Method, device, equipment and storage medium for adaptively constructing risk control models

Abstract

The present application discloses a method for adaptively constructing a risk control model, including: obtaining multi-dimensional feature data of a customer, and classifying and storing the multi-dimensional feature data, obtaining feature data of each category, creating a supervision target corresponding to the customer, and fusing and optimizing the preset business target and the supervision target to obtain a fusion target, constructing a risk control base model corresponding to each category feature data based on the feature data of each category, the preset business target and the fusion target, obtaining a target label of a new task, and inputting the target label of the new task into each risk control base model to evaluate each risk control base model, obtaining each model to be fused corresponding to a preset screening constraint condition, fusing each model to be fused based on the target label of the new task, and evaluating the fused model to select a target credit risk control model corresponding to the preset fusion constraint condition. The present application solves the technical problem of low efficiency in adaptive model construction.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology in financial technology (Fintech), and in particular to a method, device, equipment and storage medium for adaptively constructing a risk control model. Background Art

[0002] With the continuous development of financial technology, especially Internet technology finance, more and more technologies (such as distributed, artificial intelligence, etc.) are being applied in the financial field.

[0003] With the development of computer technology, the application of machine learning is becoming more and more extensive. At present, the model can be constructed by obtaining massive amounts of data from the Internet, and the risk control model can be constructed based on these data, so as to predict risks through the risk control model and help related financial enterprises reduce credit risks. Furthermore, when the model encounters a new task, it usually sets certain rules to re-screen features, and search for the optimal model parameters based on the screened features, so as to adaptively build the risk control model. However, in the process of re-screening features, it is necessary to evaluate and screen the features in all feature layers according to the goals of the new task. As the number of underlying features increases, the parameter space that the model needs to search also increases rapidly. Therefore, a lot of computing time is required, and when some underlying features migrate or change dramatically, the model needs to be readjusted, which will also consume a lot of computing resources and time costs, resulting in low efficiency in model adaptive construction. Summary of the invention

[0004] The main purpose of this application is to provide a method, device, equipment and storage medium for adaptively building a risk control model, aiming to solve the technical problem of low efficiency of adaptive model construction in the prior art.

[0005] To achieve the above objectives, the present application provides a method for adaptively constructing a risk control model, the method comprising:

[0006] Acquire multi-dimensional feature data of the customer, and classify and store the multi-dimensional feature data to obtain feature data of each category;

[0007] Create supervision targets corresponding to customers, and integrate and optimize preset business targets and the supervision targets to obtain integrated targets;

[0008] Based on each of the category feature data, the preset business goal and the fusion goal, construct a risk control base model corresponding to each of the category feature data;

[0009] Obtaining the target label of the new task, and inputting the target label of the new task into each of the risk control base models to evaluate each of the risk control base models, and obtaining each to-be-fused model corresponding to the preset screening constraint condition;

[0010] Based on the target label of the new task, the models to be integrated are integrated, and the integrated models are evaluated to select the target credit risk control model that meets the preset integration constraints.

[0011] The present application also provides a risk control model adaptive construction device, the risk control model adaptive construction device is a virtual device, and the risk control model adaptive construction device includes:

[0012] A data storage module, used to obtain multi-dimensional feature data of customers, and classify and store the multi-dimensional feature data to obtain feature data of each category;

[0013] The fusion module is used to create the supervision target corresponding to the customer, and fuse and optimize the preset business target and the supervision target to obtain the fusion target

[0014] A model building module, used to build a risk control base model corresponding to each of the category feature data based on the category feature data, the preset business goal and the fusion goal;

[0015] A screening module is used to obtain the target label of the new task and input the target label of the new task into each of the risk control base models to evaluate each of the risk control base models and obtain each model to be fused that satisfies the preset screening constraint conditions;

[0016] The model combination module is used to fuse the models to be fused based on the target label of the new task, and evaluate the fused models to select the target credit risk control model that meets the preset fusion constraint conditions.

[0017] The present application also provides a risk control model adaptive construction device, which is a physical device. The risk control model adaptive construction device includes: a memory, a processor, and a risk control model adaptive construction program stored in the memory. The risk control model adaptive construction program is executed by the processor to implement the steps of the risk control model adaptive construction method as described above.

[0018] The present application also provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores a risk control model adaptive construction program, and the risk control model adaptive construction program is executed by a processor to implement the steps of the risk control model adaptive construction method as described above.

[0019] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned risk control model adaptive construction method.

[0020] The present application provides a method, device, equipment and storage medium for adaptively constructing a risk control model. Compared with the technical means adopted in the prior art to construct a model by screening feature data for combination to obtain optimal model parameters, the present application first obtains multi-dimensional feature data of the customer, and classifies and stores the multi-dimensional feature data to obtain feature data of each category, and then creates a supervision target corresponding to the customer, and fuses and optimizes the preset business target and the supervision target to obtain a fusion target. Further, based on the feature data of each category, the preset business target and the fusion target, a risk control base model corresponding to each category feature data is constructed, and the corresponding risk control base model is jointly constructed according to different category feature data, preset business target and the fusion target, so that the risk control base model can adapt to new tasks of various categories and improve the robustness of the risk control base model. Further, the target label of the new task is obtained, and the target label of the new task is input into each risk control base model to evaluate each risk control base model to obtain each model to be fused corresponding to the preset screening constraint conditions, and then based on the The target label of the new task is obtained, each of the models to be fused is fused, and the fused model is evaluated to select the target credit risk control model corresponding to the preset fusion constraint condition, so as to evaluate and screen the existing risk control base models according to the target of the new task, and the number of risk control base models is significantly lower than the number of features, which can greatly reduce the computing time cost, and then the screened risk control base models are combined to complete the adaptation to the new task. Furthermore, when some important features change and migrate, only some risk control base models need to be retrained and the weights of related fusion models need to be adjusted, thereby reducing the search space of parameters in the screening process and reducing the consumption of computing resources. The problem in the prior art that when encountering a new task, in the process of re-screening features, it is necessary to evaluate and screen the features in all feature layers according to the target of the new task, and when some underlying features migrate or change drastically, the model needs to be readjusted, which will also consume a lot of computing resources and time costs, thereby leading to the technical defect of low efficiency of model adaptive construction, thereby improving the efficiency of model adaptive construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application 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, for those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0023] Figure 1 This is a flow chart of the first embodiment of the method for adaptively constructing a risk control model of the present application;

[0024] Figure 2 This is a flow chart of the second embodiment of the method for adaptively constructing a risk control model of the present application;

[0025] Figure 3 This is a flow chart of the third embodiment of the method for adaptively constructing a risk control model of the present application;

[0026] Figure 4 Schematic diagram of the structure of the risk control model adaptive construction device of the hardware operating environment involved in the risk control model adaptive construction method in the embodiment of the present application.

[0027] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0029] The present application embodiment provides a method for adaptively constructing a risk control model. In the first embodiment of the method for adaptively constructing a risk control model of the present application, refer to Figure 1 , the risk control model adaptive construction method includes:

[0030] Step S10, obtaining multi-dimensional feature data of the customer, and classifying and storing the multi-dimensional feature data to obtain feature data of each category;

[0031] In this embodiment, it should be noted that in the actual application scenario of credit, the category characteristic data includes the characteristic data of the People's Bank of China, internal characteristic data of the loan, characteristic data of customer static information, characteristic data of customer equipment and characteristic data corresponding to the interaction between the customer and the APP application software, etc.

[0032] Acquire multi-dimensional feature data of customers, and classify and store the multi-dimensional feature data to obtain feature data of each category. Specifically, acquire multi-dimensional feature data of customers, and then classify and store the multi-dimensional feature data according to preset classification rules to obtain feature data of each category, including rules for classification according to data sources and data characterization dimensions.

[0033] The step of classifying and storing the multi-dimensional feature data to obtain feature data of each category includes:

[0034] Step S11, classify and store the multi-dimensional feature data according to the data source and the dimension of describing the customer characteristics, and obtain the feature data of each category.

[0035] In this embodiment, it should be noted that the feature data corresponding to the characterization dimension of the customer characteristics is feature data of different dimensions of the customer, for example, customer static information feature data. Specifically, according to the data source of the original data corresponding to the multi-dimensional feature data and the characterization dimension of the customer characteristics, the multi-dimensional feature data is classified and stored to obtain the feature data of each category.

[0036] Step S20, creating a supervision target corresponding to the customer, and fusing and optimizing the preset business target and the supervision target to obtain a fusion target;

[0037] In this embodiment, it should be noted that the preset business goal is a series of multi-angle business goals based on a summary of historical business experience, such as short-term overdue, long-term overdue and fraud. If the preset business goal is the customer's short-term overdue situation, 0 and 1 are often used to represent the quality of the customer. Furthermore, unlike the preset business goal, the supervision target is the specific business information corresponding to the customer. The supervision target is a continuous value, for example, the customer's overdue amount, number of overdue days, etc.

[0038] Create a supervision target corresponding to the customer, and fuse and optimize the preset business target and the supervision target to obtain a fusion target. Specifically, create a supervision target corresponding to the customer, and then iteratively weighted optimize the existing preset business target through a preset optimization method, wherein the preset optimization method includes a gradient descent method to determine the weight coefficient corresponding to each business target in the preset business target, and then obtain the optimization target corresponding to the weight coefficient. Further, calculate the square of the Pearson correlation coefficient between the optimization target corresponding to the weight coefficient and the supervision target. When the square of the Pearson correlation coefficient reaches the maximum, the optimization target corresponding to the weight coefficient is discretized based on the numerical threshold to obtain the fusion target, so that the risk control base model constructed based on the fusion target can better adapt to new tasks.

[0039] The step of fusing and optimizing the preset business goal and the supervision goal to obtain the fusion goal includes:

[0040] Step S21, iteratively optimizing the preset business goal and the supervision goal by a preset optimization method to determine the weight coefficient corresponding to the preset business goal, and obtaining the optimization goal corresponding to the weight coefficient;

[0041] Step S22: If the square of the Pearson correlation coefficient between the optimization target and the supervision target meets the preset coefficient threshold condition, the optimization target is discretized to obtain the fusion target.

[0042] In this embodiment, it should be noted that the preset coefficient threshold condition is a condition corresponding to the square of the Pearson correlation coefficient reaching the maximum.

[0043] Specifically, the preset business objectives and the supervision objectives are iteratively optimized through a preset optimization method to obtain the weight coefficients corresponding to each business objective in the preset business objectives, and then determine whether the square of the Pearson correlation coefficient between the optimization objective corresponding to the weight coefficient and the supervision objective meets the preset coefficient threshold condition. If so, the weight coefficient is proved. Further, at least one numerical threshold is set to discretize the optimization objective corresponding to the weight coefficient to obtain the fusion objective. For example, the optimization objective less than the numerical threshold is set to 0, and the optimization objective greater than the numerical threshold is set to 1.

[0044] The optimization method is as follows: Y0n=r1*Yo1+r2*Yo2+r3*Yo3, wherein Y0n is the fusion target, Yo1 corresponds to business target 1, Yo2 corresponds to business target 2, Yo3 corresponds to business target 3, r1, r2, r3 are decimals ranging from 0 to 1, and r1+r2+r3=1, and then a set of r1, r2, r3 is obtained by the gradient descent method so that the square of the Pearson correlation coefficient between the optimization target Y0n and the supervision target Z is maximized, and then after r1, r2, r3, the optimization target is discretized, for example, a numerical threshold is set, the optimization target less than the numerical threshold is set to 0, and the optimization target greater than the numerical threshold is set to 1, and then the fusion target is obtained. The numerical threshold can be a fixed value of 0.5, or multiple sets of numerical thresholds can be taken to obtain multiple sets of fusion targets.

[0045] Step S30, constructing a risk control base model corresponding to each of the category feature data based on the category feature data, the preset business goal and the fusion goal;

[0046] In this embodiment, it should be noted that the risk control base models include LR logistic regression model, gradient boosting decision tree model (GBDT), neural network model, support vector machine SVM, random forest RF and other models, and the risk control base models corresponding to the category feature data are not the same.

[0047] Based on each of the category feature data, the preset business goal and the fusion goal, a risk control base model corresponding to each of the category feature data is constructed. Specifically, the following steps are performed for each of the category feature data:

[0048] The category feature data is associated with the preset business goal and the fusion goal, wherein the association includes associating one category feature data with a preset business goal or a fusion goal, or associating multiple category feature data with a preset business goal or a fusion goal, or associating one category feature with multiple preset business goals or multiple fusion goals, and then based on the category feature data and the preset business goals and fusion goals associated with the category feature data, a risk control base model is constructed by the base model construction method corresponding to the category feature data, so as to obtain the risk control base model corresponding to each category feature data, and the fusion target is modeled, which can improve the robustness of the base model and better adapt to new tasks. For example, the feature vector corresponding to the A category feature data is (xa1 ,xa2,xa3,xa4…), the corresponding feature data of category B are (xb1,xb2,xb3,xb4…), the corresponding feature data of category C are (xc1,xc2,xc3,xc4…), the corresponding feature data of category D are (xd1,xd2,xd3,xd4…), the preset business objectives include business objective 1, business objective 2 and business objective 3, and then the feature data of category A are combined with business objective 1 to construct the risk control base model M1 using the logistic regression method (LR), the feature data of category B are combined with business objective 2 to construct the risk control base model M2 using the gradient boosting decision tree method (GBDT), the feature data of category C are combined with business objective 3 to construct the risk control base model M3 using the neural network method (NN), and the feature data of category D are combined with the fusion objective to construct the risk control base model M4 using the neural network method (NN).

[0049] Step S40, obtaining a target tag for a new task, and inputting the target tag for the new task into each of the risk control base models to evaluate each of the risk control base models, and obtaining each to-be-fused model corresponding to a preset screening constraint condition;

[0050] In this embodiment, it should be noted that the new task is a task with a prediction target different from that of the existing model. For example, the current model predicts that the customer is short-term overdue for 3 months or is a fraud. When the prediction target needs to be the corresponding situation of long-term overdue for 6 months, the 6-month long-term overdue is taken as a new task. The evaluation method includes but is not limited to the KS (Kolmogorov-Smirnov) indicator method, the PSI (Population Stability Index) group stability indicator method, and the average inference time of the model, etc. The preset screening constraints can be set according to actual task requirements.

[0051] Obtain a target label for the new task, and input the target label for the new task into each of the risk control base models to evaluate each of the risk control base models and obtain each model to be fused that satisfies the preset screening constraints. Specifically, obtain the target label for the new task, and then input the target label corresponding to the new task into each of the risk control base models according to the target label of the new task, and calculate the evaluation indicators corresponding to each of the risk control base models, wherein the evaluation indicators include indicators such as the KS indicator and the PSI indicator. Then, according to the evaluation indicators corresponding to each of the risk control base models, and in accordance with the preset screening constraints, screen each of the risk control base models, so that the screened risk control base models can be used as the models to be fused, thereby realizing the evaluation and screening of the existing risk control base models according to the target label of the new task, so as to combine the screened risk control base models, and complete the adaptation to the new task.

[0052] The step of inputting the target label of the new task into each of the risk control base models to evaluate each of the risk control base models and obtaining each model to be fused corresponding to the preset screening constraint condition includes:

[0053] Step S41, based on the target label of the new task, evaluate each of the risk control base models to obtain the evaluation index corresponding to each of the risk control base models;

[0054] In this embodiment, based on the target label of the new task, each of the risk control base models is evaluated to obtain the evaluation indicators corresponding to each of the risk control base models. Specifically, the target label of the new task is input into each of the risk control base models, and the operating effect of each of the risk control base models is evaluated to obtain the evaluation indicators corresponding to each of the risk control base models.

[0055] Step S42: based on the preset screening constraints and the evaluation indicators corresponding to the risk control base models, the risk control base models are screened to obtain the models to be integrated.

[0056] In this embodiment, based on the preset screening constraints and the evaluation indicators corresponding to each of the risk control base models, each of the risk control base models is screened to obtain each of the models to be fused. Specifically, based on the evaluation indicators corresponding to each of the risk control base models, each of the risk control base models that satisfies the preset screening constraints is selected to obtain each of the models to be fused. For example, the preset screening constraints for the risk control base models are (f1, f2, ...), wherein f1: the model KS index is greater than 20, and f2: the model PSI index is less than 0.01. Then, according to the target Ynew of the new task input, each of the risk control base models is evaluated to obtain various evaluation indicators corresponding to each of the risk control models, such as the discrimination KS index and the stability PSI index. Furthermore, according to each of the evaluation indicators, the qualified risk control base models are screened according to the preset screening constraints (f1, f2, ...) of the risk control base models to carry out the next step of model fusion.

[0057] Step S50, based on the target label of the new task, the models to be integrated are integrated, and the integrated models are evaluated to select the target credit risk control model that meets the preset integration constraint conditions.

[0058] In this embodiment, it should be noted that the fusion constraints include constraints corresponding to the number of fusion models, the discrimination KS index, and the stability PSI index.

[0059] Based on the target label of the new task, each of the models to be fused is fused, and the fused model is evaluated to select a target credit risk control model that satisfies the preset fusion constraint conditions. Specifically, based on the target label of the new task, each of the models to be fused is weightedly fused, wherein the weighted fusion method includes linear weighted fusion and nonlinear weighted fusion methods, such as decision tree nonlinear fusion. In the present application, preferably, linear weighted fusion is selected for model fusion to obtain fusion models corresponding to different weights, and then the model evaluation indicators corresponding to the fusion models with different weights are calculated respectively, wherein the model evaluation indicators include the discrimination KS indicator, the stability PSI indicator and other indicators, and then it is judged whether the model evaluation indicators corresponding to the fusion models corresponding to the different weights respectively meet the preset fusion constraint conditions, so as to obtain a preset number of target credit risk control models that meet the preset fusion constraint conditions, wherein the preset number can be 1 or more, which is not limited here.

[0060] The step of fusing the models to be fused based on the target label of the new task and evaluating the fused model to select the target credit risk control model that satisfies the preset fusion constraint condition includes:

[0061] Step S51, based on the target label of the new task, weighted fusion is performed on the models to be fused to obtain fusion models corresponding to different weights;

[0062] In this embodiment, based on the target label of the new task, the models to be fused are weighted and fused to obtain fusion models corresponding to different weights. Specifically, based on the target label of the new task, the models to be fused are weighted and fused to obtain fusion models corresponding to different weights. For example, the models to be fused are fused using a logistic regression method.

[0063] Step S52, respectively calculating model evaluation indicators corresponding to fusion models with different weights;

[0064] In this embodiment, the model evaluation indicators corresponding to the fusion models of different weights are calculated respectively. Specifically, the model evaluation indicators corresponding to the fusion models of different weights are calculated respectively by preset indicator calculation methods, wherein the preset indicator calculation methods include KS (Kolmogorov-Smirnov) indicator method, PSI (Population Stability Index) group stability indicator method and other methods.

[0065] Step S53: Based on the model evaluation indicators, select the target credit risk control model that meets the preset fusion constraint conditions.

[0066] In this embodiment, based on each of the model evaluation indicators, a target credit risk control model corresponding to the preset fusion constraint condition is selected. Specifically, according to the model evaluation indicators corresponding to the fusion models with different weights, if there are multiple fusion models with different weights whose corresponding model evaluation indicators satisfy the preset fusion constraint condition, a preset number of target credit risk control models corresponding to the demand are obtained.

[0067] The embodiment of the present application provides a method for adaptively constructing a risk control model. Compared with the technical means adopted in the prior art to construct a model by screening feature data for combination to obtain optimal model parameters, the embodiment of the present application first obtains multi-dimensional feature data of the customer, and classifies and stores the multi-dimensional feature data to obtain feature data of each category, and then creates a supervision target corresponding to the customer, and fuses and optimizes the preset business target and the supervision target to obtain a fusion target. Further, based on the feature data of each category, the preset business target and the fusion target, a risk control base model corresponding to each category feature data is constructed, and the corresponding risk control base model is jointly constructed according to different category feature data, preset business target and the fusion target, so that the risk control base model can adapt to new tasks of various categories and improve the robustness of the risk control base model. Further, the target label of the new task is obtained, and the target label of the new task is input into each risk control base model to evaluate each risk control base model to obtain each model to be fused corresponding to the preset screening constraint condition, and then based on the new task The target label of the service is obtained, and the models to be fused are fused, and the fused models are evaluated to select the target credit risk control model corresponding to the preset fusion constraint conditions, so as to realize the evaluation and screening of the existing risk control base models according to the target of the new task, and the number of risk control base models is significantly lower than the number of features, which can greatly reduce the computing time cost, and then the screened risk control base models are combined to complete the adaptation to the new task. Furthermore, when some important features change and migrate, only some risk control base models need to be retrained and the weights of related fusion models need to be adjusted, thereby reducing the search space of parameters in the screening process and reducing the consumption of computing resources. The problem in the prior art that when encountering a new task, in the process of re-screening features, it is necessary to evaluate and screen the features in all feature layers according to the target of the new task, and when some underlying features migrate or change drastically, the model needs to be readjusted, which will also consume a lot of computing resources and time costs, thereby leading to the technical defect of low efficiency of model adaptive construction, thereby improving the efficiency of model adaptive construction.

[0068] Further, refer to Figure 2 Based on the first embodiment of the present application, in another embodiment of the present application, the step of constructing a risk control base model corresponding to each of the category feature data based on each of the category feature data, the preset business goal and the fusion goal includes:

[0069] Step A10, obtaining the model to be trained corresponding to each of the category feature data;

[0070] In this embodiment, it should be noted that the models to be trained corresponding to the category feature data are not the same. The models to be trained include LR logistic regression model, gradient boosting decision tree model (GBDT), neural network model, support vector machine SVM, random forest RF and other models to be trained.

[0071] Step A20, associate and combine each of the category feature data with the preset business goal and the fusion goal respectively, and according to the combination result, iteratively train and optimize the to-be-trained models corresponding to each of the category feature data respectively to obtain the risk control base model corresponding to each of the category feature data.

[0072] In this embodiment, each of the category feature data is associated and combined with the preset business goal and the fusion goal, so that according to the combination result, the to-be-trained model corresponding to each of the category feature data is iteratively trained and optimized to obtain the risk control base model corresponding to each of the category feature data. Specifically, the following steps are performed for each of the category feature data:

[0073] The category feature data is associated with the preset business objectives and fusion objectives, and then based on the category feature data and the preset business objectives and fusion objectives associated with the category feature data, the model to be trained is iteratively trained to optimize the model parameters corresponding to the model to be trained, and it is determined whether the optimized model to be trained meets the preset training end conditions, wherein the preset training end conditions include conditions such as loss function convergence and reaching the maximum number of iterations threshold. If satisfied, the risk control base model corresponding to the category feature data is obtained; if not satisfied, the execution step is returned to: according to the combination result, the model to be trained corresponding to each of the category feature data is iteratively trained and optimized to obtain the risk control base model corresponding to each of the category feature data, thereby realizing the construction of a risk control base model based on a combination of different categories of special data and different business objectives to adapt to new tasks of various categories and improve the robustness of the model.

[0074] The step of iteratively training and optimizing the to-be-trained model corresponding to each of the category feature data according to the combination result to obtain the risk control base model corresponding to each of the category feature data includes:

[0075] Step A21, inputting each category feature data in the combination result into the corresponding to-be-trained model, and outputting the test result corresponding to each category feature data;

[0076] In this embodiment, each category feature data in the combination result is input into the corresponding to-be-trained model, and the test result corresponding to each category feature data is output. Specifically, the following steps are performed for each category feature data:

[0077] After the category feature data and the preset business goal are associated and combined, the category feature data is input into the corresponding model to be trained, and the test result corresponding to the category feature data is output.

[0078] Step A22, calculating the model loss corresponding to each of the to-be-trained models based on the test results, the preset business objectives and the fusion objectives;

[0079] In this embodiment, based on each of the test results, the preset business objectives and the fusion objectives, the model loss corresponding to each of the models to be trained is calculated. Specifically, the following steps are performed for each of the category feature data:

[0080] Based on the test results corresponding to the category feature data, the preset business goals associated with the category feature data, and the difference between the fusion targets, the model loss corresponding to the model to be trained is calculated by a preset loss function, for example, the model loss is calculated by an L2 loss function.

[0081] Step A23, based on the gradient of the loss calculation of each model, iterative training optimization is performed on each of the models to be trained to obtain the risk control base model corresponding to each of the category feature data.

[0082] In this embodiment, based on the gradient calculated by the loss of each model, each model to be trained is iteratively trained and optimized to obtain the risk control base model corresponding to each category feature data. Specifically, based on the loss of each model, the gradient corresponding to each model to be trained is calculated, and then each model to be trained is iteratively trained and optimized to optimize the model to be trained, and it is determined whether the optimized model to be trained meets the preset training end condition. If so, the model to be trained is used as the risk control base model corresponding to the type feature data.

[0083] The embodiment of the present application provides a method for adaptively constructing a risk control model, that is, obtaining a model to be trained corresponding to each of the category feature data, and then associating and combining each of the category feature data with the preset business goal and the fusion goal, respectively, so as to iteratively train and optimize the model to be trained corresponding to each of the category feature data according to the combination result, and obtain a risk control base model corresponding to each of the category feature data, thereby realizing the construction of a corresponding risk control base model according to different category feature data and preset business goals and fusion goals, so that the risk control base model can adapt to new tasks of various categories, thereby improving the robustness of the model. Furthermore, when some important features change and migrate, only part of the risk control base model needs to be retrained, and there is no need to readjust the entire model, thereby reducing the consumption of computing resources.

[0084] Further, refer to Figure 3 Based on the first embodiment of the present application, in another embodiment of the present application, after the step of fusing the models to be fused based on the target label of the new task and evaluating the fused model to select the target credit risk control model corresponding to the preset fusion constraint condition, the risk control model adaptive construction method further includes:

[0085] Step B10, obtaining characteristic information corresponding to the customer to be predicted;

[0086] In this embodiment, it should be noted that the characteristic information includes internal characteristic information of the loan associated with the customer to be predicted, customer identity characteristic information, web browsing characteristic information, and characteristic information corresponding to the interaction between the customer and the APP application software.

[0087] Step B20: Based on the characteristic information and the target credit risk control model, risk prediction is performed on the customer to be predicted to obtain a risk prediction result.

[0088] In this embodiment, based on the characteristic information and the target credit risk control model, risk prediction is performed on the customer to be predicted to obtain a risk prediction result. Specifically, the characteristic information is input into the target credit risk control model, risk prediction is performed on the customer to be predicted, and the risk prediction result corresponding to the customer to be predicted is output. For example, based on the characteristic information x associated with the customer to be predicted, the target credit risk control model outputs the default probability prediction y corresponding to the customer to be predicted, and y has a good correlation with the customer's actual credit performance in the future. Since the target credit risk control model is adaptively constructed according to the new task, the target credit risk control model can adapt to various categories of new tasks, thereby improving the accuracy of model prediction.

[0089] Reference Figure 4 , Figure 4 It is a schematic diagram of the structure of the device for adaptively building a risk control model for the hardware operating environment involved in the embodiment of the present application.

[0090] like Figure 4 As shown, the risk control model adaptive construction device may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between the processor 1001 and the memory 1005. The memory 1005 can be a high-speed RAM memory or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0091] Optionally, the risk control model adaptive construction device may also include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, etc. The rectangular user interface may include a display screen (Display), an input submodule such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface. The network interface may optionally include a standard wired interface and a wireless interface (such as a WIFI interface).

[0092] Those skilled in the art will understand that Figure 4 The structure of the risk control model adaptive construction device shown in the figure does not constitute a limitation of the risk control model adaptive construction device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0093] like Figure 4 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, and a risk control model adaptive construction program. The operating system is a program that manages and controls the hardware and software resources of the risk control model adaptive construction device, and supports the operation of the risk control model adaptive construction program and other software and / or programs. The network communication module is used to realize the communication between the components inside the memory 1005, and the communication with other hardware and software in the risk control model adaptive construction system.

[0094] exist Figure 4 In the risk control model adaptive construction device shown, the processor 1001 is used to execute the risk control model adaptive construction program stored in the memory 1005 to implement the steps of the risk control model adaptive construction method described in any of the above items.

[0095] The specific implementation method of the risk control model adaptive construction device of the present application is basically the same as the various embodiments of the above-mentioned risk control model adaptive construction method, and will not be repeated here.

[0096] The present application also provides a risk control model adaptive construction device, the risk control model adaptive construction device comprising:

[0097] A data storage module, used to obtain multi-dimensional feature data of customers, and classify and store the multi-dimensional feature data to obtain feature data of each category;

[0098] A fusion module is used to create a supervision target corresponding to the customer, and fuse and optimize the preset business target and the supervision target to obtain a fusion target;

[0099] A model building module, used to build a risk control base model corresponding to each of the category feature data based on the category feature data, the preset business goal and the fusion goal;

[0100] A screening module is used to obtain the target label of the new task and input the target label of the new task into each of the risk control base models to evaluate each of the risk control base models and obtain each model to be fused that satisfies the preset screening constraint conditions;

[0101] The model combination module is used to fuse the models to be fused based on the target label of the new task, and evaluate the fused models to select the target credit risk control model that meets the preset fusion constraint conditions.

[0102] Optionally, the fusion module is further used for:

[0103] Iteratively optimizing the preset business objectives and the supervision objectives by a preset optimization method to determine the weight coefficient corresponding to the preset business objectives and obtain the optimization objectives corresponding to the weight coefficients;

[0104] If the square of the Pearson correlation coefficient between the optimization target and the supervision target meets the preset coefficient threshold condition, the optimization target is discretized to obtain the fusion target.

[0105] Optionally, the model building module is further used to:

[0106] Obtaining the model to be trained corresponding to each of the category feature data;

[0107] Each of the category feature data is associated and combined with the preset business goal and the fusion goal respectively, so as to iteratively train and optimize the to-be-trained models corresponding to each of the category feature data according to the combination results, so as to obtain the risk control base model corresponding to each of the category feature data.

[0108] Optionally, the model building module is further used to:

[0109] Inputting the category feature data into the corresponding models to be trained respectively, and outputting the test results corresponding to the category feature data;

[0110] Based on the test results, the preset business objectives and the fusion objectives, the model loss corresponding to each of the models to be trained is calculated;

[0111] Based on the gradient of the loss calculation of each model, each model to be trained is iteratively trained and optimized to obtain the risk control base model corresponding to each category feature data.

[0112] Optionally, the screening module is further used for:

[0113] Based on the target label of the new task, each of the risk control base models is evaluated to obtain the evaluation indicators corresponding to each of the risk control base models;

[0114] Based on the preset screening constraints and the evaluation indicators corresponding to the risk control base models, the risk control base models are screened to obtain the models to be integrated.

[0115] Optionally, the model combination module is further used for:

[0116] Based on the target label of the new task, weighted fusion is performed on the models to be fused to obtain fusion models corresponding to different weights;

[0117] Calculate the model evaluation indicators corresponding to the fusion models with different weights respectively;

[0118] Based on the evaluation indicators of each model, a target credit risk control model that meets the preset fusion constraints is selected.

[0119] Optionally, the data storage module is further used for:

[0120] The multi-dimensional feature data are classified and stored according to the data source and the characterization dimension of the customer characteristics to obtain the feature data of each category.

[0121] Optionally, the risk control model adaptive construction device is also used for:

[0122] Obtain feature information corresponding to the customer to be predicted;

[0123] Based on the characteristic information and the target credit risk control model, risk prediction is performed on the customer to be predicted to obtain a risk prediction result.

[0124] The specific implementation of the risk control model adaptive construction device of the present application is basically the same as the various embodiments of the above-mentioned risk control model adaptive construction method, and will not be repeated here.

[0125] An embodiment of the present application provides a storage medium, which is a computer-readable storage medium, and the computer-readable storage medium stores one or more programs, and the one or more programs can also be executed by one or more processors to implement the steps of the risk control model adaptive construction method described in any one of the above items.

[0126] The specific implementation methods of the computer-readable storage medium of the present application are basically the same as the various embodiments of the above-mentioned risk control model adaptive construction method, and will not be repeated here.

[0127] An embodiment of the present application provides a computer program product, and the computer program product includes one or more computer programs, and the one or more computer programs can also be executed by one or more processors to implement the steps of any of the above-mentioned methods for adaptively constructing a risk control model.

[0128] The specific implementation methods of the computer program product of the present application are basically the same as the various embodiments of the above-mentioned risk control model adaptive construction method, and will not be repeated here.

[0129] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent processing scope of the present application.

Claims

1. A method for adaptively constructing a risk control model, characterized in that: The method for adaptively constructing a risk control model includes: Acquire multi-dimensional feature data of the customer, and classify and store the multi-dimensional feature data to obtain feature data of each category; Create supervision targets corresponding to customers, and integrate and optimize preset business targets and the supervision targets to obtain integrated targets; Based on each of the category feature data, the preset business goal and the fusion goal, construct a risk control base model corresponding to each of the category feature data; Obtaining the target label of the new task, and inputting the target label of the new task into each of the risk control base models to evaluate each of the risk control base models, and obtaining each to-be-fused model corresponding to the preset screening constraint condition; Based on the target label of the new task, the models to be integrated are integrated, and the integrated models are evaluated to select the target credit risk control model that meets the preset integration constraints, wherein the integration constraints include the number of integrated models, the discrimination KS index, and the stability PSI index. The step of fusing and optimizing the preset business goal and the supervision goal to obtain the fusion goal includes: Iteratively optimizing the preset business objectives and the supervision objectives by a preset optimization method to determine the weight coefficient corresponding to the preset business objectives and obtain the optimization objectives corresponding to the weight coefficients; If the square of the Pearson correlation coefficient between the optimization target and the supervision target meets the preset coefficient threshold condition, the optimization target is discretized to obtain the fusion target, wherein the supervision target is the customer's specific business information, including the customer's overdue amount and overdue days, and the preset business target is a multi-angle business target summarized based on historical business experience, including the customer's short-term overdue situation, and the fusion target is a new optimization target obtained by combining the preset business target and the supervision target through a preset optimization method.

2. The method for adaptively constructing a risk control model according to claim 1, characterized in that: The step of constructing a risk control base model corresponding to each of the category feature data based on each of the category feature data, the preset business goal and the fusion goal includes: Obtaining the model to be trained corresponding to each of the category feature data; Each of the category feature data is associated and combined with the preset business goal and the fusion goal respectively, so as to iteratively train and optimize the to-be-trained models corresponding to each of the category feature data according to the combination results, so as to obtain the risk control base model corresponding to each of the category feature data.

3. The method for adaptively constructing a risk control model according to claim 2, characterized in that: The step of performing iterative training optimization on the to-be-trained models corresponding to the category feature data according to the combination result to obtain the risk control base model corresponding to the category feature data comprises: Inputting each category feature data in the combination result into the corresponding to-be-trained model respectively, and outputting the test result corresponding to each category feature data; Based on each of the test results, the preset business goal and the fusion goal, calculating the model loss corresponding to each of the models to be trained; Based on the gradient of the loss calculation of each model, each model to be trained is iteratively trained and optimized to obtain the risk control base model corresponding to each category feature data.

4. The method for adaptively constructing a risk control model according to claim 1, characterized in that: The step of inputting the target label of the new task into each of the risk control base models to evaluate each of the risk control base models and obtaining each to-be-fused model corresponding to the preset screening constraint condition comprises: Based on the target label of the new task, each of the risk control base models is evaluated to obtain the evaluation indicators corresponding to each of the risk control base models; Based on the preset screening constraints and the evaluation indicators corresponding to the risk control base models, the risk control base models are screened to obtain the models to be integrated.

5. The method for adaptively constructing a risk control model according to claim 1, characterized in that: The step of fusing the models to be fused based on the target label of the new task and evaluating the fused model to select the target credit risk control model that satisfies the preset fusion constraint condition comprises: Based on the target label of the new task, weighted fusion is performed on the models to be fused to obtain fusion models corresponding to different weights; Calculate the model evaluation indicators corresponding to the fusion models with different weights respectively; Based on the evaluation indicators of each model, a target credit risk control model that meets the preset fusion constraints is selected.

6. The method for adaptively constructing a risk control model according to claim 1, characterized in that: The step of classifying and storing the multi-dimensional feature data to obtain feature data of each category includes: The multi-dimensional feature data are classified and stored according to the data source and the characterization dimension of the customer characteristics to obtain the feature data of each category.

7. The method for adaptively constructing a risk control model according to any one of claims 1 to 6, characterized in that: After the step of fusing the models to be fused based on the target label of the new task and evaluating the fused model to select a target credit risk control model that satisfies the preset fusion constraint condition, the risk control model adaptive construction method further includes: Obtain feature information corresponding to the customer to be predicted; Based on the characteristic information and the target credit risk control model, risk prediction is performed on the customer to be predicted to obtain a risk prediction result.

8. A risk control model adaptive construction device, characterized in that: The risk control model adaptive construction device includes: a memory, a processor, and a risk control model adaptive construction program stored in the memory, The risk control model adaptive construction program is executed by the processor to implement the steps of the risk control model adaptive construction method as described in any one of claims 1 to 7.

9. A storage medium, the storage medium being a computer-readable storage medium, characterized in that: The computer-readable storage medium stores a risk control model adaptive construction program, and the risk control model adaptive construction program is executed by a processor to implement the steps of the risk control model adaptive construction method as described in any one of claims 1 to 7.

Images