An automated model framework and risk management method based on quasi-real-time effects

Abstract

The present invention provides an automated model framework and risk control method based on quasi-real-time effects, which belongs to the field of risk management technology. Specifically, the framework includes: a variable pool module responsible for extracting credit features of credit users' credit application information to obtain a variable pool; a quasi-real-time model module responsible for performing quasi-real-time update processing of samples based on the variable pool according to a preset period, and performing incremental training of the basic model based on samples after adversarial processing based on the basic model, and performing update processing of the credit model. When the update is completed, the method is put online; a data monitoring module is responsible for adding timestamp features to the credit model, and using timestamp features to analyze the changing trends of the credit model and customer groups, thereby improving the timeliness of the update processing of the credit model.

Description

Technical Field

[0001] The present invention belongs to the field of risk management technology, and in particular relates to an automated model framework and risk management method based on quasi-real-time effects. Background Art

[0002] In the field of credit risk control, user risk often has a lag. To address the above technical issues, existing technical solutions often determine credit risk based on the user's changing credit characteristics and the original risk model. Specifically, a similar technical solution is proposed in invention patent application CN202311379683.1 "A Multi-Objective Risk Control Strategy Optimization Method and System". However, the above technical solution has the following technical problems:

[0003] Since the risk control model is based on the risk labels of historical users, when the profile of the newly added customer group changes, the model cannot capture user risks in a timely and effective manner. At the same time, in the same model application scenario, it is necessary to consider multiple risk label modeling or evaluate the effects of multiple data features. Therefore, how to automatically build models in real time to solve the technical knowledge and background required for the practical application of machine learning methods, as well as time-consuming and repetitive steps, is a relatively important topic.

[0004] To solve the above technical problems, the present application provides an automated model framework and risk management method based on quasi-real-time effects. Summary of the Invention

[0005] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0006] Specifically, in a first aspect, the present application provides an automation model framework based on quasi-real-time effects, specifically including:

[0007] Variable pool module, quasi-real-time model module, data monitoring module;

[0008] The variable pool module is responsible for extracting the credit characteristics of the credit application information of the credit user to obtain the variable pool;

[0009] The quasi-real-time model module is responsible for performing quasi-real-time update processing of samples based on the variable pool according to a preset period, and performing incremental training of the basic model based on the basic model using samples after adversarial processing, and updating the credit model. After the update is completed, the model is put online.

[0010] The data monitoring module is responsible for adding timestamp features to the credit model and using the timestamp features to analyze the credit model and the changing trends of the customer base.

[0011] A further technical solution is that the basic model is constructed by building an AutoML framework.

[0012] In a second aspect, the present application provides a risk management method, which is applied to the aforementioned automated model framework based on quasi-real-time effects, and specifically includes:

[0013] S1 divides the credit users into different customer groups based on their credit application information, and determines adversarial processing customer groups within the customer groups based on the distribution of the customer groups in the training data of the credit model and the distribution data of credit users on different dates;

[0014] S2 uses the adversarial processing customer group of the current date as the matching processing customer group, and determines that the training processing risk of the credit model does not meet the requirements based on the changes in the amplified data of different matching processing customer groups, and then proceeds to the next step;

[0015] S3 determines the target influencing customer group and its influence correlation coefficient with the different matching customer groups based on the changes in the credit user data of different customer groups, the credit user data of different matching customer groups, and the changes in the augmented data;

[0016] S4 determines the verification processing strategy for different target influence customer groups during the training of different matching processing customer groups based on the influence correlation coefficient between the target influence customer group and different matching processing customer groups and the data amplification status on the current date, determines the training processing sequence for different matching processing customer groups based on the verification processing strategy, and performs credit risk management and control based on the updated credit model.

[0017] A further technical solution is to divide the credit users into different customer groups, specifically including:

[0018] Determining the number of identical information items in the credit application information of different credit users based on similarities in the credit application information of the credit users;

[0019] The credit users are divided into different customer groups according to the number of identical information items.

[0020] A further technical solution is to classify credit users whose number of identical information items is greater than a preset threshold value of the number of identical information items into the same customer group.

[0021] A further technical solution is that the distribution of the customer group in the training data of the credit model includes the number of credit users of the customer group in the credit model and the proportion of credit users in the total number of credit users in the credit model.

[0022] A further technical solution is that the method for determining the confrontation processing customer group in the customer group is:

[0023] Based on the distribution of the customer group in the training data of the credit model, determine the proportion of credit users of the customer group in the historical training data of the credit model, and use this proportion as the proportion of the number of credit users of the customer group;

[0024] Determine the date on which the credit users in the customer group exist based on the distribution data of the credit users on different dates, and use the date as the matching credit date;

[0025] Based on the ratio of the number of matching credit dates to the number of users in the customer group, the adversarial processing demand coefficient of the customer group is determined, and based on the adversarial processing demand coefficient, it is determined whether the customer group belongs to the adversarial processing customer group.

[0026] A further technical solution is that when the confrontation processing requirement coefficient is greater than a preset processing requirement coefficient threshold, it is determined that the customer group belongs to the confrontation processing customer group.

[0027] A further technical solution is that the method for determining the training processing order of the matching processing customer group is:

[0028] Using verification strategies for different matching customer groups, determine the target influencing customer groups that need verification and processing for different matching customer groups, and use them as verification influencing customer groups;

[0029] According to the verification processing strategies of different verification-affecting customer groups in different matching processing customer groups, the proportion of matching processing customer groups that have been verified in the matching processing customer groups is determined, and the proportion is used as the verification demand coefficient;

[0030] Based on the sum of the verification requirement coefficients of different verification influencing customer groups of different matching processing customer groups, the verification requirement values ​​of different matching processing customer groups are determined, and the training processing order of the different matching processing customer groups is determined based on the verification requirement values.

[0031] A further technical solution is to determine the training processing sequence for different matching processing customer groups based on the verification requirement value, specifically including:

[0032] The training processing order of different matching processing customer groups is determined from large to small according to the verification requirement value, and after all matching processing customer groups are trained according to the training processing order, the credit model training processing is performed based on all customer groups except the matching processing customer groups.

[0033] A further technical solution is to manage and control credit risk based on the updated credit model, specifically including:

[0034] Based on the updated credit model and in combination with the credit application information of different credit users as input, the credit risk of the credit user is determined based on the output of the updated credit model.

[0035] The beneficial effects of the present invention are:

[0036] Based on the influence correlation coefficient between the target influence customer group and different matching processing customer groups and the data amplification situation on the current date, the verification processing strategy for different target influence customer groups during the training processing of different matching processing customer groups is determined. This not only ensures the reliability of the verification processing of the target influence customer groups that are highly influenced by the matching processing customer groups, but also avoids the technical problem that the verification processing efficiency cannot meet the requirements when verification processing is performed on different matching processing customer groups during training processing.

[0037] Based on the verification processing strategy, the training processing order of different matching processing customer groups is determined, and the number of target influence customer groups that need to be verified and processed during the training processing of different matching processing customer groups and the verification association between the target influence customer groups and other matching processing customer groups are fully taken into consideration. This realizes the screening of matching processing customer groups with a large number of target influence customer groups that need to be verified and processed and a large number of verification associations between the target influence customer groups and other matching processing customer groups, and performs verification processing in a timely manner, thereby improving the efficiency and reliability of positioning processing when there are abnormalities in the verification processing results.

[0038] Other features and advantages will be described in the following description. The objectives and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and drawings.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the accompanying drawings.

[0041] Figure 1 It is a framework diagram of an automation model framework based on quasi-real-time effects;

[0042] Figure 2 It is a flowchart of a multi-source data verification and reach method for uninsured persons;

[0043] Figure 3 It is a flow chart of the method for determining the uninsured persons who are difficult to visit;

[0044] Figure 4It is a flow chart to determine whether the difficulty of visiting and handling the target area meets the requirements;

[0045] Figure 5 It is a flowchart for determining the optimization processing of the reach processing strategy that needs to be performed. DETAILED DESCRIPTION

[0046] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this specification without creative work should fall within the scope of protection of this specification.

[0047] In the field of credit risk control, user risk often has a lag. Since the risk control model is based on the risk labels of historical users, when the profile of the newly added customer group changes, the model cannot capture the user risk in a timely and effective manner. At the same time, in the same model application scenario, it is necessary to consider multiple risk label modeling or to evaluate the effects of multiple data features. Therefore, how to automatically build a model in real time to solve the technical knowledge and background required for the practical application of machine learning methods, as well as time-consuming and repetitive steps, is a relatively important topic. The present invention proposes a model update method based on an automated modeling framework with quasi-real-time effects, which can quickly respond to the impact of customer group fluctuations on the online model and ensure the stability of the model effect.

[0048] The user data and other data involved in this application are obtained with full consent and authorization, and the collection, use and processing of relevant information comply with the relevant laws, regulations and standards of relevant countries and regions.

[0049] 1) Implementation plan:

[0050] 1. Variable pool module: Establish a candidate feature pool as a variable screening benchmark. On this basis, build an indicator screening system including stability, discrimination, and versatility, and ultimately form a standardized variable pool for automatic modeling.

[0051] 2. Quasi-real-time model module:

[0052] a. Automatic sample update: A sliding window + incremental data update mode is used to achieve quasi-real-time sample updates;

[0053] b. Automatic model update: First, the AutoML framework is built to train the base model. Then, real-time sample information is introduced through adversarial weighting to perform incremental training on the base model, thus achieving near-real-time model updates.

[0054] c. Automatic model online: Automatically convert offline model files into online format files for output, thus automating the modeling process;

[0055] d. Time-series data monitoring: Adding timestamps to the tree model's features is crucial when traffic changes. By parsing the tree structure, we can detect how and when time affects data. Furthermore, we can create splits based on timestamps and analyze the differences in traffic before and after the split.

[0056] Example 1

[0057] like Figure 1 As shown, the present application provides an automation model framework based on quasi-real-time effects, specifically including:

[0058] Variable pool module, quasi-real-time model module, data monitoring module;

[0059] The variable pool module is responsible for extracting the credit characteristics of the credit application information of the credit user to obtain the variable pool;

[0060] The quasi-real-time model module is responsible for performing quasi-real-time update processing of samples based on the variable pool according to a preset period, and performing incremental training of the basic model based on the basic model using samples after adversarial processing, and updating the credit model. After the update is completed, the model is put online.

[0061] The data monitoring module is responsible for adding timestamp features to the credit model and using the timestamp features to analyze the credit model and the changing trends of the customer base.

[0062] Furthermore, the basic model is constructed by building an AutoML framework.

[0063] Example 2

[0064] Second, as Figure 2 As shown, the present application provides a risk management method, which is applied to the above-mentioned automated model framework based on quasi-real-time effects, specifically including:

[0065] S1 divides the credit users into different customer groups based on their credit application information, and determines adversarial processing customer groups within the customer groups based on the distribution of the customer groups in the training data of the credit model and the distribution data of credit users on different dates;

[0066] The adversarial processing demand coefficient of the customer group is determined by the ratio of the proportion of the number of dates of credit users existing in the customer group to the proportion of the number of credit users of the customer group in the historical training data of the credit model. When the adversarial processing demand coefficient is greater than the preset processing demand coefficient threshold, it is determined that the customer group belongs to the adversarial processing customer group.

[0067] S2 uses the adversarial processing customer group of the current date as the matching processing customer group, and determines that the training processing risk of the credit model does not meet the requirements based on the changes in the amplified data of different matching processing customer groups, and then proceeds to the next step;

[0068] When there is a matching processing customer group whose data volume of augmented data is greater than a preset data volume threshold, it is determined that the training processing risk of the credit model does not meet the requirements.

[0069] S3 determines the target influencing customer group and its influence correlation coefficient with the different matching customer groups based on the changes in the credit user data of different customer groups, the credit user data of different matching customer groups, and the changes in the augmented data;

[0070] Determine the number of credit users in a customer group based on the changes in credit user data for different customer groups;

[0071] The changes in the credit user data and expansion data of different matching processing customer groups are used to determine the changed number of credit users and the number of expanded users of different matching processing customer groups. Based on the average value of the ratio of the changed number of credit users and the number of expanded users of different matching processing customer groups to the changed number of credit users of the customer groups, the influence correlation coefficients of the different matching processing customer groups are determined. When the sum of the influence correlation coefficients of the different matching processing customer groups is greater than the preset correlation coefficient threshold, it is determined that the customer group belongs to the target influence customer group.

[0072] Table 1 shows the change data of credit users in different customer groups on different dates

[0073]

[0074] S4 determines the verification processing strategy for different target influence customer groups during the training of different matching processing customer groups based on the influence correlation coefficient between the target influence customer group and different matching processing customer groups and the data amplification status on the current date, determines the training processing sequence for different matching processing customer groups based on the verification processing strategy, and performs credit risk management and control based on the updated credit model.

[0075] The target influenced customer group is verified in the matching processing customer group whose influence correlation coefficient is greater than the preset influence correlation coefficient threshold or the matching processing customer group whose number of expanded credit users is greater than the preset expansion number threshold.

[0076] The training processing order of the matching processing customer groups is determined from large to small based on the sum of the influence correlation coefficients with different target influencing customer groups.

[0077] Furthermore, the credit users are divided into different customer groups, including:

[0078] Determining the number of identical information items in the credit application information of different credit users based on similarities in the credit application information of the credit users;

[0079] The credit users are divided into different customer groups according to the number of identical information items.

[0080] Specifically, credit users whose number of identical information items is greater than a preset threshold value of the number of identical information items are classified into the same customer group.

[0081] It should be noted that the distribution of the customer group in the training data of the credit model includes the number of credit users of the customer group in the credit model and the proportion of credit users in the total number of credit users in the credit model.

[0082] It is understandable that if Figure 3 As shown, the method for determining the adversarial processing customer group in the customer group is:

[0083] Based on the distribution of the customer group in the training data of the credit model, determine the proportion of credit users of the customer group in the historical training data of the credit model, and use this proportion as the proportion of the number of credit users of the customer group;

[0084] Determine the date on which the credit users in the customer group exist based on the distribution data of the credit users on different dates, and use the date as the matching credit date;

[0085] Based on the ratio of the number of matching credit dates to the number of users in the customer group, the adversarial processing demand coefficient of the customer group is determined, and based on the adversarial processing demand coefficient, it is determined whether the customer group belongs to the adversarial processing customer group.

[0086] Furthermore, when the confrontation processing requirement coefficient is greater than a preset processing requirement coefficient threshold, it is determined that the customer group belongs to the confrontation processing customer group.

[0087] In another possible embodiment, the method for determining the adversarial processing customer group in the customer group is:

[0088] Based on the distribution of the customer group in the training data of the credit model, determine the number of credit users of the customer group in the historical training data of the credit model, and use it as the number of matching users;

[0089] Determine the date on which the credit users in the customer group exist based on the distribution data of the credit users on different dates, and use the date as the matching credit date;

[0090] Based on the matching credit date and the number of matching users, it is determined whether the customer group belongs to the adversarial processing customer group.

[0091] Furthermore, when the number of matched credit dates of the customer group is greater than a preset credit date number threshold and the number of matched users is less than a preset matching user number threshold, it is determined that the customer group belongs to the adversarial processing customer group.

[0092] Furthermore, the expanded data is determined based on the expanded credit users after the matching processing customer group is subjected to counter-processing based on the credit data of the original credit users.

[0093] Specifically, such as Figure 4 As shown, it is determined that the training processing risk of the credit model does not meet the requirements, including:

[0094] Determine the number of credit users expanded by different matching customer groups based on changes in the expansion data of different matching customer groups, and use this as the number of expanded users;

[0095] According to the number of expanded users of different matching processing customer groups, a matching processing customer group with an expanded user number greater than a preset expansion number threshold is determined, and the expanded customer group is used as the expansion processing customer group;

[0096] Based on the number of the expanded customer base, it is determined whether the training processing risk of the credit model meets the requirements.

[0097] Furthermore, when the number of the expanded customer group is greater than a preset expanded customer group number threshold, it is determined that the training processing risk of the credit model does not meet the requirements.

[0098] It is understandable that when the training processing risk of the credit model meets the requirements, the training processing of the credit model is carried out directly after the expansion processing of all matching processing customer groups, and there is no need to verify the credit model after the training processing is completed.

[0099] In another possible embodiment, determining that the training processing risk of the credit model does not meet the requirements specifically includes:

[0100] Determine the number of credit users expanded by different matching customer groups based on changes in the expansion data of different matching customer groups, and use this as the number of expanded users;

[0101] According to the number of expanded users of different matching processing customer groups, the sum of the number of expanded users of different matching processing customer groups is calculated and used as the total number of expansions;

[0102] Based on the total amount of expansion, it is determined whether the training processing risk of the credit model meets the requirements.

[0103] Furthermore, when the total number of expansions is greater than a preset value of the expansion number, it is determined that the training processing risk of the credit model does not meet the requirements.

[0104] In another possible embodiment, determining that the training processing risk of the credit model does not meet the requirements specifically includes:

[0105] S41 determines the number of credit users expanded by the different matching customer groups based on the changes in the expansion data of the different matching customer groups, and uses the number as the number of expanded users. Based on the number of expanded users of the different matching customer groups, the number of expanded users of the different matching customer groups is summed, and the sum is used as the total number of expanded users.

[0106] Optionally, before entering step S42, it is also necessary to determine whether the total number of expansions meets the requirements. When the total number of expansions does not meet the requirements, it is determined that the training processing risk of the credit model does not meet the requirements. If and only if the total number of expansions meets the requirements, proceed to step S42.

[0107] If the total number of expansions does not meet the requirements, it means that the number of customers undergoing expansion processing is greater than a certain threshold, which will inevitably lead to deviations in the training data of the credit model. Therefore, it can be determined without a doubt that the training processing risk of the credit model at this time does not meet the requirements.

[0108] S42 determines the expansion impact value of each customer group based on the number of credit users of each customer group and the ratio of the total expansion number to the number of credit users of each customer group;

[0109] Optionally, before entering step S43, it is also necessary to determine whether there is a customer group with a relatively large amplification influence value. When there is no customer group with an amplification influence value greater than a preset threshold, it can be directly determined that the training processing risk of the credit model meets the requirements.

[0110] Optionally, before entering step S43, it is also necessary to obtain the number of customer groups whose amplification influence values ​​do not meet the requirements. When the number of customer groups whose amplification influence values ​​do not meet the requirements is greater than a preset amplification influence customer group number threshold, it is determined that the training processing risk of the credit model does not meet the requirements.

[0111] If and only if the number of customer groups whose amplification influence values ​​do not meet the requirements is not greater than the preset amplification influence customer group number threshold, then proceed to step S43.

[0112] S43 determines the matching processing customer groups whose number of expanded users is greater than a preset expansion number threshold according to the number of expanded users in different matching processing customer groups, and uses it as the expansion processing customer group. According to the ratio of the number of expanded users in different expansion processing customer groups to the number of credit users in different customer groups, different expansion processing customer groups are determined, and the training processing risk coefficient of the credit model is determined in combination with the expansion influence value of different customer groups. Based on the training processing risk coefficient, it is determined whether the training processing risk of the credit model meets the requirements.

[0113] Optionally, the above step S43 includes the following contents:

[0114] Before evaluating the amplification impact coefficient, it is also necessary to determine whether the number of the amplified customer group meets the requirements. If so, it is determined that the training processing risk of the credit model does not meet the requirements. If not, the amplification impact coefficient is evaluated again.

[0115] Furthermore, when the training process risk coefficient is greater than a preset training risk coefficient threshold, it is determined that the training process risk of the credit model does not meet the requirements.

[0116] Specifically, such as Figure 5 As shown, the method for determining the target influencing customer group in the customer group is:

[0117] Determining the number of credit users in a customer group based on changes in credit user data for different customer groups;

[0118] Determine the number of credit users changed and the number of expanded users for different matching customer groups based on the changes in the credit user data and expansion data for different matching customer groups;

[0119] Based on the average value of the change in the number of credit users of different matching processing customer groups and the ratio of the number of expanded users to the change in the number of credit users of the customer groups, the influence correlation coefficients of different matching processing customer groups are determined, and based on the influence correlation coefficients of different matching processing customer groups, it is determined whether the customer group belongs to the target influence customer group.

[0120] Furthermore, the changed number of credit users in the customer group is determined based on the number of credit users in the customer group during the initial training process of the credit model and the current number of credit users in the customer group.

[0121] It is understandable that determining whether a customer group belongs to the target influenced customer group according to the influence correlation coefficient of different matching processing customer groups specifically includes:

[0122] When there is a matching customer group whose influence correlation coefficient is greater than a preset influence correlation coefficient threshold, the customer group is determined to be a target influence customer group.

[0123] In another possible embodiment, the method for determining the target influencing customer group in the customer group is:

[0124] S31 determines the changed number of credit users in different customer groups based on changes in the credit user data of the customer groups, determines the changed number of credit users and the number of expanded users in different matching customer groups based on changes in the credit user data and expanded data of different matching customer groups, and determines the impact correlation coefficients of different matching customer groups based on the ratios of the changed number of credit users and the number of expanded users in different matching customer groups to the changed number of credit users in the customer groups;

[0125] S32 determines the change impact coefficients of different matching customer groups based on the changed number of credit users and the number of expanded users of the different matching customer groups, and in combination with the current number of credit users of the customer groups;

[0126] S33 determines the comprehensive influence value of the customer group according to the influence correlation coefficient and the correlation influence coefficient of different matching processing customer groups, and determines whether the customer group belongs to the target influence customer group based on the comprehensive influence value.

[0127] Furthermore, when the comprehensive influence value of the customer group is greater than a preset comprehensive influence threshold, the customer group is determined to be a target influence customer group.

[0128] Optionally, the above step S31 includes the following contents:

[0129] S311 determines the number of credit users in a customer group based on the change in credit user data of the customer group. When the number of credit users in the customer group is less than a preset change threshold, the process proceeds to step S312. When the number of credit users in the customer group is not less than the preset change threshold, the process proceeds to step S313.

[0130] S312: When the current number of credit users in the customer group is less than the preset credit user number threshold, it is determined that the customer group belongs to the target influence customer group. When the current number of credit users in the customer group is not less than the preset credit user number threshold, the process proceeds to step S313;

[0131] S313 determines the changed number of credit users and the number of expanded users for each of the different matching customer groups based on the changes in the credit user data and the expanded data for each of the different matching customer groups, and determines the impact correlation coefficients for each of the different matching customer groups based on the ratios of the changed number of credit users and the number of expanded users for each of the different matching customer groups to the changed number of credit users for the customer groups;

[0132] S314: When there is a matching customer group with an influence correlation coefficient greater than the preset influence correlation coefficient threshold, it is determined that the customer group belongs to the target influence customer group. When there is no matching customer group with an influence correlation coefficient greater than the preset influence correlation coefficient threshold, the process proceeds to step S315;

[0133] S315: When the sum of the influence correlation coefficients of different matching customer groups does not meet the requirement, it is determined that the customer group belongs to the target influence customer group. When the sum of the influence correlation coefficients of different matching customer groups meets the requirement, the process proceeds to step S32.

[0134] Optionally, the above step S32 includes the following contents:

[0135] S321 determines the change impact coefficients of different matching customer groups based on the changed number of credit users and the number of expanded users in different matching customer groups, combined with the current number of credit users in the customer groups. When the change impact coefficients of different matching customer groups are all less than a preset change impact coefficient threshold, the process proceeds to step S322. When there is a matching customer group whose change impact coefficient is not less than the preset change impact coefficient threshold, the process proceeds to step S323.

[0136] S322: When the sum of the influence correlation coefficients of different matched customer groups is less than the preset influence correlation coefficient value, it is determined that the customer group does not belong to the target influence customer group. When the sum of the influence correlation coefficients of different matched customer groups is not less than the preset influence correlation coefficient value, the process proceeds to step S323.

[0137] S323 When the number of matched processed customer groups whose change influence coefficient is not less than the preset change influence coefficient threshold does not meet the requirements, it is determined that the customer group belongs to the target influence customer group. When the number of matched processed customer groups whose change influence coefficient is not less than the preset change influence coefficient threshold meets the requirements, proceed to step S33.

[0138] Furthermore, the method for determining the verification processing strategy of the target influence customer group during the different matching customer group training processes is as follows:

[0139] Determine the number of expanded users of the matching processing customer group on the current date based on the data expansion of the matching processing customer group on the current date;

[0140] Determine the customer group weight coefficients of different matching customer groups based on the number of expanded users of different matching customer groups on the current date and their proportion in the number of expanded users on the current date;

[0141] The customer group influence value of the target influence customer group is determined based on the sum of the products of the influence correlation coefficients between the target influence customer group and different matching processing customer groups and the customer group weight coefficients. Based on the customer group influence value, the verification processing strategy of the target influence customer group during the training processing of different matching processing customer groups is determined.

[0142] Specifically, determining the verification processing strategy of the target influential customer group during different matching customer group training processes based on the customer group influence value specifically includes:

[0143] When the customer group influence value is greater than the preset customer group influence value threshold, after the different matching customer group training processes and all customer group training processes are completed, the target influence customer group verification process needs to be performed;

[0144] When the customer group influence value is not greater than the preset customer group influence value threshold, or when the customer group influence value is less than the preset influence threshold, no verification process of the target influence customer group is required during different matching customer group training processes;

[0145] When the customer group influence value is not less than the preset influence threshold, the verification processing strategy of the target influence customer group during the training process of different matching processing customer groups is determined according to the product of the influence correlation coefficient of different matching processing customer groups and the target influence customer group and the customer group weight coefficient.

[0146] Furthermore, when the product of the influence correlation coefficient between the matching processing customer group and the target influence customer group and the customer group weight coefficient is greater than the preset weight coefficient threshold, it is determined that the matching processing customer group needs to be verified in the target influence customer group during the training process.

[0147] It should be noted that the method for determining the training processing order of the matching processing customer group is:

[0148] Using verification strategies for different matching customer groups, determine the target influencing customer groups that need verification and processing for different matching customer groups, and use them as verification influencing customer groups;

[0149] According to the verification processing strategies of different verification-affecting customer groups in different matching processing customer groups, the proportion of matching processing customer groups that have been verified in the matching processing customer groups is determined, and the proportion is used as the verification demand coefficient;

[0150] Based on the sum of the verification requirement coefficients of different verification influencing customer groups of different matching processing customer groups, the verification requirement values ​​of different matching processing customer groups are determined, and the training processing order of the different matching processing customer groups is determined based on the verification requirement values.

[0151] Furthermore, the training processing sequence of different matching processing customer groups is determined based on the verification requirement value, specifically including:

[0152] The training processing order of different matching processing customer groups is determined from large to small according to the verification requirement value, and after all matching processing customer groups are trained according to the training processing order, the credit model training processing is performed based on all customer groups except the matching processing customer groups.

[0153] It is understandable that credit risk management and control based on the updated credit model specifically includes:

[0154] Based on the updated credit model and in combination with the credit application information of different credit users as input, the credit risk of the credit user is determined based on the output of the updated credit model.

[0155] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, refer to the descriptions of the method embodiments.

[0156] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0157] The foregoing description is merely one or more embodiments of this specification and is not intended to limit this specification. It will be apparent to those skilled in the art that various modifications and variations may be made to one or more embodiments of this specification. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of one or more embodiments of this specification are intended to be within the scope of the claims of this specification.

Claims

1. A risk management method, characterized in that: Specifically include: Divide the credit users into different customer groups based on their credit application information, and determine the adversarial processing customer group within the customer groups based on the distribution of the customer groups in the training data of the credit model and the distribution data of credit users on different dates; The adversarial customer group for the current date is used as the matching customer group. If the changes in the augmented data for different matching customer groups are used to determine whether the training processing risk of the credit model does not meet the requirements, proceed to the next step. Determine the target influencing customer group and its influence correlation coefficient with the different matched customer groups based on the changes in the credit user data of different customer groups, the credit user data of different matched customer groups, and the changes in the augmented data; Determining, based on the influence correlation coefficients between the target influence customer group and different matching customer groups and the data amplification status on the current date, verification processing strategies for different target influence customer groups during the training of different matching customer groups; determining, based on the verification processing strategies, the training processing sequence for different matching customer groups; and performing credit risk management and control based on the updated credit model; The method for determining the adversarial processing customer group in the customer group is: Based on the distribution of the customer group in the training data of the credit model, determine the proportion of credit users of the customer group in the historical training data of the credit model, and use this proportion as the proportion of the number of credit users of the customer group; Determine the date on which the credit users in the customer group exist based on the distribution data of the credit users on different dates, and use the date as the matching credit date; determining an adversarial processing requirement coefficient of the customer group based on a ratio of the number of matched credit dates to the number of users in the customer group, and determining whether the customer group belongs to an adversarial processing customer group based on the adversarial processing requirement coefficient; The method for determining the target influencing customer group in the customer group is: Determining the number of credit users in a customer group based on changes in credit user data for different customer groups; Determine the number of credit users changed and the number of expanded users for different matching customer groups based on the changes in the credit user data and expansion data for different matching customer groups; Determining the influence correlation coefficients of different matched customer groups based on the average value of the change in the number of credit users and the ratio of the number of expanded users to the change in the number of credit users of the customer groups, and determining whether the customer groups belong to the target influenced customer group based on the influence correlation coefficients of the different matched customer groups; The method for determining the verification processing strategy of the target impact customer group during the different matching customer group training processes is as follows: Determine the number of expanded users of the matching processing customer group on the current date based on the data expansion of the matching processing customer group on the current date; Determine the customer group weight coefficients of different matching customer groups based on the number of expanded users of different matching customer groups on the current date and their proportion in the number of expanded users on the current date; The customer group influence value of the target influence customer group is determined based on the sum of the products of the influence correlation coefficients between the target influence customer group and different matching processing customer groups and the customer group weight coefficients. Based on the customer group influence value, the verification processing strategy of the target influence customer group during the training processing of different matching processing customer groups is determined.

2. The risk management method according to claim 1, characterized in that: Divide the credit users into different customer groups, including: Determining the number of identical information items in the credit application information of different credit users based on similarities in the credit application information of the credit users; The credit users are divided into different customer groups according to the number of identical information items.

3. The risk management method according to claim 2, characterized in that: Credit users whose number of identical information items is greater than a preset threshold of the number of identical information items are classified into the same customer group.

4. The risk management method according to claim 1, wherein: The distribution of the customer group in the training data of the credit model includes the number of credit users of the customer group in the credit model and the proportion of credit users in the total number of credit users in the credit model.

5. The risk management method according to claim 1, wherein: When the confrontation processing requirement coefficient is greater than a preset processing requirement coefficient threshold, it is determined that the customer group belongs to the confrontation processing customer group.

6. The risk management method according to claim 1, characterized in that: The method for determining the training processing order of the matching processing customer group is: Using verification strategies for different matching customer groups, determine the target influencing customer groups that need verification and processing for different matching customer groups, and use them as verification influencing customer groups; According to the verification processing strategies of different verification-affecting customer groups in different matching processing customer groups, the proportion of matching processing customer groups that have been verified in the matching processing customer groups is determined, and the proportion is used as the verification demand coefficient; Based on the sum of the verification requirement coefficients of different verification influencing customer groups of different matching processing customer groups, the verification requirement values ​​of different matching processing customer groups are determined, and the training processing order of the different matching processing customer groups is determined based on the verification requirement values.

7. The risk management method according to claim 6, characterized in that: Determining the training processing sequence for different matching processing customer groups based on the verification requirement value specifically includes: The training processing order of different matching processing customer groups is determined from large to small according to the verification requirement value, and after all matching processing customer groups are trained according to the training processing order, the credit model training processing is performed based on all customer groups except the matching processing customer groups.

8. An automated model framework based on quasi-real-time effects, using a risk management method according to any one of claims 1 to 7, characterized in that: Specifically include: Variable pool module, quasi-real-time model module, data monitoring module; The variable pool module is responsible for extracting the credit characteristics of the credit application information of the credit user to obtain the variable pool; The quasi-real-time model module is responsible for performing quasi-real-time update processing of samples based on the variable pool according to a preset period, and performing incremental training of the credit model based on the credit model using samples after adversarial processing, and performing update processing of the credit model. After the update is completed, the model is put online; The data monitoring module is responsible for adding timestamp features to the credit model and using the timestamp features to analyze the credit model and the changing trends of the customer base; Use samples after adversarial processing to perform incremental training of the credit model. That is, introduce real-time sample information through adversarial weights to perform incremental training on the credit model. By analyzing the timestamp features, the time of the credit model and incremental training can be detected, and then the changing trends of the credit model and customer base can be analyzed.

9. The automation model framework based on quasi-real-time effect according to claim 8, characterized in that: The basic model is constructed by building the AutoML framework.

Images