scoring model, and further improving the accuracy of the approval of renewal loan applications.
[0121] The embodiments of the present application also provide corresponding computer devices and computer-readable media for implementing the model training method provided in the embodiments of the present application.
[0122] Among them, the computer device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute instructions or codes so that the device executes a model training method of any embodiment of the present application.
[0123] See also Figure 4, which is a schematic diagram of a computer readable medium provided in an embodiment of the present application. The computer readable medium 400 stores a computer program 411, which implements the above-mentioned Figure 1 The steps of the model training method.
[0124] It should be noted that in the context of the present application, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0125] It should be noted that the machine-readable medium mentioned above in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.
[0126] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.
[0127] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.
[0128] Although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present application. Certain features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination.
[0129] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.
Claims
1. A weight updating method, characterized in that: The method comprises: Acquire data to be evaluated, wherein the data to be evaluated is customer data that is determined by an expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions; Inputting the data to be evaluated into a risk scoring model, performing counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determining the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally; According to the evaluation factors, the scoring factor weights of the expert scoring model are updated.
2. The method according to claim 1, characterized in that The risk scoring model is obtained in the following way: Get a sample customer dataset of users who apply for loan renewal; According to the expert scoring model, a positive sample data set is determined from the sample customer data set, wherein the positive sample data set is a customer data set determined by the expert scoring model to meet the loan renewal conditions; The positive sample data set is input into the counterfactual model, and the convolutional neural network model and the counterfactual attention module of the counterfactual model are trained by the positive sample data set to obtain a risk scoring model.
3. The method according to claim 2, characterized in that The training of the convolutional neural network model and the counterfactual attention module of the counterfactual model by using the positive sample data set includes: Establishing a feature channel for each evaluation factor in the positive sample data set through the convolutional neural network model of the counterfactual model; Assigning an attention weight to each of the feature channels by a counterfactual attention module of the counterfactual model, wherein the attention weight is positively correlated with the evaluation factor weight of the evaluation factor; Performing an element-by-element multiplication operation on the feature channel and the attention weight to obtain a weighted feature channel; The counterfactual model is trained based on the weighted feature channels.
4. The method according to claim 3, characterized in that The step of inputting the positive sample data set into the counterfactual model comprises: Generate a positive sample image set according to the positive sample data set, wherein the positive sample image set represents each evaluation factor in the positive sample data set; The positive sample image set is input into the counterfactual model.
5. The method according to claim 2, characterized in that: The step of inputting the positive sample data set into the counterfactual model comprises: When all positive sample data in the positive sample data set are numerical values, obtaining the median and quartile range of multiple positive sample data; Determining a normalized positive sample data set according to the positive sample data set, the median, and the interquartile range; The normalized positive sample data set is input into the counterfactual model.
6. A weight updating device, characterized in that: The device comprises: an acquisition module, a determination module, and an update module; The acquisition module is used to acquire the data to be evaluated, wherein the data to be evaluated is customer data that is determined by the expert scoring model to be unqualified for loan renewal but actually meets the loan renewal conditions; The determination module is used to input the data to be evaluated into a risk scoring model, perform counterfactual inference calculation on the data to be evaluated through the risk scoring model, and determine the evaluation factors in the data to be evaluated that cause the expert scoring model to score abnormally; The updating module is used to update the scoring factor weights of the expert scoring model according to the evaluation factors.
7. The device according to claim 6, characterized in that The risk scoring model is obtained through the following units: A data acquisition unit, used to acquire sample customer data of users applying for loan renewal; A data determination unit, configured to determine positive sample data from the sample customer data according to the expert scoring model, wherein the positive sample data is customer data that is determined by the expert scoring model to meet the loan renewal conditions; A model training unit is used to input the positive sample data into the counterfactual model, and train the convolutional neural network model and the counterfactual attention module of the counterfactual model through the positive sample data to obtain a risk scoring model.
8. The device according to claim 7, characterized in that The model training unit is specifically used to: establish a feature channel for each evaluation factor in the positive sample data set through the convolutional neural network model of the counterfactual model; An attention weight is assigned to each of the feature channels through the counterfactual attention module of the counterfactual model, wherein the attention weight is positively correlated with the evaluation factor weight of the evaluation factor; an element-by-element multiplication operation is performed on the feature channel and the attention weight to obtain a weighted feature channel; and the counterfactual model is trained based on the weighted feature channel.
9. A computer device, characterized in that: The device comprises: a memory and a processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the weight updating method as described in any one of claims 1 to 5.
10. A computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the weight updating method according to any one of claims 1 to 5 is implemented.