Method, device, terminal equipment and storage medium for grouping risk control customers

Abstract

The present invention relates to the field of financial technology, and discloses a method, device, terminal device, and storage medium for grouping risk control customers. The method for grouping risk control customers of the present invention obtains a first label prediction value of the feature sample data by inputting the feature sample data of the risk control customer into a preset first prediction model; determines whether the feature sample data belongs to a difficult sample according to the first label prediction value; if not, the risk control customer is classified into an easy sample customer group; if so, the risk control customer is classified into a difficult sample customer group, and calls a preset second prediction model to perform secondary grouping processing on the risk control customer. The use of the technical solution of the present invention can avoid the reliance on expert experience in the existing risk control business scenarios for grouping customers, improve the quality and universality of the grouping of risk control customers, and thus improve the risk control effect on customers in various risk control business scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of financial technology (Fintech), and in particular to a method, apparatus, terminal device, and computer storage medium for grouping risk-controlled customers. Background Art

[0002] With the development of computer technology, more and more technologies are being applied in the financial field. Traditional financial industry is gradually transforming into financial technology. However, due to the increasing requirements of security, real-time performance and stability in the financial industry, higher requirements are also placed on technology.

[0003] Currently, in the process of risk control for customers in the financial industry, due to the existence of Simpson's paradox, it is usually necessary to group customers when the customer base composition is more complex, and then establish corresponding sub-models according to different situations. This is conducive to providing overall risk control prediction capabilities.

[0004] Currently, the mainstream methods for customer segmentation can be divided into two categories: experience-based segmentation and technical segmentation. However, experience-based segmentation relies heavily on the accuracy of expert experience. Using inappropriate features to segment customers can reduce the ultimate risk control effectiveness. Furthermore, experience-based segmentation is not universally applicable and requires the selection of appropriate segmentation dimensions for different data conditions and business scenarios. Furthermore, while technical segmentation has been automated to a certain extent, it is sensitive to the number of clusters and requires setting an appropriate number of clusters to achieve good segmentation quality. Furthermore, because technical segmentation is primarily data-driven compared to experience-based segmentation, customer segments segmented through technical segmentation still need to be combined with actual business analysis to ensure segmentation interpretability. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method, device, terminal equipment and computer storage medium for grouping risk control customers, aiming to avoid the reliance on expert experience in grouping customers in existing risk control business scenarios, improve the quality and universality of grouping risk control customers, and thus improve the risk control effect on customers in various risk control business scenarios.

[0006] To achieve the above objectives, the present invention provides a method for grouping risk control customers, the method comprising:

[0007] Inputting the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data;

[0008] Determining whether the feature sample data is a difficult sample according to the first label prediction value;

[0009] If not, the risk control customers will be classified into the easy sample customer group;

[0010] If so, the risk control customer is classified into a difficult sample customer group, and the preset second prediction model is called to perform secondary grouping processing on the risk control customer.

[0011] Optionally, the difficult sample customer group includes multiple customer groups with different sample difficulty levels;

[0012] The step of calling the preset second prediction model to perform secondary grouping processing on the risk control customers includes:

[0013] Inputting the feature sample data into the preset second prediction model to obtain a second label prediction value of the feature sample data;

[0014] Determining the sample difficulty of the feature sample data according to the second label prediction value;

[0015] The risk control customers are divided into target customer groups among the multiple customer groups with different sample difficulties according to the sample difficulty.

[0016] Optionally, before the step of inputting the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data, the method further includes:

[0017] Collect customer feature data to construct overall training samples;

[0018] A first initial model training is performed based on the overall training samples to obtain a preset first prediction model that calculates a first label prediction value based on the customer feature data.

[0019] Optionally, the method further includes:

[0020] A second initial model is trained based on the first label prediction value and the overall training sample to obtain a preset second prediction model for calculating the second label prediction value based on the customer feature data, wherein the algorithm used for the second initial model training is different from the algorithm used for the first initial model training.

[0021] Optionally, the step of performing a second initial model training according to the first label prediction value and the overall training samples includes:

[0022] Determine the dividing point between difficult and easy samples;

[0023] Dividing the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult-easy sample split point;

[0024] A second initial model training is performed based on the difficult samples.

[0025] Optionally, the sample difficulty / easy dividing point includes: a first dividing point and a second dividing point, the value of the first dividing point is greater than 0 and less than the value of the second dividing point, and the value of the second dividing point is less than 1.

[0026] Optionally, the step of dividing the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult and easy sample dividing point includes:

[0027] Determine, from the entire training sample, a first target sample whose first label prediction value is greater than 0 and less than the value of the first split point, and determine a second target sample whose first label prediction value is greater than the value of the second split point and less than 1;

[0028] dividing the first target sample and the second target sample into easy samples;

[0029] Determine, from the entire training sample, a third target sample whose first label prediction value is greater than the value of the first split point and is less than or equal to the value of the second split point;

[0030] The third target sample is divided into difficult samples.

[0031] In addition, to achieve the above-mentioned purpose, the present invention further provides a device for grouping risk-controlled customers. The device for grouping risk-controlled customers of the present invention includes:

[0032] A model prediction module, configured to input the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data;

[0033] a sample difficulty determination module, configured to determine whether the feature sample data is a difficult sample based on the first label prediction value;

[0034] A first grouping module, configured to classify the risk control customer into an easy sample customer group if the sample difficulty determination module determines that the customer is not satisfied;

[0035] The second clustering module is used to classify the risk control customer into a difficult sample customer group if the sample difficulty determination module determines that it is yes, and call the preset second prediction model to perform secondary clustering processing on the risk control customer.

[0036] Among them, each functional module of the device for grouping risk control customers of the present invention implements the steps of the risk control customer grouping method described above during operation.

[0037] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal device, which includes: a memory, a processor, and a risk control customer grouping program stored on the memory and runnable on the processor. When the risk control customer grouping program is executed by the processor, the steps of the risk control customer grouping method described above are implemented.

[0038] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer storage medium, on which a risk control customer grouping program is stored. When the risk control customer grouping program is executed by a processor, the steps of the risk control customer grouping method described above are implemented.

[0039] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the risk control customer grouping method as described above.

[0040] The present invention provides a method, apparatus, terminal device, computer storage medium and computer program product for clustering risk control customers, which obtains a first label prediction value of the feature sample data by inputting the feature sample data of the risk control customer into a preset first prediction model; determines whether the feature sample data belongs to a difficult sample based on the first label prediction value; if not, the risk control customer is classified into an easy sample customer group; if so, the risk control customer is classified into a difficult sample customer group, and calls a preset second prediction model to perform secondary clustering processing on the risk control customer.

[0041] When the present invention performs clustering processing on risk control customers in various risk control business scenarios in the financial field, the characteristic sample data of the risk control customer is input into a preset first prediction model to calculate the first label prediction value corresponding to the characteristic sample data. Thereafter, a sample difficulty analysis is performed based on the first label prediction value to determine whether the characteristic sample data belongs to a difficult sample. Thus, when it is determined that the characteristic sample data does not belong to a difficult sample, the risk control customer is directly classified into the easy sample customer group. Otherwise, when it is determined that the characteristic sample data belongs to a difficult sample, the risk control customer is immediately classified into the difficult sample customer group and the preset second prediction model is called at the same time to perform secondary clustering processing on the risk control customer.

[0042] Compared to traditional customer clustering methods based on expert experience or using technical clustering, the present invention divides risk control customers into difficult sample groups or easy sample groups based on the dimension of sample difficulty. Subsequently, the characteristic sample data classified as difficult sample groups is further subjected to model training and further clustering. In this way, the present invention not only avoids the reliance on expert experience when clustering customers in risk control business scenarios, but also improves the overall distinguishing power of risk control customer clustering to ensure customer clustering instructions. In addition, the method of clustering risk control customers proposed by the present invention can be extended to various business scenarios in the risk control field, effectively improving the universality of customer clustering.

[0043] In addition, the present invention groups risk control customers into difficult and easy samples, so that the weight distribution of each feature dimension in the two customer groups can better fit the corresponding customer groups, thereby avoiding the problem of strong features having too large weights and weak features being unable to take effect when the overall risk control model is trained based on customer group characteristics, improving the robustness of the risk control model, and effectively improving the risk control effect on customers in risk control business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the device structure of the hardware operating environment of the terminal device involved in the embodiment of the present invention;

[0045] Figure 2 This is a flow chart of the first embodiment of the method for grouping risk control customers of the present invention;

[0046] Figure 3 This is a schematic diagram of the application process involved in an embodiment of the method for grouping risk control customers of the present invention;

[0047] Figure 4 This is a schematic diagram of an application scenario involved in an embodiment of a method for grouping risk control customers of the present invention;

[0048] Figure 5 This is a schematic diagram of the functional modules of an embodiment of a device for grouping risk control customers according to the present invention.

[0049] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0050] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment of the terminal device involved in the embodiment of the present invention.

[0052] The terminal device of the embodiment of the present invention can be a business application device used in the financial field for customer risk management. Specifically, the terminal device can be a smart phone, PC (Personal Computer), tablet computer, portable computer, etc.

[0053] like Figure 1 As shown, the terminal device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may optionally also be a storage device independent of the aforementioned processor 1001.

[0054] Those skilled in the art will understand that Figure 1 The terminal device structure shown in the figure does not constitute a limitation on the terminal device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0055] like Figure 1 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a grouping program for risk control customers.

[0056] exist Figure 1 In the terminal shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the client and communicate data with the client; and the processor 1001 can be used to call the risk control customer grouping program stored in the memory 1005 and perform the following operation steps:

[0057] Inputting the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data;

[0058] Determining whether the feature sample data is a difficult sample according to the first label prediction value;

[0059] If not, the risk control customers will be classified into the easy sample customer group;

[0060] If so, the risk control customer is classified into a difficult sample customer group, and the preset second prediction model is called to perform secondary grouping processing on the risk control customer.

[0061] Furthermore, the difficult sample customer group includes multiple customer groups with different sample difficulty levels; the processor 1001 can also be used to call the risk control customer grouping program stored in the memory 1005 and perform the following operation steps:

[0062] Inputting the feature sample data into the preset second prediction model to obtain a second label prediction value of the feature sample data;

[0063] Determining the sample difficulty of the feature sample data according to the second label prediction value;

[0064] The risk control customers are divided into target customer groups among the multiple customer groups with different sample difficulties according to the sample difficulty.

[0065] Furthermore, the processor 1001 may also be configured to call a risk control customer grouping program stored in the memory 1005, and before executing the step of inputting the feature sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the feature sample data, further perform the following operation steps:

[0066] Collect customer feature data to construct overall training samples;

[0067] A first initial model training is performed based on the overall training samples to obtain a preset first prediction model that calculates a first label prediction value based on the customer feature data.

[0068] Furthermore, the processor 1001 may also be configured to call a risk control customer grouping program stored in the memory 1005 and perform the following steps:

[0069] A second initial model is trained based on the first label prediction value and the overall training sample to obtain a preset second prediction model for calculating the second label prediction value based on the customer feature data, wherein the algorithm used for the second initial model training is different from the algorithm used for the first initial model training.

[0070] Furthermore, the processor 1001 may also be configured to call a risk control customer grouping program stored in the memory 1005 and perform the following steps:

[0071] Determine the dividing point between difficult and easy samples;

[0072] Dividing the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult-easy sample split point;

[0073] A second initial model training is performed based on the difficult samples.

[0074] Furthermore, the sample difficulty and easy cutting points include: a first cutting point and a second cutting point, the value of the first cutting point is greater than 0 and less than the value of the second cutting point, and the value of the second cutting point is less than 1.

[0075] Furthermore, the processor 1001 may also be configured to call a risk control customer grouping program stored in the memory 1005 and perform the following steps:

[0076] Determine, from the entire training sample, a first target sample whose first label prediction value is greater than 0 and less than the value of the first split point, and determine a second target sample whose first label prediction value is greater than the value of the second split point and less than 1;

[0077] dividing the first target sample and the second target sample into easy samples;

[0078] Determine, from the entire training sample, a third target sample whose first label prediction value is greater than the value of the first split point and is less than or equal to the value of the second split point;

[0079] The third target sample is divided into difficult samples.

[0080] Based on the above hardware structure, various embodiments of the risk control customer grouping method of the present invention are proposed.

[0081] It should be noted that, at present, in the process of risk control for customers in the financial industry, due to the existence of Simpson's paradox, it is usually necessary to group customers when the customer base composition is more complex, and then establish corresponding sub-models according to different situations, which is conducive to providing overall risk control prediction capabilities.

[0082] Currently, the mainstream methods for customer segmentation can be divided into two categories: experience-based segmentation and technical segmentation. However, experience-based segmentation relies heavily on the accuracy of expert experience. Using inappropriate features to segment customers can reduce the ultimate risk control effectiveness. Furthermore, experience-based segmentation is not universally applicable and requires the selection of appropriate segmentation dimensions for different data conditions and business scenarios. Furthermore, while technical segmentation has been automated to a certain extent, it is sensitive to the number of clusters and requires setting an appropriate number of clusters to achieve good segmentation quality. Furthermore, because technical segmentation is primarily data-driven compared to experience-based segmentation, customer segments segmented through technical segmentation still need to be combined with actual business analysis to ensure segmentation interpretability.

[0083] In response to the above phenomenon, the present invention provides a method for clustering risk control customers. When clustering risk control customers in various risk control business scenarios in the financial field, the characteristic sample data of the risk control customer is input into a preset first prediction model to calculate a first label prediction value corresponding to the characteristic sample data. Thereafter, a sample difficulty analysis is performed based on the first label prediction value to determine whether the characteristic sample data belongs to a difficult sample. Thus, when it is determined that the characteristic sample data does not belong to a difficult sample, the risk control customer is directly classified into the easy sample customer group. Otherwise, when it is determined that the characteristic sample data belongs to a difficult sample, the risk control customer is immediately classified into the difficult sample customer group and the preset second prediction model is called at the same time to perform secondary clustering processing on the risk control customer.

[0084] Compared to traditional customer clustering methods based on expert experience or using technical clustering, the present invention divides risk control customers into difficult sample groups or easy sample groups based on the dimension of sample difficulty. Subsequently, the characteristic sample data classified as difficult sample groups is further subjected to model training and further clustering. In this way, the present invention not only avoids the reliance on expert experience when clustering customers in risk control business scenarios, but also improves the overall distinguishing power of risk control customer clustering to ensure customer clustering instructions. In addition, the method of clustering risk control customers proposed by the present invention can be extended to various business scenarios in the risk control field, effectively improving the universality of customer clustering.

[0085] In addition, the present invention groups risk control customers into difficult and easy samples, so that the weight distribution of each feature dimension in the two customer groups can better fit the corresponding customer groups, thereby avoiding the problem of strong features having too large weights and weak features being unable to take effect when the overall risk control model is trained based on customer group characteristics, improving the robustness of the risk control model, and effectively improving the risk control effect on customers in risk control business scenarios.

[0086] Please refer to Figure 2 , Figure 2 This is a flow chart of the first embodiment of the method for grouping risk control customers of the present invention. It should be noted that although a logical order is shown in the flow chart, in some cases, the method for grouping risk control customers of the present invention can also perform the steps shown or described in a different order than shown here.

[0087] In this embodiment, the risk control customer grouping method of the present invention is applied to the above-mentioned terminal device. The risk control customer grouping method of the present invention includes:

[0088] Step S10: inputting the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data;

[0089] In this embodiment, the terminal device inputs the characteristic sample data of the risk control customer into a first prediction model that has been trained and constructed in advance in any risk control business scenario in the financial field, so that the first prediction model performs model calculation based on the characteristic sample data to obtain the first label prediction value corresponding to the characteristic sample data.

[0090] It should be noted that in this embodiment, a risk control customer is a customer who needs risk control management in a risk control business scenario. Typically, the terminal device can directly determine the risk control customer based on parsing the risk control business command. In addition, the characteristic sample data of the risk control customer includes but is not limited to: the user's education level, age, account opening time, loan product type (such as consumer loan, business loan, and car loan), etc. The terminal device can specifically crawl and obtain various characteristic sample data of the risk control customer through the data interface between the terminal device and the external data platform.

[0091] Furthermore, in a feasible embodiment, before the above step S10, the risk control customer grouping method of the present invention may further include:

[0092] Collect customer feature data to construct overall training samples;

[0093] A first initial model training is performed based on the overall training samples to obtain a preset first prediction model that calculates a first label prediction value based on the customer feature data.

[0094] In this embodiment, the risk control customer grouping method of the present invention can be used in advance by the terminal device in the same manner as the above-mentioned crawling of the risk control customer feature sample data, and through the same data interface to perform customer feature data collection operations, thereby crawling and obtaining feature sample data such as the education level, age, account opening time, and loan product type of multiple customers, thereby constructing an overall training sample based on the feature sample data of each customer and each customer.

[0095] Afterwards, the terminal device can use the GBDT (Gradient Boosting Decision Tree) algorithm to perform the first initial model training based on the overall training sample, thereby obtaining the preset first prediction model that can calculate the first label prediction value based on the customer's feature sample data.

[0096] For example, Figure 3In the application process shown in this embodiment, it is assumed that the risk control customer clustering method of the present invention collects customer feature data in advance through the terminal device to construct the overall training sample [S1, S2, S3, ... Sn-1, Sn], where Sn represents the nth sample, and each sample corresponds to a customer and the customer's feature sample data. In this way, the terminal device uses the GBDT algorithm based on this overall training sample to train an initial model F1—the preset first prediction model.

[0097] Furthermore, in this embodiment, the method for grouping risk control customers may further include:

[0098] A second initial model is trained based on the first label prediction value and the overall training sample to obtain a preset second prediction model for calculating the second label prediction value based on the customer feature data, wherein the algorithm used for the second initial model training is different from the algorithm used for the first initial model training.

[0099] In this embodiment, after the terminal device constructs an overall training sample and uses the GBDT algorithm to perform a first initial model training to obtain the above-mentioned preset first prediction model, the terminal device further divides the overall training sample into difficult and easy samples based on the first label prediction value calculated for the overall training sample using the preset first prediction model, and then uses the difficult samples in the overall training sample to perform a second initial model training, thereby obtaining a preset second prediction model that can calculate the second label prediction value based on the customer's feature data.

[0100] It should be noted that, in this embodiment, the algorithm used by the terminal device when performing the second initial model training using the difficult samples in the overall training samples is different from the GBDT algorithm used by the terminal device when performing the first initial model training based on the overall training samples.

[0101] Furthermore, in this embodiment, the step of “performing second initial model training according to the first label prediction value and the overall training samples” may specifically include:

[0102] Determine the dividing point between difficult and easy samples;

[0103] Dividing the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult-easy sample split point;

[0104] A second initial model training is performed based on the difficult samples.

[0105] In this embodiment, after the terminal device constructs the overall training sample and uses the GBDT algorithm to perform the first initial model training to obtain the above-mentioned preset first prediction model, it further calculates the first label prediction values ​​of the overall training sample based on the preset first prediction model, thereby determining the appropriate difficult and easy sample splitting points based on the first label prediction values.

[0106] In this way, the terminal device divides the training samples corresponding to the first label prediction values ​​into difficult samples or easy samples according to the difficult and easy sample dividing point, and so on until the entire training sample is divided into two parts, difficult samples and easy samples. Finally, the terminal device uses all the difficult samples in the entire training sample to perform the second initial model training.

[0107] It should be noted that, in this embodiment, the terminal device may specifically use an optimization search method to determine the difficult and easy sample split points. Furthermore, the difficult and easy sample split points include: a first split point and a second split point, wherein the value of the first split point is greater than 0 and less than the value of the second split point, and the value of the second split point is less than 1, that is, the value a of the first split point is less than the value b of the second split point, and 0<a<b is less than 1.

[0108] Furthermore, in this embodiment, the step of “dividing the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult and easy sample split point” may specifically include:

[0109] Determine, from the entire training sample, a first target sample whose first label prediction value is greater than 0 and less than the value of the first split point, and determine a second target sample whose first label prediction value is greater than the value of the second split point and less than 1;

[0110] dividing the first target sample and the second target sample into easy samples;

[0111] Determine, from the entire training sample, a third target sample whose first label prediction value is greater than the value of the first split point and is less than or equal to the value of the second split point;

[0112] The third target sample is divided into difficult samples.

[0113] In this embodiment, after determining the difficult and easy sample splitting point, the terminal device can split the training samples corresponding to the first label prediction value in the overall training sample into easy samples or difficult samples according to the size between each first label prediction value and the value of the first splitting point and the value of the second splitting point in the difficult and easy sample splitting point.

[0114] That is: the terminal device first determines from the entire training sample all first target samples whose first label prediction value is greater than 0 and less than the value a of the first split point, and determines from the entire training sample all second target samples whose first label prediction value is greater than the value b of the second split point and less than 1; then, the terminal device divides all the first target samples and all the second target samples into easy samples.

[0115] In addition, the terminal device also determines all third target samples from the overall training samples whose first label prediction value is greater than the value a of the first split point and less than the value b of the second split point, and divides all third target samples into difficult samples.

[0116] For example, Figure 3 As shown in the application process, in this embodiment, the risk control customer clustering method of the present invention calculates the first label prediction value [P1, P2, P3, ..., Pn-1, Pn] of each training sample [S1, S2, S3, ... Sn-1, Sn] in the overall training samples through the terminal device according to the initial model F1 trained by the GBDT algorithm, where Pn = F1 (S1), which represents the first label prediction value of the output corresponding to the nth sample input initial model F1.

[0117] Afterwards, the terminal device finds appropriate difficult and easy sample splitting points based on the respective first label prediction values: the first splitting point a and the first splitting point b, so as to split the entire training sample into difficult samples and easy samples.

[0118] It should be noted that, in this embodiment, Figure 4 As shown in the figure, the label prediction value Pn in the risk control field ranges from 0 to 1. If Pn is closer to 1, it means that the risk control model has a greater confidence in predicting the sample label as positive. Conversely, if Pn is closer to 0, it means that the risk control model has a greater confidence in predicting the sample label as negative. In this way, the samples at the two ends (that is, the samples with label prediction values ​​less than a or greater than b) can be divided into easy samples, and the middle samples (samples with label prediction values ​​between a and b) can be divided into difficult samples.

[0119] Furthermore, when the terminal device can determine the first split point a and the second split point b through optimal search, assuming the terminal device uses a grid search method for optimal search, the terminal device uses a minimum step size of 0.01 and sets the difficulty threshold to T (the terminal device can set an appropriate threshold T based on actual risk control business scenarios to ensure that the number of difficult samples cut out can support the next step of training). The search space for the first split point a is [0, 0.01, 0.02, ... 1], with specific values ​​ranging from small to large, while the search space for b is [1, 0.99, 0.98, ... 0], with specific values ​​ranging from large to small. At the same time, the terminal device optimizes the target min(|K1-K2-T|), where K1 is the KS Kolmogorov-Smirnov (KS) value of the easy sample cut out, and K2 is the KS value of the corresponding difficult sample. In this way, when the set difficulty threshold T is larger, the number of difficult samples cut out from the overall training samples will be smaller, and the difference in KS values ​​between difficult and easy samples will be larger; conversely, the smaller the set difficulty threshold T is, the more difficult samples cut out from the overall training samples will be, and the difference in KS values ​​between difficult and easy samples will be smaller.

[0120] In this way, the terminal device can further divide the overall training samples into two parts, difficult samples and easy samples, based on the first dividing point a and the second dividing point b obtained by the optimal search. That is, the terminal device will divide the training samples [Sh1, Sh2, Sh3,…, Shi] whose first prediction values ​​calculated by the above-mentioned initial model F1 are greater than or equal to a and less than or equal to b into difficult samples, and will divide the training samples [Se1, Se2, Se3,…, Sej] whose first prediction values ​​calculated by the corresponding initial model F1 are less than a or greater than b into easy samples, where i represents the i-th difficult sample and j represents the j-th easy sample.

[0121] After that, the terminal device further uses other algorithms different from the GBDT algorithm, such as the Multi-Layer Perception MLP (Multi-Layer Perception, also known as Multi-Layer Perceptron), to perform a second initial model training based on the difficult samples [Sh1, Sh2, Sh3,…, Shi] cut out from the overall training samples to obtain an initial model F2 - the preset second prediction model. In this way, the risk control processing ability to distinguish difficult samples can be improved.

[0122] Finally, the terminal device can further choose to perform model fusion on the preset first prediction model and the preset second prediction model obtained through training. In this way, for the characteristic sample data of a new risk control customer, the terminal device uses the data as the test sample Snew, and adopts the following calculation process for the test sample Snew: calculate the initial model prediction value Pf1=F1(Snew) and the difficult sample optimization model prediction value Pf2=F2(Snew), if Pf1 is less than a or greater than b, output Pf1, otherwise output Pf2.

[0123] It should be noted that in this embodiment, the MLP algorithm used by the terminal device to train the above-mentioned preset second prediction model can also be replaced by other machine learning algorithms, such as logistic regression LR, support vector machine SVM, etc. The risk control customer grouping method of the present invention is not limited to the specific type of the machine learning algorithm, as long as the MLP algorithm used to train the above-mentioned preset second prediction model is different from the GBDT algorithm used to train the above-mentioned preset first prediction model.

[0124] In addition, when the terminal device uses an optimization search method to determine the above-mentioned difficult and easy sample splitting points, in addition to using a specific grid search method, the terminal device can also use other optimization search methods such as genetic algorithms, simulated annealing, etc.

[0125] Step S20, determining whether the feature sample data is a difficult sample according to the first label prediction value;

[0126] In this embodiment, the terminal device performs model calculation based on the characteristic sample data of the risk control customer through the above-mentioned first prediction model, thereby obtaining the first label prediction value corresponding to the characteristic sample data, and then further determines whether the characteristic sample data belongs to a difficult sample based on the size of the first label prediction value.

[0127] Exemplarily, the terminal device can specifically compare the first label prediction value corresponding to the characteristic sample data of the risk control customer with the values ​​of the difficult sample splitting points determined in the pre-model training: the value a of the first splitting point and the value b of the second splitting point, so that when it is compared that the first label prediction value is less than the first splitting point a (or greater than the second splitting point b), it is determined that the characteristic sample data of the risk control customer corresponding to the first label prediction value does not belong to a difficult sample but to an easy sample; conversely, when it is compared that the first label prediction value is greater than the first splitting point a but less than the second splitting point b, it is determined that the characteristic sample data of the risk control customer corresponding to the first label prediction value belongs to a difficult sample.

[0128] Step S30: If not, classify the risk control customer into an easy sample customer group;

[0129] Step S40: If yes, the risk control customer is classified into a difficult sample customer group, and a preset second prediction model is called to perform secondary grouping processing on the risk control customer.

[0130] In this embodiment, when the terminal device determines that the characteristic sample data of the risk control customer does not belong to the difficult sample group, it can directly classify the risk control customer into the easy sample customer group. When the terminal device determines that the characteristic sample data belongs to the difficult sample group, the terminal device will classify the risk control customer into the difficult sample customer group. Thereafter, the terminal device further calls the pre-trained preset second prediction model to calculate the second label prediction value of the risk control customer, and performs a secondary clustering process on the risk control customer based on the second label prediction value to further classify the risk control customer into a more accurate customer group.

[0131] Furthermore, in a feasible embodiment, the difficult sample customer group includes multiple customer groups with different sample difficulty levels. Based on this, the above step S40 of "calling the preset second prediction model to perform secondary grouping processing on the risk control customers" may specifically include:

[0132] Inputting the feature sample data into the preset second prediction model to obtain a second label prediction value of the feature sample data;

[0133] Determining the sample difficulty of the feature sample data according to the second label prediction value;

[0134] The risk control customers are divided into target customer groups among the multiple customer groups with different sample difficulties according to the sample difficulty.

[0135] In this embodiment, after the terminal device classifies the risk control customer into the difficult sample customer group, if the difficult sample customer group still includes multiple customer groups with different sample difficulty, the terminal device needs to further perform secondary grouping processing on the risk control customer, that is, the terminal device calls the preset second prediction model trained through the above process, and then inputs the feature sample data of the risk control customer belonging to the difficult sample into the second prediction model, and the second prediction model calculates and outputs the second label prediction value based on the feature sample data.

[0136] Afterwards, the terminal device can use the second label prediction value as the new first label prediction value, and further determine the sample difficulty of the feature sample data based on the same logic as the process described in steps S20 to S40 above, based on the size of the second label prediction value, that is, determine whether the feature sample data belongs to which difficult sample of different sample difficulties among the above-mentioned difficult samples. Repeat the sample difficulty determination process once to determine which part of the difficult samples with higher sample difficulty among the multiple sample difficulties the feature sample data belongs to. In this way, after finally determining the sample difficulty of the feature sample data, the terminal device can classify the risk control customer corresponding to the feature sample data into the target customer group corresponding to the sample difficulty among the multiple customer groups with different sample difficulties.

[0137] An embodiment of the present invention provides a method for clustering risk control customers. In any risk control business scenario in the financial field, a terminal device inputs the characteristic sample data of the risk control customer into a pre-trained and constructed first prediction model, whereby the first prediction model performs a model calculation based on the characteristic sample data to obtain a first label prediction value corresponding to the characteristic sample data. Subsequently, the terminal device further determines whether the characteristic sample data belongs to a difficult sample based on the magnitude of the first label prediction value. Thus, when the terminal device determines that the characteristic sample data of the risk control customer does not belong to a difficult sample, the risk control customer can be directly classified into an easy sample customer group. When the terminal device determines that the characteristic sample data belongs to a difficult sample, the terminal device will classify the risk control customer into a difficult sample customer group. Subsequently, the pre-trained preset second prediction model is further called to calculate the second label prediction value of the risk control customer, and a secondary clustering process is performed on the risk control customer based on the second label prediction value to further classify the risk control customer into a more accurate customer group.

[0138] Compared to traditional customer clustering methods based on expert experience or using technical clustering, the present invention divides risk control customers into difficult sample groups or easy sample groups based on the dimension of sample difficulty. Subsequently, the characteristic sample data classified as difficult sample groups is further subjected to model training and further clustering. In this way, the present invention not only avoids the reliance on expert experience when clustering customers in risk control business scenarios, but also improves the overall distinguishing power of risk control customer clustering to ensure customer clustering instructions. In addition, the method of clustering risk control customers proposed by the present invention can be extended to various business scenarios in the risk control field, effectively improving the universality of customer clustering.

[0139] In addition, the present invention groups risk control customers into difficult and easy samples, so that the weight distribution of each feature dimension in the two customer groups can better fit the corresponding customer groups, thereby avoiding the problem of strong features having too large weights and weak features being unable to take effect when the overall risk control model is trained based on customer group characteristics, improving the robustness of the risk control model, and effectively improving the risk control effect on customers in risk control business scenarios.

[0140] Furthermore, the present invention also provides a risk control customer grouping device, please refer to Figure 5 , Figure 5 This is a schematic diagram of the functional modules of an embodiment of a device for grouping risk control customers according to the present invention.

[0141] like Figure 5 As shown, the risk control customer grouping device of the present invention includes:

[0142] The model prediction module 10 is used to input the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data;

[0143] A sample difficulty determination module 20, configured to determine whether the feature sample data is a difficult sample based on the first label prediction value;

[0144] A first grouping module 30 is configured to classify the risk control customer into an easy sample customer group if the sample difficulty determination module determines that the risk control customer is not satisfied;

[0145] The second grouping module 40 is configured to classify the risk control customer into a difficult sample customer group if the sample difficulty determination module determines that the result is yes, and call a preset second prediction model to perform secondary grouping processing on the risk control customer.

[0146] Furthermore, the difficult sample customer group includes multiple customer groups with different sample difficulty levels;

[0147] The second clustering module 40 is further used to input the feature sample data into the preset second prediction model to obtain a second label prediction value of the feature sample data; determine the sample difficulty of the feature sample data according to the second label prediction value; and divide the risk control customers into target customer groups among the multiple customer groups with different sample difficulties according to the sample difficulty.

[0148] Furthermore, the risk control customer grouping device of the present invention further includes:

[0149] A model building module is used to collect customer feature data to construct an overall training sample; and perform a first initial model training based on the overall training sample to obtain a preset first prediction model that calculates a first label prediction value based on the customer feature data.

[0150] Furthermore, the model construction module of the risk control customer grouping device of the present invention is also used to perform a second initial model training based on the first label prediction value and the overall training sample to obtain a preset second prediction model for calculating the second label prediction value based on the customer characteristic data, wherein the algorithm used for the second initial model training is different from the algorithm used for the first initial model training.

[0151] Furthermore, the model building module includes:

[0152] Determination unit, used to determine the cutting point of difficult and easy samples;

[0153] a sample segmentation unit, configured to segment the entire training sample into difficult samples and easy samples according to the first label prediction value and the difficult and easy sample segmentation point;

[0154] A model training unit is used to perform a second initial model training based on the difficult sample.

[0155] Furthermore, the sample difficulty and easy cutting points include: a first cutting point and a second cutting point, the value of the first cutting point is greater than 0 and less than the value of the second cutting point, and the value of the second cutting point is less than 1.

[0156] Furthermore, the sample segmentation unit is also used to determine a first target sample from the overall training samples, whose first label prediction value is greater than 0 and less than the value of the first segmentation point, and to determine a second target sample whose first label prediction value is greater than the value of the second segmentation point and less than 1; to segment the first target sample and the second target sample into easy samples; to determine a third target sample from the overall training samples, whose first label prediction value is greater than the value of the first segmentation point and less than or equal to the value of the second segmentation point; and to segment the third target sample into a difficult sample.

[0157] Among them, the functional implementation of each module in the above-mentioned risk control customer grouping device corresponds to the various steps in the above-mentioned risk control customer grouping method embodiment, and its functions and implementation processes will not be repeated here one by one.

[0158] The present invention also provides a computer storage medium, which stores a risk control customer grouping program. When the risk control customer grouping program is executed by a processor, the steps of the risk control customer grouping method described in any of the above embodiments are implemented.

[0159] The specific embodiments of the computer storage medium of the present invention are basically the same as the embodiments of the risk control customer grouping method described above, and will not be described in detail here.

[0160] The present invention also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the risk control customer grouping method as described in any of the above embodiments.

[0161] The specific embodiments of the computer storage medium of the present invention are basically the same as the embodiments of the risk control customer grouping method described above, and will not be described in detail here.

[0162] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0163] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0164] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0165] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for grouping risk control customers, characterized in that: The method for grouping risk control customers includes: Collect customer feature data to construct overall training samples; Based on the overall training samples and using a gradient boosting decision tree algorithm, a first initial model training is performed to obtain a preset first prediction model that calculates a first label prediction value based on customer feature data; Inputting the characteristic sample data of the risk control customer into a preset first prediction model to obtain a first label prediction value of the characteristic sample data; Determining whether the feature sample data belongs to a difficult sample according to the first label prediction value, wherein the difficult sample refers to feature sample data whose first label prediction value is greater than a value of a first split point and less than a value of a second split point, the value of the first split point is greater than 0 and less than the value of the second split point, and the value of the second split point is less than 1; If not, the risk control customers are classified into the easy sample customer group; If so, the risk control customer is classified into a difficult sample customer group, and a preset second prediction model is called to perform secondary grouping processing on the risk control customer, wherein the preset second prediction model is trained using a multi-layer perceptron algorithm.

2. The method for grouping risk control customers according to claim 1, characterized in that: The difficult sample customer group includes multiple customer groups with different sample difficulty levels; The step of calling the preset second prediction model to perform secondary grouping processing on the risk control customers includes: Inputting the feature sample data into the preset second prediction model to obtain a second label prediction value of the feature sample data; Determining the sample difficulty of the feature sample data according to the second label prediction value; According to the sample difficulty, the risk control customers are divided into target customer groups among the multiple customer groups with different sample difficulties.

3. The method for grouping risk control customers according to claim 2, characterized in that: The method further comprises: A second initial model training is performed according to the first label prediction value and the overall training sample to obtain a preset second prediction model for calculating the second label prediction value based on the customer feature data, wherein the algorithm used for the second initial model training is different from the algorithm used for the first initial model training.

4. The method for grouping risk control customers according to claim 3, characterized in that: The step of performing a second initial model training according to the first label prediction value and the overall training sample comprises: Determine the dividing point between difficult and easy samples; Dividing the entire training sample into a difficult sample and an easy sample according to the first label prediction value and the difficult and easy sample dividing point; A second initial model training is performed based on the difficult samples.

5. The method for grouping risk control customers according to claim 4, characterized in that: The difficult and easy sample dividing points include: a first dividing point and a second dividing point, the value of the first dividing point is greater than 0 and less than the value of the second dividing point, and the value of the second dividing point is less than 1.

6. The method for grouping risk control customers according to claim 5, characterized in that: The step of dividing the overall training sample into a difficult sample and an easy sample according to the first label prediction value and the difficult and easy sample dividing point comprises: Determine a first target sample from the overall training samples whose first label prediction value is greater than 0 and less than the value of the first split point, and determine a second target sample whose first label prediction value is greater than the value of the second split point and less than 1; dividing the first target sample and the second target sample into easy samples; Determine, from the overall training samples, a third target sample whose first label prediction value is greater than the value of the first split point and is less than or greater than the value of the second split point; The third target sample is divided into difficult samples.

7. A risk control customer grouping device, characterized in that: The risk control customer grouping device includes: A model prediction module is used to collect customer feature data to construct an overall training sample; perform a first initial model training based on the overall training sample and use a gradient boosting decision tree algorithm to obtain a preset first prediction model that calculates a first label prediction value based on the customer feature data; input the feature sample data of the risk control customer into the preset first prediction model to obtain a first label prediction value of the feature sample data; a sample difficulty determination module, configured to determine whether the feature sample data belongs to a difficult sample according to the first label prediction value, wherein the difficult sample refers to feature sample data whose first label prediction value is greater than a value of a first split point and less than a value of a second split point, the value of the first split point is greater than 0 and less than the value of the second split point, and the value of the second split point is less than 1; A first grouping module, configured to classify the risk control customer into an easy sample customer group if the sample difficulty determination module determines that the risk control customer is not a sample customer group; The second grouping module is used to classify the risk control customer into a difficult sample customer group if the sample difficulty determination module determines that it is yes, and call the preset second prediction model to perform secondary grouping processing on the risk control customer, wherein the preset second prediction model is trained using a multi-layer perceptron algorithm.

8. A terminal device, characterized in that: The terminal device includes: a memory, a processor, and a risk control customer grouping program stored in the memory and executable on the processor. When the risk control customer grouping program is executed by the processor, the steps of the risk control customer grouping method described in any one of claims 1 to 6 are implemented.

9. A computer storage medium, characterized in that: The computer storage medium stores a risk control customer grouping program, and when the risk control customer grouping program is executed by the processor, the steps of the risk control customer grouping method according to any one of claims 1 to 6 are implemented.

Images