Data processing method and device under financial service

Through sample bucketing and subset division, the sample difference is calculated and object matching is performed, which solves the problem of insufficient sample randomness in financial business and improves the accuracy and credibility of data processing.

CN118967328APending Publication Date: 2024-11-15WEBANK (CHINA)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411064520.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Due to laws and regulations and social influences in financial business, it is difficult to directly collect highly random sample sets, which affects the equivalence between the experimental group and the control group, which in turn affects the accuracy and credibility of the data processing results.

Method used

By obtaining the sample set and dividing the sample into buckets and subsets, the sample difference is calculated. If the difference is greater than or equal to the preset value, object matching is performed to build the target sample set to enhance sample randomness.

Benefits of technology

It effectively reduces the impact of model processing accuracy on sample randomness, improves the enhancement of sample randomness, and ensures the accuracy and credibility of financial business data processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967328A_ABST
    Figure CN118967328A_ABST
Patent Text Reader

Abstract

The invention discloses a data processing method and device under financial businesses, and the method comprises the steps: obtaining a sample set, and enabling a business object in the sample set to be an experiment object or a contrast object; according to the historical transaction amount of each business object under the target financial business, carrying out sample bucket division to obtain a plurality of sample buckets, and then according to the characteristic value of each business object in the sample set under the target characteristic dimension, carrying out subset division on the sample set to obtain a plurality of sample subsets divided based on the target characteristic dimension; according to the number distribution of the experimental objects in the plurality of sample buckets in each sample subset and the number distribution of the contrast objects in the plurality of sample buckets, calculating a sample difference degree of a plurality of sample subsets divided based on the target feature dimension, and when the sample difference degree is greater than or equal to a preset difference degree, determining that the sample subsets are different; object matching is performed in each sample subset divided based on the target feature dimension to obtain a target sample set, so that the sample randomness can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data processing method and device for financial services. Background Art

[0002] Many business processes under financial business (such as the formulation of marketing strategies) usually need to be completed based on random experiments. However, due to restrictions such as laws and regulations and social influence, it is difficult to directly collect a sample set with a high degree of randomness from financial business, and thus it is impossible to ensure the equivalence between the experimental group and the control group, which affects the accuracy and credibility of the data processing results.

[0003] At present, in order to enhance the randomness of samples, regression models are usually used to predict the propensity scores of each sample in the sample set, and then samples with similar propensity scores are selected as samples for data processing. Since this method only screens samples based on the propensity score, it is very dependent on the model processing accuracy of the propensity score model. When the model processing accuracy is poor, the randomness of the selected samples will also be greatly reduced. In view of this, how to reduce the impact of model processing accuracy on sample randomness in financial business and effectively enhance sample randomness has become an important research topic at present. Summary of the invention

[0004] The embodiments of the present application provide a data processing method and device for financial business, which can reduce the impact of model processing accuracy on sample randomness in financial business and effectively enhance sample randomness.

[0005] On the one hand, an embodiment of the present application provides a data processing method for financial services, including:

[0006] Acquire a sample set, wherein the sample set includes multiple business objects under the target financial business, the business objects have feature values ​​under the reference feature dimension, and each business object is an experimental object or a control object;

[0007] According to the historical transaction amount of each business object in the sample set under the target financial business, the sample set is divided into sample buckets to obtain multiple sample buckets, and the historical transaction amounts corresponding to the business objects in different sample buckets are different;

[0008] Each reference feature dimension is used as a target feature dimension respectively, and the sample set is divided into subsets according to the feature value of each business object in the sample set under the target feature dimension, so as to obtain a plurality of sample subsets divided based on the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different;

[0009] Calculating the sample difference of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets;

[0010] If the sample difference is greater than or equal to the preset difference, each sample subset divided based on the target feature dimension will be used as a reference sample set, and object matching will be performed based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein, the target sample set includes each sample pair that has been successfully matched, and a sample pair includes an experimental object and a control object.

[0011] On the other hand, an embodiment of the present application provides a data processing device for financial services, including:

[0012] A sample acquisition unit, used to acquire a sample set, wherein the sample set includes a plurality of business objects under a target financial business, wherein the business objects have feature values ​​under a reference feature dimension, and each business object is an experimental object or a control object;

[0013] A sample bucketing unit, configured to bucket the sample set according to the historical transaction amount of each business object in the sample set under the target financial business, to obtain a plurality of sample buckets, wherein the business objects in different sample buckets correspond to different historical transaction amounts;

[0014] A subset division unit is used to use each reference feature dimension as a target feature dimension, and to perform subset division on the sample set according to the feature value of each business object in the sample set under the target feature dimension, so as to obtain a plurality of sample subsets divided under the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different;

[0015] A difference calculation unit, configured to calculate the difference of samples after subset division based on the target feature dimension according to a first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and a second quantity distribution of the control objects in each sample subset in the multiple sample buckets;

[0016] A sample screening unit is used to, if the sample difference is greater than or equal to a preset difference, use multiple sample subsets divided under the target feature dimension as reference sample sets, and perform object matching based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein the target sample set contains each business object that has been successfully matched.

[0017] On the other hand, an embodiment of the present application provides a computer device, including:

[0018] a processor adapted to implement one or more computer programs;

[0019] A storage medium storing one or more computer programs, wherein the one or more computer programs are suitable for being loaded by the processor and executing the method according to the first aspect.

[0020] On the other hand, an embodiment of the present application provides a storage medium, wherein the storage medium stores one or more computer programs, and the one or more computer programs are suitable for being loaded by a processor and executing the method of the first aspect.

[0021] On the other hand, an embodiment of the present application provides a program product, which includes a computer program, and the computer program is suitable for being loaded by a processor and executing the method of the first aspect.

[0022] In an embodiment of the present application, the computer device performs sample bucketing on the sample set based on the historical transaction amount to obtain multiple sample buckets, and performs subset division on the sample set based on the target feature dimension to obtain multiple sample subsets, and then determines the ability of the computer device to distinguish the control object from the experimental object based on the target feature dimension based on the number distribution of the experimental object and the control object in each sample bucket in the sample subset, and uses the sample difference of the sample subset to characterize the ability. Among them, the greater the sample difference, the stronger the ability to distinguish, and the feature corresponding to the target feature dimension can also be used as a key feature to distinguish the control object from the experimental object. However, it is obvious that the experimental object and the control object in the same sample subset have similar or identical feature values ​​under the target feature dimension. Then, matching the experimental object and the control object in the sample subset with a large sample difference to construct a target sample set can make the experimental object and the control object contained in each sample pair in the target sample set have similarities in key features, thereby confusing the computer device to distinguish the experimental object from the control object, and effectively enhancing the sample randomness of the target sample set. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a schematic diagram of the principle of enhancing sample randomness provided by this application;

[0025] Figure 2It is a flowchart of a data processing method under a financial business provided by this application;

[0026] Figure 3 This is a schematic diagram of the principle of selecting a target experimental object and a target control object provided by the present application;

[0027] Figure 4 It is a schematic diagram of a principle of obtaining a target sample set by recursive subset partitioning provided by the present application;

[0028] Figure 5a It is a schematic diagram of a recursive subset partitioning method provided by the present application;

[0029] Figure 5b It is a schematic diagram of another recursive subset partitioning method provided by the present application;

[0030] Figure 6 It is a flowchart of another data processing method under financial business provided by this application;

[0031] Figure 7 It is a schematic diagram of a data processing flow in an information promotion scenario provided by this application;

[0032] Figure 8a It is a schematic diagram of the distribution difference of the propensity scores of the experimental subject group and the control subject group provided by the present application;

[0033] Figure 8b is a schematic diagram of the distribution difference of the propensity scores of another experimental subject group and the control subject group provided by the present application;

[0034] Fig. 9 It is a Qini Curve schematic diagram provided by this application;

[0035] Fig.10 This is a schematic diagram of an Uplift Bin provided by this application;

[0036] Fig.11 It is a structural schematic diagram of a data processing device for financial services provided by the present application;

[0037] Fig.12 It is a structural schematic diagram of a computer device provided by this application. DETAILED DESCRIPTION

[0038] The embodiment of the present application provides a data processing scheme suitable for financial business scenarios, which can be used to improve the randomness of the sample customer group, so that the target business processing in the financial business scenario can obtain a processing result with a higher degree of credibility. In practical applications, the target business processing may refer to the business processing related to the control experiment based on the sample customer group, and the sample customer group is composed of multiple business objects under the financial business. The randomness of the sample customer group refers to the degree to which the probability of each business object being selected in the sample set meets the condition of being independent and equal. The higher the randomness, the closer the probability of each business object being selected, and each business object can be divided into an experimental object or a control object according to the corresponding rules. In other words, the sample set in the embodiment of the present application may contain multiple business objects, and any business object is an experimental object or a control object.

[0039] For ease of understanding, the following combination Figure 1 The principle shown briefly explains how this scheme enhances the randomness of the sample group. Figure 1 After obtaining the sample set, the sample set is first divided into sample buckets and subsets to obtain corresponding sample buckets and sample subsets. The sample bucketing is based on the historical transaction amount of each business object in the sample set under the target financial business. Each sample bucket obtained by the sample bucketing corresponds to a historical transaction amount interval, so that the historical transaction amounts corresponding to business objects in different sample buckets are different. The subset division is based on the feature values ​​of each business object in the sample set under the target feature dimension. The target feature dimension is selected from multiple reference feature dimensions. The reference feature dimension is the feature dimension where each business object has a corresponding feature value. Each selected target feature dimension is divided to obtain a set of sample subsets.

[0040] Furthermore, for each sample subset in each group of sample subsets, the number distribution of the experimental objects and the control objects in the sample subset in each sample bucket is determined respectively, so as to calculate the sample difference of the group of sample subsets according to the determined number distribution, and the sample difference is used to measure the distribution balance degree of the control objects and the experimental objects. When the sample difference is not less than the preset difference, each sample subset in the group of sample subsets is used as a reference sample set to obtain the target sample set by matching the experimental objects and the control objects in the reference sample set. Among them, the samples in the target sample set can be called pseudo-random samples, and the preset difference refers to the sample difference that a group of sample subsets that are allowed to be used to construct pseudo-random samples need to meet. In addition, the target sample set contains each sample pair that is successfully matched, and each sample pair consists of a control object and an experimental object.

[0041] In this scheme, the quantity distribution of experimental objects and control objects in each sample bucket in the sample subset can be used to measure whether the distribution of experimental objects and control objects in the sample subset is balanced. The more unbalanced the distribution is, the greater the sample difference is. Since the sample subset is divided based on the feature values ​​of each business object in the sample set under the target feature dimension, the sample difference corresponding to the sample subset can further reflect the ability of the target feature dimension to distinguish experimental objects from control objects. Moreover, the greater the sample difference is, the more effective the target feature dimension is in distinguishing experimental objects from control objects. The features corresponding to the target feature dimension can also be used as key features to distinguish control objects from experimental objects. It is obvious that the experimental subjects and control subjects in the same sample subset have similar or identical feature values ​​under the target feature dimension. Therefore, matching the experimental subjects and control subjects in the sample subset with large sample differences can make the experimental subjects and control subjects in each successfully matched sample pair have strong similarities in key features, confusing the computer equipment to distinguish between the experimental subjects and control subjects, thereby eliminating the influence of the aforementioned key features on the sample randomness and enhancing the randomness of the samples in each sample pair. Then the target sample set composed of such sample pairs also has strong randomness. It can be seen that when enhancing the randomness of the sample, this scheme does not need to rely too much on the propensity score of each business object, which reduces the influence of the processing accuracy of the propensity score model on the randomness of the sample.

[0042] In addition, by referring to the historical transaction amount of the business object for sample bucketing, a sample bucket can contain each business object in a historical transaction amount range, and the historical transaction amount can be used to reflect the transaction capacity of the business object. Therefore, based on the distribution differences of the business objects in the sample subset in each historical transaction amount range, the final distribution differences of the sample subset can be comprehensively determined. The final distribution differences can be comprehensively determined with reference to multiple groups of objects with transaction capabilities, so they have higher credibility and accuracy, and thus the constructed pseudo-random samples have stronger reference value.

[0043] Based on the above scheme principle, the present application embodiment proposes a data processing method under financial business. The flowchart of the method can be seen in Figure 2 , which is mainly used to enhance the randomness of samples by screening samples in the sample set. In actual application scenarios, the method can be executed by a terminal device or a server, or by a composite computing device composed of a terminal device and a server. For ease of explanation, the devices that can execute the method are collectively referred to as computer devices. Figure 2 The execution process of the method may specifically include S201-S205:

[0044] S201. Obtain a sample set, where the sample set includes multiple business objects under a target financial business. The business objects have feature values ​​under a reference feature dimension, and each business object is an experimental object or a control object.

[0045] In a specific embodiment, the target financial business refers to a financial business that requires business processing involving a control experiment. A business object refers to an object that has generated business data under the target financial business, usually a user (or service object) of the target financial business. In a control experiment scenario, each business object can be used as a control object or an experimental object. Among them, the experimental object is usually a business object that needs to accept a certain business processing to observe changes in its certain data, while the control object is usually a business object that does not need to accept the business processing to use this data as reference data. For example, in a scenario where it is necessary to determine the growth of transaction volume brought about by an information promotion strategy based on a control experiment, the experimental object can be a business object that accepts information promotion, the control object can be a business object that does not receive information promotion, and the transaction amount is an observed variable. By observing the difference in transaction amount growth between the two objects, it can be determined whether the information promotion strategy is effective.

[0046] In addition, the business data generated by the business object under the target financial business may include transaction data, object data, business tag data, etc. Among them, transaction data can be used to indicate characteristic information such as historical transaction amount and transaction frequency, and object data can include object attribute data and object behavior data. As an example, object attribute data can include object information under dimensions such as object age and city, and object behavior data can include behavior information under dimensions such as business browsing information, business participation information, and click-through rate of promotion information.

[0047] It can be seen that in the target financial business, each business object will generate data in one or more dimensions, and based on the data generated in one or more dimensions, the characteristic information (or characteristic value) of the business object in one or more characteristic dimensions can be analyzed. Then, in one implementation, the computer device can use the characteristic dimension in which each business object has a characteristic value as a reference characteristic dimension to obtain one or more reference characteristic dimensions, so that each business object in the sample set has a characteristic value in the reference characteristic dimension.

[0048] However, in real life, the differences in attributes or behaviors between different business objects will result in limited feature dimensions that are commonly possessed by different business objects. In order to enrich the selected reference feature dimensions, the embodiment of the present application proposes another implementation method. In this implementation method, the computer device can select the feature dimensions in which most business objects have feature values ​​as reference feature dimensions, so as to obtain rich reference feature dimensions. Among them, for business objects in which feature values ​​of a certain reference feature dimension do not exist in the sample set, the computer device fills the missing feature values ​​of the business object under the reference feature dimension according to a certain data filling method, and can also make each business object in the sample set have feature values ​​under the reference feature dimension. Moreover, data filling can effectively reduce the impact of the lack of feature values ​​on data processing accuracy. Optionally, the data filling method can be mean filling, mode filling, or model prediction value filling, etc., which are not limited here.

[0049] S202. Bucket the sample set according to the historical transaction amount of each business object in the sample set under the target financial business to obtain multiple sample buckets, where the business objects in different sample buckets have different corresponding historical transaction amounts.

[0050] In a specific embodiment, sample bucketing can be understood as sample grouping or sample binning, and its main purpose is to divide the sample population into multiple groups of samples (i.e., multiple sample buckets) according to the characteristic value interval of a certain dimension. The corresponding characteristic values ​​under the same group of samples belong to the same interval, so that the computer device can divide the sample population into multiple small groups with different trading capabilities. In this case, if the sample distribution of the sample population is comprehensively determined based on the sample distribution in each small group, the sample distribution of the sample population can have a higher reference value. Optionally, the sample bucketing method can be equal frequency bucketing (i.e., bucketing by ensuring that the amount of data in each bucket is the same), equidistant bucketing (i.e., bucketing by ensuring that the width of the characteristic value interval corresponding to each bucket is the same), cluster bucketing (i.e., bucketing based on the similarity of characteristic values), etc., which are not limited here.

[0051] In one implementation, the sample bucketing of the embodiment of the present application can be implemented by equal frequency bucketing based on the historical transaction amount of the business object. Among them, the historical transaction amount can refer to: the total transaction amount generated by the business object under a certain type or multiple types of transaction business within a certain period of time (such as the total withdrawal amount under the withdrawal business). In addition, each sample bucket obtained by the sample bucketing can contain one or more business objects in the sample set, and the historical transaction amount of each business object in the same sample bucket belongs to the same transaction amount interval. The number of business objects contained in different sample buckets is not necessarily different, but the transaction amount intervals corresponding to different sample buckets must be different. For example, assuming that two sample buckets are obtained after sample bucketing, the transaction amount interval corresponding to the first sample bucket is [10,000, 50,000), and the transaction amount interval corresponding to the second sample bucket is (50,000, 200,000], which means that the historical transaction amount of each business object under the first sample bucket is greater than or equal to 10,000 and less than 50,000, and the historical transaction amount of each business object under the second sample bucket is greater than or equal to 50,000 and less than 200,000.

[0052] S203, taking each reference feature dimension as a target feature dimension, and dividing the sample set into subsets according to the feature value of each business object in the sample set under the target feature dimension, to obtain multiple sample subsets divided based on the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different.

[0053] In a specific embodiment, when there are multiple reference feature dimensions, the computer device can use each reference feature dimension as a target feature dimension, and perform subset division of the sample set under the target feature dimension to obtain multiple sample subsets divided based on the target feature dimension (hereinafter referred to as a group of sample subsets corresponding to the target feature dimension), so that business objects in the same sample subset have the same or similar feature values ​​under the target feature dimension.

[0054] As an example, each group of sample subsets may include two sample subsets, and when subset division is performed based on the target feature dimension, the division rule adopted may be "whether the feature value of the business object under the target feature dimension is in a preset feature value range", so that business objects in the sample set whose feature values ​​are in the preset feature value range can be divided into the same sample subset, and business objects that are not in the preset feature value range can be divided into another sample subset.

[0055] S204. Calculate the sample differences of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets.

[0056] In a specific embodiment, the first quantity distribution is used to indicate the distribution of the experimental objects in the sample subset in each sample bucket, and the second quantity distribution is used to indicate the distribution of the control objects in the sample subset in each sample bucket. As an example, the first quantity distribution may specifically indicate the ratio between the number of target experimental objects in each sample bucket and the total number of experimental objects in the sample subset, while the second quantity distribution may specifically indicate the ratio between the number of target control objects in each sample bucket and the total number of control objects in the sample subset.

[0057] The target experimental object in any sample bucket refers to the experimental object contained in both the sample subset and the sample bucket. Figure 3 Taking the sample bucket and sample subset shown as an example, Figure 3 The target experimental objects contained in the sample bucket in can be as follows Figure 3 As shown in the left figure. Figure 3 It is not difficult to see that, assuming that the set of experimental objects in the sample bucket is the first set, and the set of experimental objects in the sample subset is the second set, the target experimental objects in the sample bucket are actually the objects contained in the intersection of the first set and the second set. Figure 3 The right picture in the figure shows the target control object, which will not be elaborated here.

[0058] That is to say, the first quantity distribution of each sample subset can be determined in the following manner: first, the total number of experimental objects contained in the sample subset (referring to the sample subset for which the first quantity distribution is currently to be determined) is obtained, and Figure 3 The principle shown in the middle left figure determines the number of target experimental objects contained in each sample bucket. Then, the ratio between the number of target experimental objects in each sample bucket and the aforementioned total number is used as the experimental object ratio corresponding to the sample bucket, and the experimental object ratio corresponding to all sample buckets is used as the first quantity distribution corresponding to the sample subset.

[0059] Correspondingly, the second quantity distribution of each sample subset can be determined after the sample subset for which the second quantity distribution is currently required is determined, and the total number of control subjects included in the sample subset is obtained, and the total number of control subjects included in the sample subset is determined as described above. Figure 3 The principle shown in the middle right figure determines the number of target control objects contained in each sample bucket. Then, the ratio between the number of target control objects in each sample bucket and the total number of control objects is used as the control object ratio corresponding to the sample bucket, and the control object ratio corresponding to all sample buckets is used as the second quantity distribution corresponding to the sample subset.

[0060] In the embodiment of the present application, the first quantity distribution and the second quantity distribution of the sample subset are mainly used to determine the sample difference of the sample subset, and then determine the sample difference of a group of sample subsets divided based on the target feature dimension based on the sample difference of each sample subset. The sample difference is used to indicate the difference in distribution between the control object and the experimental object.

[0061] In a specific implementation method, when determining the sample difference of any sample subset, the computer device can first determine the proportion of control objects and the proportion of experimental objects corresponding to the sample subset under each sample bucket, and then calculate the difference between the proportion of experimental objects and the proportion of control objects corresponding to each sample bucket, and then use the sum of the squares of the differences corresponding to the sample subset under all sample buckets as the sample difference of the sample subset.

[0062] As an example, the sample difference of each sample subset can be calculated as shown in Formula 1. In Formula 1, R i represents the sample difference of the i-th sample subset, Q represents the total number of sample buckets, j represents the j-th sample bucket, Indicates the proportion of experimental subjects in the j-th sample bucket corresponding to the sample subset, Indicates the proportion of control subjects in the j-th sample bucket corresponding to the i-th sample subset.

[0063]

[0064] Furthermore, the sample differences of the multiple sample subsets divided based on the target feature dimension can be weighted and summed to obtain the sample differences corresponding to the multiple sample subsets. As an example, the sample differences of the multiple sample subsets divided based on the target feature dimension can be calculated in the manner of Formula 2. In Formula 2, R represents the sample differences of the multiple sample subsets divided based on the target feature dimension, b represents the number of sample subsets in the multiple sample subsets (i.e., b sample subsets are multiple sample subsets), n i represents the number of business objects in the i-th sample subset, Sum represents the number of business objects in the sample set, and R i Represents the sample difference of the i-th sample subset.

[0065]

[0066] It is worth mentioning that, in order to simplify the calculation, in general, the computer device will divide the sample set into two sample subsets based on the target feature dimension, so that the sample difference of the multiple sample subsets divided based on the target feature dimension can be calculated as shown in Formula 3. In Formula 3, is the number of business objects in the first sample subset, represents the sample difference of the first sample subset, is the number of business objects in the second sample subset, Represents the sample difference of the second sample subset.

[0067]

[0068] S205. If the sample difference is greater than or equal to the preset difference, each sample subset divided based on the target feature dimension is used as a reference sample set, and object matching is performed based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein the target sample set includes each business object that has been successfully matched.

[0069] In a specific embodiment, the preset difference degree may be a fixed value determined based on historical experience values, or may be a value dynamically determined based on the sample difference degree R corresponding to each reference feature dimension.

[0070] In one implementation, when the preset difference is a fixed value, the computer device can directly use each sample subset obtained by dividing the target feature dimension as a reference sample set when the sample difference corresponding to the current target feature dimension is greater than or equal to the preset difference, so as to obtain the target sample set by matching objects in the reference sample set. When the sample difference corresponding to the current target feature dimension is less than the preset difference, other reference feature dimensions can be reselected as the target feature dimension, and the aforementioned S203 can be triggered to execute until the target feature dimension whose corresponding sample difference is greater than or equal to the preset difference is obtained. In this way, in the process of enhancing the randomness of samples, both computational efficiency and sample randomness of the constructed target samples can be taken into account.

[0071] In another implementation, when the preset difference is a dynamically determined value, the computer device may use the maximum sample difference corresponding to each reference feature dimension as the preset difference. Specifically, the computer device may use each reference feature dimension as the target feature dimension in turn, and trigger the execution of the above-mentioned S203-S204 processes to obtain the sample difference of a group of sample subsets corresponding to each reference feature dimension, and finally use the maximum sample difference among the obtained sample differences as the preset difference.

[0072] In practical applications, the sample difference corresponding to the target feature dimension can be used to indicate whether the distribution of experimental objects and control objects in each sample subset is balanced after the subset is divided based on the target feature dimension. Among them, the more unbalanced the distribution is, the more the target feature dimension can be used to accurately distinguish between experimental objects and control objects. In this case, it can be considered that the control objects and experimental objects contained in each sample subset divided based on the target feature dimension have escaped the distinction of the target feature dimension, so that the control objects in the sample subset and the experimental controls have a high degree of similarity under the target feature dimension, thereby effectively confusing the computer device to distinguish between experimental objects and control objects. Then, further, the computer device can obtain a sample customer group with a strong randomness by matching objects in the sample subset divided by the target feature dimension.

[0073] Based on this, it is not difficult to see that by selecting the maximum sample difference as the preset difference, and using each sample subset corresponding to the target feature dimension that is not less than the preset difference to construct the target sample set, the constructed experimental objects and control objects can cause strong confusion to the computer device, making it difficult for the computer device to distinguish the control objects and experimental objects in the target sample set based on the target feature dimension. This also ensures that the probability of the computer device selecting the control sample and the experimental sample from the target sample set is the same and independent, which ensures the sample randomness of the target sample set.

[0074] Then, it can be understood that in order to further enhance the sample randomness of the target sample set, in the actual implementation process, after determining the target feature dimension whose sample difference meets the preset difference, the reference sample sets obtained based on the target feature dimension can be further subsetted based on other feature dimensions to obtain a new reference sample set, so that the similarity between the control objects and the experimental objects in the reference sample set can be maximized as much as possible, thereby maximizing the sample randomness of the target sample set constructed based on the reference sample set.

[0075] As an example, the principle of constructing a target sample set by recursively partitioning the sample set into subsets can be found in Figure 4 .like Figure 4, after acquiring the sample set, the computer device can perform sample bucketing on the sample set to obtain multiple sample buckets, and subset the sample set under each reference feature dimension to obtain a group of sample subsets corresponding to each reference feature dimension. Then, the sample difference corresponding to each group of sample subsets is calculated according to the implementation method shown in the aforementioned S204, and a group of sample subsets whose sample difference is greater than or equal to the preset difference is determined. Further, each sample subset in the group of sample subsets is used as a reference sample set, and it is determined whether the reference sample set meets the preset conditions. When the reference sample set meets the preset conditions, object matching can be performed in the reference sample set to obtain one or more sample pairs. When the reference sample set does not meet the preset conditions, the reference sample set is used as a sample set, and the step of performing sample bucketing on the sample set is triggered.

[0076] The preset condition may include one or more of the following conditions (1)-(2): (1) the reference sample set is a sample subset of the Mth layer, where M is an integer greater than 1; (2) the number of business objects in the reference sample set is less than a first preset value. To facilitate understanding and simplify the operation logic, the following is an example in which the preset condition includes one condition.

[0077] In one implementation, when the preset condition is that the reference sample set is a sample subset of the Mth layer (M=2), the principle of recursive subset division of the sample set can be exemplarily referred to as Figure 5a . Among them, the M-th layer sample subset refers to: the sample subset obtained by performing the M-th subset division based on the initial sample set (i.e., the sample population). It should be noted that the feature dimensions used when dividing sample subsets of different layers are different, or in other words, the feature dimensions used when using the reference sample set as the sample set for subset division need to be different from the feature dimensions used by the ancestral set of the reference sample set. The ancestral set refers to the set that contains the reference sample set before the reference sample set is divided.

[0078] by Figure 5a For example, Figure 5a The sample set 1.1.1 in is the reference sample set used to construct the target sample set. Since the reference sample set is obtained by subset division of sample set 1.1, and sample set 1.1 is obtained by subset division of sample set 1, the ancestor set of sample set 1.1.1 includes sample set 1.1 and sample set 1. Therefore, the feature dimension (i.e., feature dimension 2.1) used when subset division of sample set 1.1 needs to be different from the aforementioned feature dimension 1, but the same feature dimension can be used for the division of sample subsets at the same level, i.e., feature dimension 2.1 can be the same as feature dimension 2.2.

[0079] In addition, it is not difficult to see that Figure 5aIn the example, sample set 1 is the initial sample set, which contains all the collected samples. Since this sample set has not been divided into subsets, it belongs to the 0th layer sample subset. Sample sets 1.1 and 1.2 are obtained by performing the first subset division on sample set 1 based on feature dimension 1. Therefore, sample sets 1.1 and 1.2 are both 1st layer sample subsets. Similarly, it can be understood that Figure 5a Then, when M=2, it means that the second layer sample subset is a reference sample set that meets the preset conditions, and the computer device should perform object matching based on each second layer sample subset to obtain the target sample set.

[0080] In another implementation, when the preset condition is that the number of business objects in the reference sample set is less than a first preset value, the principle of recursively dividing the sample set into subsets can be exemplarily referred to in Figure 5b .like Figure 5b , assuming that after the sample set 1 is subsetted, sample set 1.1 and sample set 1.2 are obtained, and the number of business objects in sample set 1.1 is less than (or equal to) the first preset value, the computer device will no longer subset the sample set 1.1, and use sample set 1.1 as a reference sample set for constructing the target sample set. If the number of business objects in sample set 1.2 is greater than the first preset value, the computer device will continue to subset the sample set 1.2 until a sample subset in which the number of business objects is less than the first preset value is obtained. By limiting the minimum sample size (i.e., the number of business objects) in the reference sample set, the excessive imbalance between the experimental objects and the control objects in the reference sample set can be minimized, thereby improving the success rate of object matching.

[0081] In addition, the aforementioned object matching can be achieved by bidirectional matching based on the propensity score of the business object, and the propensity score of the business object can be obtained by pre-predicting using a propensity score model, which is not limited here. In one implementation, the computer device can first select the control object with the closest propensity score to the experimental object for each experimental object in a reference sample set in a matching manner without replacement (i.e., the successfully matched business object is no longer used for subsequent matching), and the experimental object and the control object are used as a sample pair to obtain one or more forward matching sample pairs. Then, in the reference sample set, in a matching manner without replacement, select the experimental object with the closest propensity score to the control object for each control object, so as to obtain one or more reverse matching sample pairs. Then, the business objects in each forward matching sample pair and the business objects in each reverse matching sample pair are all used as the business objects contained in the target sample set, so as to construct the target sample set.

[0082] In an embodiment of the present application, the computer device performs sample bucketing on the sample set based on the historical transaction amount to obtain multiple sample buckets, and performs subset division on the sample set based on the target feature dimension to obtain multiple sample subsets, and then determines the ability of the computer device to distinguish the control object from the experimental object based on the target feature dimension based on the number distribution of the experimental object and the control object in each sample bucket in the sample subset, and uses the sample difference of the sample subset to characterize the ability. Among them, the greater the sample difference, the stronger the ability to distinguish, and the feature corresponding to the target feature dimension can also be used as a key feature to distinguish the control object from the experimental object. However, it is obvious that the experimental object and the control object in the same sample subset have similar or identical feature values ​​under the target feature dimension. Then, matching the experimental object and the control object in the sample subset with a large sample difference to construct a target sample set can make the experimental object and the control object contained in each sample pair in the target sample set have similarities in key features, thereby confusing the computer device to distinguish the experimental object from the control object, and effectively enhancing the sample randomness of the target sample set. It can be seen that this scheme does not need to rely too much on the propensity scores of each business object when enhancing the randomness of the sample, which reduces the impact of the processing accuracy of the propensity score model on the randomness of the sample.

[0083] Based on the above scheme principle, the present application embodiment proposes a data processing method under financial business. The flowchart of the method can be seen in Figure 6 This method is mainly used to dynamically weight each business object in the target sample set to further enhance the randomness of the sample customer group in the gain prediction process. Among them, gain prediction refers to predicting the gain transaction amount generated by the target object after receiving the promotion information. In actual application, Figure 6 The method shown can still be executed by the aforementioned computer device, and as Figure 6 As shown, the execution process of the method may specifically include S601-S604:

[0084] S601. Obtain the weight of each business object in the target sample set, and select a preset number of target sample subsets from the target sample set; wherein the weight is the inverse of the propensity score of the business object, and the propensity score is used to indicate the probability of the business object receiving the promotion information.

[0085] In a specific embodiment, the target sample subset refers to a subset of the target sample set, and the business objects in the target sample subset can be randomly selected from the target sample set. Different target sample subsets can contain the same business objects, such as target sample subset 1 and target sample subset 2 can both contain business object A. In addition, in order to determine the degree of reference to each business object during gain prediction, each business object can be weighted (i.e., weights can be added). In practical applications, the weight of each business object is the inverse of the propensity score of the business object. Since the propensity score is used to indicate the probability of the business object being subjected to information promotion, the weight generated based on the inverse of the probability can also be called an inverse probability weight.

[0086] Optionally, the propensity score of the business object may be the propensity score used when matching objects in the aforementioned S205, or may be the propensity score obtained by optimizing the propensity score model based on the target sample set and predicting the score of the business object using the optimized propensity score model, which is not limited here. It can be understood that in a sample set with strong randomness, the propensity scores of various business objects are similar and tend to 0.5.

[0087] S602. Construct a preset number of decision trees based on a preset number of target sample subsets and the weight of each business object in each target sample subset; wherein each decision tree is used to indicate an object classification rule constructed based on a target sample subset.

[0088] In a specific embodiment, the decision tree is generally a binary tree, and each node in the tree is used to store one or more business objects, and the bifurcation point in the tree is used to indicate a split feature, which generally includes a feature value interval under a feature dimension. It is not difficult to understand that in a binary tree, each split feature is associated with three nodes, one is a parent node, and the other two are child nodes. Business objects in the parent node that meet the split feature are grouped into one child node, and business objects that do not meet the split feature are grouped into another child node. In this case, the bifurcation point in a decision tree can constitute a set of object classification rules, and the business object in a leaf node in the decision tree is a business object under a certain category divided based on the set of object classification rules.

[0089] In one implementation, a target sample subset is used to construct a decision tree. The construction logic of the decision tree is similar to the division logic of the aforementioned sample subsets, both of which are obtained through recursive subset division, and the difference is that the subset division in the decision tree is implemented based on the gain value difference. Taking the decision tree as a binary tree as an example, specifically, when constructing a decision tree, the computer device can use the feature dimensions (which can be the aforementioned reference feature dimensions) that each business object has in common to construct split features (one split feature corresponds to one feature dimension), and then use the target sample subset as the root node, and divide the root node into sub-nodes based on whether each business object in the root node satisfies the split feature, so as to obtain two sub-nodes of the root node (respectively called left sub-nodes and right sub-nodes), and then calculate the reference gain value corresponding to each sub-node, so as to further calculate the gain value difference between the left sub-node and the right sub-node, and use it as the gain value difference corresponding to the split feature.

[0090] The reference gain value represents the gain transaction amount that each business object in the subset may generate after receiving the promotion information. The calculation method of the reference gain value can be exemplarily shown in Formula 4.

[0091]

[0092] In Formula 4, Node represents a child node (or subset), Indicates the reference gain value of the child node, n Node Indicates the number of business objects in this child node, W i Indicates the weight of the i-th business object in the child node, Y i represents the historical transaction amount of the i-th business object, T i Represents the type characteristics of the i-th business object, T i The value of is 1 or 0. i =1 means that the i-th business object is the experimental object, T i =0 indicates that the i-th business object is a control object. The historical transaction amount Y of all business objects in the child node i The mean of The type characteristics T of all business objects in the child node i The mean of .

[0093] Further, the gain value difference may be calculated in a manner as shown in Formula 5. In Formula 5, C1 represents the left child node, C2 represents the right child node, and Δ(C1, C2) represents the gain value difference between the left child node and the right child node. Indicates the number of business objects in the left subsection, Indicates the number of business objects of the right child node, n pIndicates the number of business objects in the parent node of the left child node and the right child node. Represents the reference gain value of the left child node, Indicates the reference gain value of the right child node.

[0094]

[0095] Based on the above principles, the gain value differences corresponding to the split features constructed by each feature dimension can be calculated. Then, the computer device can use the split feature with the largest gain value difference as the bifurcation condition used when dividing the root node into subnodes, and trigger the execution of the subnode division of the subnode according to the same principle until the node division stop condition is reached, thereby constructing a decision tree. Among them, the node division stop condition may include any of the following: (1) The subnode obtained by the current division is the X-order subnode of the root node, and X is an integer greater than 1; (2) The number of business objects contained in the subnode obtained by the current division is less than the second preset value. It should be noted that there is no absolute correlation between the second preset value and the aforementioned first preset value, and they can be set independently.

[0096] S603. Based on a preset number of decision trees, the propensity score of each business object in the target sample set is updated, and the step of obtaining the weight of each business object in the target sample set is triggered until the Nth batch of decision trees is obtained, where N is an integer greater than 1.

[0097] In a specific embodiment, each leaf node in the decision tree is used to store each business object of the same category in the corresponding target sample subset. In this case, when the computer device updates the propensity score of each business object in the target sample set based on a preset number of decision trees, it can first classify the business object using the object classification rule indicated by each decision tree to determine the category corresponding to the business object under each decision tree (i.e., determine the leaf node to which the business object belongs), and treat the business object belonging to the same category as the business object as a similar object of the business object under the corresponding decision tree. Then, the propensity score of the business object is updated based on each decision tree.

[0098] In one implementation, the computer device may first determine, under each decision tree, the sum of weights of similar objects of the business object under the decision tree (called the first weight sum), and determine the sum of weights of experimental objects among the similar objects (called the second weight sum), and then use the ratio of the first weight sum to the second weight sum as the propensity score of the business object under the decision tree. Finally, use the sum of the propensity scores of the business object under a preset number of decision trees as the new propensity score of the business object.

[0099] In actual application, the computer device can select N batches of target sample subsets (each batch contains a preset number of target sample subsets) in turn to construct N batches of decision trees, and update the propensity scores of the business objects N-1 times to determine a stable propensity score. Among them, the m+1th batch of target sample subsets is used to construct the m+1th batch of decision trees. The weight of the business object referenced when constructing the m+1th batch of decision trees is based on the inverse of the propensity score after the mth batch of decision trees are updated. The constructed m+1th batch of decision trees is used to update the propensity score of the business object for the m+1th time. In other words, the update of the m+1th propensity score is based on the propensity score after the mth update. By continuously constructing new decision trees to update the propensity score of the business object, the propensity score of the business object can be gradually stabilized, thereby ensuring that the propensity score has a higher degree of credibility.

[0100] For ease of understanding, the following describes the implementation of each update of the propensity score with specific formulas and examples. Taking the (m+1)th update of the propensity score as an example, the propensity score of the i-th business object in the target sample subset under the b-th decision tree can be calculated as shown in Formula 6.

[0101]

[0102] In formula 6, p i,b (m+1) represents the propensity score of the ith business object under the bth decision tree during the process of updating the propensity score for the m+1th time. sample W represents the number of business objects contained in the target sample subset used to construct the b-th decision tree, which belongs to the m+1-th batch of decision trees. k (m) is the weight corresponding to the kth business object in the target sample subset after the mth propensity score update, L b (x i ) represents the set of similar objects of the i-th business object. k ∈L b (x i ), T k =1}) is the indicative function. When the kth business object is a similar object to the ith business object and the kth business object is an experimental object, the value of the indicative function is 1, otherwise it is 0. k ∈L b (x i )}) is similar, but when the kth business object is a similar object to the ith business object, the value of the indicator function is 1, otherwise it is 0.

[0103] Furthermore, the new propensity score of the i-th business object can be calculated as shown in Formula 7. In Formula 7, Pi (m+1) represents the new propensity score of the ith business object after the (m+1)th propensity score update, B represents the preset number, and p i,b (m+1) means: the propensity score of the i-th business object under the b-th decision tree during the m+1-th propensity score update process.

[0104]

[0105] In addition, after the (m+1)th propensity score is updated, the weight of the i-th business object is calculated according to Formula 8.

[0106]

[0107] S604: Based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, the gain of the target object is predicted to obtain the gain transaction amount of the target object.

[0108] In a specific embodiment, the target object refers to the object for which the gain transaction amount needs to be calculated. Optionally, the target object can be a business object in the target sample set, or it can be another object other than the business object in the target sample set. Based on the above description, it can be known that a leaf node of the decision tree is used to store business objects of the same category in the target sample subset for constructing the decision tree. In this case, the computer device can calculate the historical transaction amount of the target object according to the following steps (1)-(3):

[0109] (1) According to the object classification rule indicated by each decision tree in the Nth batch of decision trees, the target object is classified based on the object features of the target object to obtain similar objects of the target object in each decision tree. The similar objects of the target object refer to business objects that belong to the same category as the target object.

[0110] (2) Based on the weight of each business object in the target sample set and the weight of similar objects in each decision tree when constructing the Nth batch of decision trees (i.e., the weight corresponding to the N-1th propensity score update), the reference weight of each business object is calculated. The reference weight of the business object is used to indicate the degree of reference to the business object when predicting the gain of the target object.

[0111] In one implementation, when calculating the reference weight of a business object, the computer device may first determine the similar objects of the business object in each decision tree, and then obtain the sum of the current weights of the similar objects under each decision tree to obtain the sum of the weights corresponding to each decision tree in the Nth batch of decision trees. Further, the ratio of the current weight of the business object to the sum of the weights is determined respectively, and the average of the determined ratios is used as the reference weight of the business object.

[0112] As an example, the ratio between the sum of weights corresponding to the ith business object and the bth decision tree can be regarded as the weight of the ith business object under the bth decision tree, so that the reference weight of the ith business object can be understood as the average value of the weights of the business object under the Nth batch of decision trees. Then, illustratively, the calculation method of the reference weight of the ith business object can be as shown in Formula 9.

[0113]

[0114] In Formula 9, x represents the target object, α i (x) represents the reference weight of the ith business object relative to the target object (i.e., the reference weight of the ith business object), B represents the preset number in the Nth batch of decision trees, α bi (x) represents the weight of the i-th business object relative to the target object under the b-th decision tree, and its specific calculation method can be as shown in Formula 10.

[0115]

[0116] In formula 10, W i (N-1) represents the weight corresponding to the i-th business object after the N-1th propensity score update, L b (x) represents the set of similar objects of the i-th business object in the b-th decision tree, W j (N-1 represents the weight corresponding to the jth similar object in the similar object set after the N-1th propensity score update.

[0117] (3) Based on the reference weight of each business object and the historical transaction amount of each business object, a gain prediction is performed on the target object to obtain the gain transaction amount of the target object.

[0118] As an example, the computer device may calculate the gain transaction amount of the target object in the manner as shown in Formula 11.

[0119]

[0120] In formula 11, Indicates the gain transaction amount of the target object, N mahcjrepresents the total number of business objects in the target sample set, i is the i-th business object in the target sample set, It represents the weighted mean of the historical transaction amount of each business object in the target sample set, and its calculation method can be shown in Formula 12. The weighted mean of the feature values ​​representing the type features of each business object in the target sample set can be calculated as shown in Formula 13.

[0121]

[0122] Among them, the understanding of each variable in Formula 12 can refer to the corresponding description in the previous text, which will not be repeated here.

[0123]

[0124] The understanding of each variable in Formula 13 can also refer to the corresponding description in the previous text, which will not be repeated here.

[0125] exist Figure 6 In the illustrated method embodiment, the reference weight of the business object is the average of the weights of the business object under the Nth batch of decision trees, which is equivalent to the reference weight of the business object being the weight after equalization, thereby avoiding the situation where the reference weight of the business object is too large or too small. Further, it can also avoid the computer device from excessively referring to or ignoring the historical transaction amount of certain business objects during the gain prediction process for the target object, thereby effectively improving the accuracy and credibility of the gain prediction results.

[0126] In addition, the weight of the business object under the Nth batch of decision trees is the inverse of the propensity score of each business object in the target sample set after the N-1th update. The propensity score of the business object is then updated based on this weight, which is equivalent to applying the technical principle of inverse probability weighting, which can make the propensity score update in the direction close to 0.5, further confusing the computer equipment to distinguish between the experimental objects and the control objects in the target sample set. Then, the reference weight determined when the propensity score approaches 0.5 can also enable the computer equipment to refer to the control objects and the experimental objects as equally as possible during the gain prediction process, which also enhances the randomness of the target sample set in the gain prediction stage.

[0127] It should also be noted that in the specific application scenario of the embodiment of the present application, the business processing under the target financial business mentioned in the aforementioned S205 may include information promotion processing. In order to increase the transaction amount brought about by information promotion, the embodiment of the present application proposes to predict the confidence of the gain transaction amount, so as to determine whether to promote information to the target object in combination with the gain transaction amount and its corresponding confidence.

[0128] Specifically, when determining whether it is necessary to promote information to the target object, the computer device can first obtain the similarity between the target object and each business object in the initial sample set, and then use the business object whose similarity is greater than or equal to the similarity threshold as the target business object. Then, when the target business object exists in the target sample set of S205, the confidence of the gain transaction amount of the target object is determined based on the similarity of the target business object according to the principle that the confidence and similarity are positively correlated. When the target business object does not exist in the target sample set, the confidence of the gain transaction amount of the target object is determined based on the similarity of the target business object according to the principle that the confidence and similarity are negatively correlated. Finally, when the confidence of the gain transaction amount is greater than the confidence threshold, and the gain transaction amount is greater than the transaction amount threshold, the promotion information is sent to the target object.

[0129] In order to facilitate the rapid application of the various methods proposed in this application, the following Figure 7 And specific application scenarios, an implementation process of an embodiment of the present application is briefly described.

[0130] In this example, the business process under the target financial business is to formulate an information promotion strategy that can increase the transaction amount of the target object. In this case, the business process can specifically include two stages: gain prediction and information promotion strategy formulation. Among them, the information promotion strategy is used to indicate the conditions for pushing interest-free coupons to the target object. By pushing interest-free coupons to the target object, the target object can be encouraged to withdraw money under the target financial business and increase the transaction amount of the target object.

[0131] Before describing the implementation process of this example in detail, let's first explain the terms in the example. First, let the characteristics of the user (i.e., business object) be X. i , i is the user's number, X i is a vector used to measure whether the user meets the conditions for the issuance of interest-free coupons. Whether user i is issued an interest-free coupon is recorded as T i , T i =1 means that user i is issued an interest-free coupon (the user is the experimental subject at this time), T i =0 means that user i has not been issued an interest-free coupon (in this case, the user is a control object). i represents the amount of money the user withdraws after being issued the interest-free coupon. Therefore, a user can be simply represented by a vector (X i , T i , Y i ).

[0132] For example, the user's age, historical withdrawal amount, city level, and years of work experience are all factors that may affect the issuance of interest-free coupons. So, for user i who is 28 years old, has a historical withdrawal amount of 100,000, lives in a second-tier city, has 3 years of work experience, and has been issued an interest-free coupon, the withdrawal amount after issuance is 50,000, his characteristics can be expressed as X i =(28, 100000, second-tier cities, 3), T i =1,Y i =50000. The user can be represented as: (X i , T i , Y i ) = (28, 100,000, second-tier cities, 3, 1, 50,000)

[0133] Under this definition, if Figure 7 , the process of formulating information promotion strategy mainly includes S701-S705:

[0134] S701. Sample preprocessing and construction of propensity score model.

[0135] Specifically, a plurality of users are first collected under the target financial business according to the aforementioned user representation method, and then an initial sample set is obtained, such as the sample set example in Table 1.

[0136] Table 1

[0137]

[0138] Since the above-mentioned original samples often have missing values ​​due to some reasons, if they are not processed, the accuracy and reliability of the model will be directly affected. At the same time, there are discrete features such as "city level". Intelligent models often only support numerical input. The existence of this variable causes the sample to not be directly input into the model for training. Therefore, before the data is input into the model, a series of data preprocessing steps must be performed to ensure the quality and applicability of the data. This requires a comprehensive inspection and processing of each feature to solve various problems in the data and process the data into a form that can be recognized by the model. The embodiment of the present application processes the original sample according to the following steps 1 to 5.

[0139] Step 1: Fill in missing values.

[0140] For missing features, specific methods are used to fill in, including but not limited to: mean filling, mode filling, deletion or prediction through models. For example, if the working years of user No. 2 in Table 1 are missing, mean filling can be used to fill in. Specifically, all users whose cities are first-tier cities, whose ages are 30-35 years old, and whose historical withdrawal amounts are between 450,000 and 550,000 are taken, and those whose working years are missing are eliminated, and then the mean working years of these users are calculated, and the mean is 5. Finally, this value is used to fill in the working years of user 2, so that the representation of this user can be changed from X2=(31, 500,000, first-tier cities, missing, 1, 150,000) to X2=(31, 500,000, first-tier cities, 5, 1, 150,000).

[0141] Step 2: One-hot encoding of discrete features.

[0142] For discrete features such as the city level, it is necessary to digitize it by one-hot encoding so that it can be smoothly input into the model and ensure the interpretability of the original features. One-hot encoding is a conventional encoding method and will not be described in detail here. After the processing of steps 1 and 2 is completed, the processed samples shown in Table 2 can be obtained, and the samples shown in Table 2 can be input into the propensity score model to train the model.

[0143] Table 2

[0144]

[0145] Step 3: Input the processed samples into the propensity score model for model training.

[0146] Specifically, after the aforementioned data processing is completed, a binary classification model is constructed to calculate a propensity score for each user, which reflects the possibility (or probability) of the user being issued an interest-free coupon. During the model training process, the model loss function used is as shown in Formula 15.

[0147]

[0148] In formula 15, i represents the user number, p i =P(T i =1|X i ) is the propensity score of user i, indicating the probability of user i being issued a coupon, T i Indicates whether the user is actually issued an interest-free coupon, T i =1 means user i is issued an interest-free coupon, otherwise it is recorded as T i = 0. The computer device can maximize the classification accuracy of the model by minimizing the model loss, thereby improving the accuracy of the predicted propensity score.

[0149] Step 4: Evaluate the propensity score model.

[0150] Specifically, after the propensity score model is trained, the "non-randomness" of the sample set can be evaluated by calculating the AUC value of the propensity score model. Among them, AUC is a common indicator used to measure the accuracy of classification models, and its range is 0.5-1.0. The closer the AUC value is to 1, the more confident the propensity score model is in determining whether a user is issued an interest-free coupon. For example, after training the model, the estimated AUC is 0.95, which means that the model is very confident in correctly identifying which users are likely to be issued coupons.

[0151] Step 5: Predict the propensity score for each user i.

[0152] Specifically, when the AUC is low, the propensity score model can be retrained until a propensity score model with a higher AUC is obtained. Then, all users are input into the model to obtain the propensity score estimate p for each user. i For example, the processed feature X1 of user 1 = (28, 100000, (0, 1, 0, 0, 0), 3) is input into the model, and the model outputs p i =0.93 means that the user's propensity score is 0.93, so that each user can be expressed as (X i , T i , Y i , p i ).

[0153] Furthermore, based on the prediction results p of each user i , evaluate whether the distribution of the predicted results of the experimental object group and the control object group is slightly different. If the difference is small (such as Figure 8a The distribution shown in the figure indicates that the randomness of the sample population is strong. If the difference is large (such as Figure 8b If the randomness of the sample population is weak, a pseudo-random sample needs to be constructed, and S702 is triggered.

[0154] S702: Construct a pseudo-random sample (ie, a target sample set) based on the principle of adaptive stratification.

[0155] After obtaining the propensity score of each user, similar users can be matched to obtain a pseudo-random sample. The embodiment of the present application proposes a hierarchical bidirectional matching method without replacement, and the matching process may include the following steps 1 to 8:

[0156] Step 1: Preprocessing and screening of matching samples.

[0157] First, according to p lower <p i <pupper Conditional filtering samples, p lower =0.01, p upper =0.99. i Samples that do not meet this condition are considered to have a “completely deterministic” strategy allocation and are not random. For example, when p i =0.99, it can be considered that the business party has almost never stopped issuing interest-free coupons to user i, so the sample is extremely lacking in control users (control users refer to users with extremely similar features in numerical values ​​but who have not been issued coupons). Therefore, these users need to be filtered out from the sample. The filtered sample set is recorded as D random .

[0158] Step 2: Adaptively select key features for stratification and automatically stratify the sample set.

[0159] In the sample set D random In the stratification, users are further automatically divided into multiple layers to maximize the randomness of different subsets within the layer. Before stratification, considering that the response variable (i.e., withdrawal amount) is a continuous variable, the amount of calculation required to directly calculate the distribution distance is large. Therefore, the withdrawal amount is first discretized and the extreme values ​​are logarithmized to reduce the impact on accuracy.

[0160] The discretization process is: random All U i Perform equal frequency bucketing (assuming the number of buckets is 10) and record the bucket number that Ui falls into. Logarithmic processing is: i Perform the process as in formula 16, so that the withdrawal amount is U i express.

[0161] U i =log(1+Y i ) Formula 16 then enters the adaptive stratification stage. The core idea of ​​adaptive stratification is to seek the optimal split point so that after the layer is split into two sub-layers based on the split point (equivalent to splitting a sample set into two sample subsets), the sum of the "differences" in the two layers is maximized. The sub-layers are continuously divided recursively based on this criterion until the stopping condition is reached. The stopping condition can be the number of divisions, the minimum number of people in the sub-layer, etc. In the embodiment of the present application, the calculation of the difference can be implemented as in the above-mentioned formulas 1 to 3, which will not be repeated here.

[0162] For example, suppose that D randomThe optimal stratification point calculated is "whether the city level is a second-tier city". After division on this basis, users in second-tier cities will be classified into the same sub-layer, such as user 1 and user 4; users in non-second-tier cities will be classified into another layer, such as user 3 and user 5. Assuming that the optimal division point within the layer calculated for second-tier cities is "whether the age is less than 36 years old", then user 1 and user 4 will be divided into different layers; the optimal division point within the layer of non-second-tier cities is "whether the years of work are less than 5 years", then user 3 and user 5 will also be separated. In the end, these four users all fell into different layers (i.e. subsets).

[0163] Step 3: In each layer that reaches the stopping condition, match the control group population with the experimental group population.

[0164] By performing automatic recursive partitioning according to the above method, several layers that meet the stopping conditions can be obtained. Then, two-way matching is performed based on the user's propensity score to further improve the randomness of the sample customer group. Specifically, for D random The experimental group population in each layer is matched with the control group population in a non-replacement matching method (each experimental group user is matched with only one closest control group user). Optionally, the population of the control group in the layer that is successfully matched is recorded as A0, and the maximum acceptable distance d for matching is max =0.01, and users in the control group exceeding this distance will not be counted in A0.

[0165] Taking user 1 as an example, since his strategy is to issue coupons, we need to find a control group user whose strategy is not to issue coupons and whose propensity score is closest to that of user 1 in the layer. Assume that control group user 11 falls into the same layer as user 1, and his propensity score p 11 =0.92, closest to user 1, with a distance of (0.93-0.92) 2 =0.0001 <d max , then user 11 can be classified as A0. And because it is a match without replacement, user 11 needs to be removed from the to-be-matched population. For all experimental group users in the layer, use the same logic to match the control group, so that A0 can be obtained.

[0166] Step 4: In each stratum, match the control group population with the experimental group population.

[0167] Similarly, for each layer of the control group population, match the experimental group population in the same way as step 3 (matching the maximum acceptable distance d max =0.01), recorded as A1. Specifically, taking user 11 mentioned in step 3 as an example, since user 1 is closest to user 11, user 1 will be classified into A1.

[0168] Step 5: Merge the sample groups that have been successfully matched within the layer.

[0169] The above two successfully matched sample groups are merged to obtain the intra-layer matching result A match =A0∪A1. For example, if user 1 and user 11 are successfully matched, both users will be classified into A. match .

[0170] Step 6: Combine the matched populations in all strata to obtain a pseudo-random experimental sample.

[0171] For all layers that meet the stop condition, match them according to steps 3 to 5, and then merge the matching results A in each layer match , we can get the final pseudo-random experimental sample D as shown in Table 3 match (i.e., the target sample set).

[0172] Table 3

[0173]

[0174] Step 7: Randomness evaluation of matching data.

[0175] Specifically, after obtaining the matching data D match After that, the matching effect can be evaluated. In a data set with good matching effect, we expect to see that the predictive ability of the propensity score model will be significantly reduced, and the prediction AUC should be close to 0.5. This is because in a sample customer group with strong randomness, the users in the experimental group and the control group are similar in key features (features that can effectively distinguish experimental subjects from control subjects), so it is difficult for the propensity score model to predict whether a user will be issued an interest-free coupon based on these features.

[0176] Step 8: Keep the best matching samples, retrain the propensity score model and make estimates.

[0177] Based on step 6 match Retrain the propensity score model and update D based on the trained propensity score model match The propensity score of each user in .

[0178] S703: construct a gain prediction model based on the target sample set to predict the gain transaction amount of the target object.

[0179] Specifically, this application proposes a self-learning weighted random forest (Learnable WeightedCausalForest, LWCF) and a self-learning dual machine learning causal forest (Learable Weighted CausalForest DML, LWCF-DML). The core of this type of model is the ability to adaptively iteratively calculate sample weights and further reduce the non-random influence of samples based on this weight.

[0180] In actual application, it is necessary to assign an initial weight to each sample (i.e., the user in the target sample set) before building the model. The initial weight can be the inverse of the propensity score recalculated in S702. When training the model, a certain upper limit F of the number of iterations is pre-set, and the model will adaptively update the weight after each iteration. For example, when the model iterates to the m+1th time, it will be based on the weights W calculated at the mth iteration. i (m) Build a decision tree. After the construction is completed, the model will update the weight of each sample to W based on the above formula 6-8 i (m+1), and this weight will be input into the construction of the decision tree in the m+2th iteration, and the whole process will be repeated until the weight W is updated. i (F) until.

[0181] Specifically, the following uses LWCF to construct a gain prediction model as an example for explanation. The process of using LWCF to construct a gain prediction model can be summarized as the following steps 1 to 8:

[0182] Step 1: Build a base learner.

[0183] The gain forest is often composed of multiple base learners, which are usually decision trees. Before building each tree, the original matching data set D match (The sample size is N match ) randomly select a subset D sample (i.e., the target sample subset), which is obtained by sampling with replacement, and then randomly extracting d features from all possible features, and then we can sample Build a decision tree with d features. sample All are stored in one node.

[0184] For example, if the sample sampling ratio is set to 70%, the input model feature X_i contains 4 features, and the feature sampling ratio is set to 50%. Then, we need to match 70% of the users are randomly selected with replacement, and 50% of the features are randomly selected from the four features. These selected users and the extracted features can be included in the same node, forming D sample .

[0185] Step 2: Determine D sample The optimal partition point.

[0186] The method of determining the optimal partition node in this step is similar to the adaptive stratification mentioned in S702, mainly in the difference of the partitioning criteria. The partitioning criteria of this step can be referred to in the above Figure 6 The relevant embodiments of S602 are not described in detail here.

[0187] Step 3: Recursive partitioning.

[0188] For each node, the algorithm will evaluate all possible splitting features and division points, and select the features and splitting points that maximize the splitting criteria to automatically split. When selecting splitting features for child nodes, the splitting features that can be selected must be different from the splitting features corresponding to their respective ancestor nodes.

[0189] For example, assuming that the consumption amount is the partition feature and the partition point is the consumption amount is less than 500, then the samples in the parent node will be split into two parts and fall into two child nodes, where the left child node C1 contains all samples with a consumption amount of less than 500, and the right child node C2 contains all samples with a consumption amount greater than 500. This process will be performed recursively. Specifically, the left child node C1 and the right child node C2 are used as new nodes to be divided. Each node calculates the optimal partitioning in accordance with the method of step 2 to achieve the splitting of the C1 node and the C2 node. This splitting process will continue automatically until the stopping condition is met. The stopping condition may be such as reaching the preset tree depth, the number of samples in the node is lower than a certain threshold, or the gain brought by the split is less than a certain threshold. Nodes that meet the stopping condition can no longer be split, and they can be called leaf nodes. It can be understood that D sample Each user in will fall into a leaf node, thus forming a tree. The bifurcation in the tree is a split feature.

[0190] Step 4: Build multiple base learners.

[0191] Referring to the above process, multiple decision trees are constructed. Each decision tree is constructed using the same logic, and the construction of the trees is independent of each other.

[0192] Step 5: Update the weight of each sample.

[0193] Specifically, the weight of each sample is updated by referring to the aforementioned formulas 6-8, which will not be described in detail here.

[0194] Step 6: Repeat steps 1 to 5 until F times are reached, and calculate the final weight W of each sample i (F).

[0195] Step 7: Weight each sample W i (F) is aggregated to obtain the reference weight of the sample, and the gain value of the user to be predicted (i.e., the gain transaction amount of the target object) is estimated based on the reference weight of each sample.

[0196] Specifically, the method of aggregating the weights of each sample can be referred to in Formulas 9-10, and the method of predicting the gain value can be referred to in Formulas 11-13, which will not be described in detail here. match All users in the data are input into the gain prediction model, and the gain value of each user is estimated in the above-mentioned way. The gain value can reflect the user's "sensitivity" to the issuance of interest-free coupons. The higher the gain value, the greater the increase in withdrawal amount after the coupon is issued.

[0197] It is worth mentioning that the principle of using LWCF-DML to build a gain prediction model is similar to that of LWCF. The only difference is that T needs to be added before entering the model. i With Y i Decentralized processing. Decentralized processing may include: for sample i (expressed as (X i ,T i ,Y i ), first remove it from the data set and use all samples X except sample i j (j=1,2,3…,N,j≠i) respectively fit T j and Y j , and then estimate the result of i based on the LWCF-DML model (assuming that it is ), T i Replace with We can represent user i as (X i ,T i new ,Y i new )

[0198] Step 8: Model evaluation.

[0199] The evaluation needs to be calculated based on the user's estimated gain value, whether to issue coupons, and the withdrawal amount. Evaluation generally refers to measuring the reliability of the gain prediction model through Qini Curve and Uplift Bin. Among them, Qini Curve is a cumulative gain graph. After training the gain prediction model, the gain value of each sample is estimated, which is recorded as uplift. i Sort in descending order according to the estimated gain value, denote k as the first k observations, and calculate R(k) for each observation according to Formula 17.

[0200]

[0201] In formula 17, the total sample size is N, is the sample size of the experimental group among the first k largest gain values. is the sample size of the control group among the first k largest gain values. is the sum of the responses of the experimental group among the first k maximum gain values. is the sum of the responses of the control group among the first k maximum gain values.

[0202] To describe the baseline in the same way as AUC, calculate the cumulative value R(N) at the last point and connect the origin to it to get Fig. 9 The line marked by 02 in the figure gives us the Qini Curve. Fig. 9 As shown. AUUC is calculated based on Qini Curve. Some improvements have been made here. AUUC in this application is calculated based on Fig. 9 The area enclosed by the line marked with 01 is Fig. 9 The line marked with 02 corresponds to the ratio between the areas of the rectangles.

[0203] Uplift Bin is a way to measure whether the evaluation order is accurate. First, the bins are divided into equal-frequency bins in descending order according to the gain value, and each bin is recorded as Bin. j , calculate the CATE value in each bucket to satisfy Equation 18.

[0204]

[0205] At the same time, the model estimate in each bucket is calculated according to Formula 19.

[0206]

[0207] This will give you an Uplift Bin, such as Fig.10 CATE j Corresponding to the line marked by 12 in the figure, Model j The corresponding line is marked by 11 in the figure. Generally, when the two lines fit together and the line marked by 12 shows a strict downward trend, it means that the model accuracy is high. When the evaluation result meets the expectations, the estimated gain result of each user is retained, and then the confidence model is constructed to judge the credibility of the estimated gain result.

[0208] S704: Construct a confidence model to predict the confidence of the gain transaction amount obtained by the gain prediction model.

[0209] The gain prediction model in S703 is based only on the matching result D matchHowever, the matching results are usually only applicable to those users who are successfully matched, which means that it may not cover all potential user groups. In addition, user behaviors and preferences are not static, but change over time, environment, personal experience and other factors. This dynamic change may cause the matching results to become less accurate over time, introducing uncertainty. To deal with this uncertainty, we introduce a mechanism to evaluate the reliability (i.e., confidence) of the gain value of each sample.

[0210] Specifically, a confidence model (binary classification model) can be trained to predict how similar the current sample is to the successfully matched sample. In this model, we mark the successfully matched samples as 1 and the unsuccessfully matched samples as 0. By training this confidence model, we can not only identify which users are similar to the successfully matched samples, but also output a confidence score for the gain value of each sample, which reflects the model's confidence in the gain value output.

[0211] As an example, when training a confidence model, you can use steps one to five:

[0212] Step 1: Feature selection.

[0213] Based on the characteristics of the propensity score model established in S701, p i As a new feature, because the matching is based on p i Therefore, the introduction of this feature can improve the classification AUC. For example, S703 uses X i As a feature, the feature used to build the confidence model can be expressed as (X i ,p i ).

[0214] Step 2: Model construction.

[0215] Use the users who have successfully matched and the users who have failed to match to build a binary confidence model. Assuming that user 2 is discarded due to a failed match, it should be marked as 0. Users 1, 3, 4, and 5 have successfully matched the corresponding users and should be marked as 1. All other users are marked according to the same logic, and the sample D used to build the binary confidence model can be obtained. confidence .

[0216] Step 3: Confidence assessment.

[0217] The confidence score output by the confidence model can be used to measure the matching quality of each sample. A high classification AUC means that the matching result is more reliable, while a classification AUC lower than expected may require further review or adjustment. For example, the classification AUC of the coupon issuance scenario is 0.91, which is a very high value, indicating that the confidence model is relatively reliable.

[0218] Step 4: Model estimation.

[0219] The confidence level of each sample is estimated and the result is saved. The result is combined with the gain value estimated in S703 to obtain the estimated result required for formulating the interest-free coupon distribution strategy.

[0220] S705: Perform information promotion based on the gain transaction amount of the target object and the confidence level of the gain transaction.

[0221] With the confidence level, you can combine it with the gain value to build a hybrid strategy to accurately identify which users should be reached by marketing activities (distribute interest-free coupons) and which users should not be reached (not distribute interest-free coupons). In actual applications, there are many ways to build a hybrid strategy, depending on business needs. The following is an example of a strategy combination:

[0222] (1) A 7-day interest-free coupon will be issued to users whose confidence level is greater than 0.8 and whose estimated gain is greater than 2000.

[0223] (2) Coupons will not be issued to users whose confidence level is greater than 0.8 and whose estimated gain value is less than 2000.

[0224] (3) For all people whose confidence level is less than 0.8, the original coupon issuance strategy is maintained, because the gain value of this group of people has a lower credibility.

[0225] After obtaining the hybrid strategy, you can launch it online and judge the effectiveness of the hybrid strategy by observing the changes in the withdrawal amount after the new strategy is launched.

[0226] The application process proposed in the embodiment of this application can help financial institutions (such as banks) effectively predict customer marketing gain effects in actual scenarios, and then make more informed decisions in applications. This method is suitable for actual scenarios where the strategy allocation mechanism is non-random and needs to process dynamically changing data, especially scenarios such as bank interest-free coupons and similar marketing placements. After all, coupon issuance strategies are often based on "strong rules", and user groups are always changing over time. The use of this application can improve marketing effects.

[0227] Based on the above Figure 2 and Figure 6 Related to various method embodiments, the present application also discloses a data processing device under financial services, which is used to perform the following Figure 2 and Figure 6 In the various steps shown in , the device may be a computer program (including program code) running on a computer device. And, specifically, see Fig.11 , the device may include Fig.11The sample acquisition unit 1101, the sample bucketing unit 1102, the subset division unit 1103, the difference calculation unit 1104 and the sample screening unit 1105 in the embodiment include:

[0228] The sample acquisition unit 1101 is used to acquire a sample set, wherein the sample set includes multiple business objects under the target financial business, the business objects have feature values ​​under the reference feature dimension, and each business object is an experimental object or a control object;

[0229] The sample bucketing unit 1102 is used to bucket the sample set according to the historical transaction amount of each business object in the sample set under the target financial business, so as to obtain multiple sample buckets, wherein the historical transaction amounts corresponding to the business objects in different sample buckets are different;

[0230] The subset division unit 1103 is used to use each reference feature dimension as a target feature dimension, and to divide the sample set into subsets according to the feature value of each business object in the sample set under the target feature dimension, so as to obtain a plurality of sample subsets divided based on the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different;

[0231] The difference calculation unit 1104 is used to calculate the sample difference of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets;

[0232] The sample screening unit 1105 is used to, if the sample difference is greater than or equal to a preset difference, use each sample subset divided based on the target feature dimension as a reference sample set, and perform object matching based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein the target sample set includes each successfully matched sample pair, and one sample pair includes an experimental object and a control object.

[0233] In one implementation, when used to determine the first quantity distribution corresponding to any sample subset, the difference calculation unit 1104 may execute:

[0234] Determine the total number of experimental objects in any of the sample subsets and the number of target experimental objects in each sample bucket; wherein the target experimental object refers to: an experimental object that exists in both the corresponding sample bucket and any of the sample subsets;

[0235] The ratio of the number of target experimental objects in each sample bucket to the total number is respectively used as the proportion of experimental objects corresponding to each sample bucket, and the proportion of experimental objects corresponding to the multiple sample buckets is used as the first quantity distribution corresponding to any sample subset.

[0236] In another embodiment, the second quantity distribution includes the proportion of control objects corresponding to each sample bucket; when the difference calculation unit 1104 is used to calculate the sample difference of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets, it can execute:

[0237] Under each of the sample subsets, respectively, determine the difference between the proportion of experimental subjects and the proportion of control subjects corresponding to each sample bucket;

[0238] The sum of squares of the differences corresponding to the multiple sample buckets under each sample subset is used as the sample difference degree of the corresponding sample subset;

[0239] The sample differences of the multiple sample subsets divided under the target feature dimension are weightedly summed to obtain the sample differences of the multiple sample subsets divided based on the target feature dimension.

[0240] In another embodiment, the number of reference feature dimensions is a target number; the sample screening unit 1105 may also be used to perform:

[0241] Obtain the sample difference corresponding to each reference feature dimension to obtain the target number of sample differences;

[0242] The largest sample difference among the target number of sample differences is used as the preset difference.

[0243] In yet another embodiment, the sample screening unit 1105 may further perform:

[0244] The reference sample set is used as the sample set, and the step of bucketing the sample set according to the historical transaction amount of each business object in the sample set under the target financial business is triggered until the obtained reference sample set meets the preset conditions; wherein the preset conditions include that the reference sample set is a sample subset of the Mth layer or the number of business objects in the reference sample set is less than a first preset value, M is an integer greater than 1, and the target feature dimensions used by the reference sample sets of different layers obtained by division are different;

[0245] The performing object matching based on the business objects in the reference sample set to obtain the target sample set includes:

[0246] Object matching is performed based on the business objects in the reference sample set that meet the preset condition to obtain the target sample set.

[0247] In yet another embodiment, the business processing includes gain prediction for the target object, and the gain prediction is used to predict the gain transaction amount generated by the target object after receiving the promotion information; Fig.11 The device shown may also include a weight acquisition unit 1106, a decision tree construction unit 1107, a data updating unit 1108, and a gain prediction unit 1109, wherein:

[0248] The weight acquisition unit 1106 is used to acquire the weight of each business object in the target sample set, and select a preset number of target sample subsets from the target sample set; wherein the weight is the inverse of the propensity score of the business object, and the propensity score is used to indicate the probability of the business object receiving the promotion information;

[0249] A decision tree construction unit 1107 is used to construct a preset number of decision trees based on the preset number of target sample subsets and the weights of each business object in each target sample subset; wherein each decision tree is used to indicate an object classification rule constructed based on a target sample subset;

[0250] The data updating unit 1108 is used to update the propensity score of each business object in the target sample set based on the preset number of decision trees, and trigger the step of obtaining the weight of each business object in the target sample set until the Nth batch of decision trees is obtained, where N is an integer greater than 1;

[0251] The gain prediction unit 1109 is used to perform gain prediction on the target object based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, so as to obtain the gain transaction amount of the target object.

[0252] In another embodiment, a leaf node of each decision tree is used to store each business object of the same category in the corresponding target sample subset; when the gain prediction unit 1109 performs gain prediction on the target object based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, and obtains the gain transaction amount of the target object, it can execute:

[0253] According to the object classification rule indicated by each decision tree in the Nth batch of decision trees, classify the target object based on the object feature of the target object to obtain similar objects of the target object in each decision tree, where the similar objects refer to business objects belonging to the same category as the target object;

[0254] Based on the weight of each business object in the target sample set and the weight of similar objects in each decision tree when constructing the Nth batch of decision trees, calculating the reference weight of each business object, wherein the reference weight is used to indicate: the degree of reference to the corresponding business object when performing gain prediction on the target object;

[0255] Based on the reference weights of the various business objects and the historical transaction amounts of the various business objects, a gain prediction is performed on the target object to obtain the gain transaction amount of the target object.

[0256] In yet another implementation manner, when calculating the reference weight of any service object in the target sample set, the gain prediction unit 1109 may perform:

[0257] Obtaining the sum of weights of similar objects in each decision tree;

[0258] The ratio between the weight of any business object and the sum of the weights corresponding to each decision tree is determined, and the average of the determined ratios is used as the reference weight of any business object.

[0259] In another embodiment, a leaf node of each decision tree is used to store each business object of the same category in the corresponding target sample subset; when the data updating unit 1108 updates the propensity score of any business object in the target sample set based on the preset number of decision trees, it can execute:

[0260] Classifying the any business object by using the object classification rule indicated by each decision tree, and determining the category of the any business object corresponding to each decision tree;

[0261] Determine similar objects of any business object under each decision tree, and use the ratio between the sum of weights of the similar objects and the sum of weights of the experimental objects in the similar objects as the propensity score of any business object under the corresponding decision tree; wherein the similar objects refer to business objects that belong to the same category as any business object;

[0262] The sum of the propensity scores of any business object under the preset number of decision trees is used as a new propensity score of any business object.

[0263] In yet another implementation, the business processing further includes information promotion processing; the gain prediction unit 1109 may also be used to perform:

[0264] Acquire the similarity between the target object and each business object in the sample set, and take the business object whose similarity is greater than or equal to the similarity threshold as the target business object;

[0265] If the target business object exists in the target sample set, the confidence of the gain transaction amount of the target object is determined based on the similarity of the target business object according to the principle that confidence and similarity are positively correlated;

[0266] If the target business object does not exist in the target sample set, the confidence of the target object's gain transaction amount is determined based on the similarity of the target business object according to the principle that confidence and similarity are negatively correlated;

[0267] When the confidence of the gain transaction amount is greater than the confidence threshold, and the gain transaction amount is greater than the transaction amount threshold, the promotion information is sent to the target object.

[0268] In a specific embodiment of the present application, Fig.11 The various units in the device shown are divided based on logical functions, and the above-mentioned various units can be separately or all combined into one or several other units to constitute, or one (some) of the units can be further divided into multiple smaller functional units to constitute, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. In other embodiments of the present application, the above-mentioned device may also include other units, and in actual applications, these functions can be implemented with the assistance of one or more other units.

[0269] In another specific embodiment of the present application, the program can be executed on a general computing device such as a computer device including a central processing unit (CPU), a random access memory (RAM), a read-only memory (ROM), and other processing elements and storage elements. Figure 2 or Figure 6 The computer program (including program code) of each step involved in the method shown in FIG. Fig.11 The apparatus shown in the figure is used to implement the corresponding method in the embodiment of the present application. The computer program can be recorded on a storage medium, for example, and loaded into the above-mentioned computer device through the storage medium, and can be run therein.

[0270] Based on the relevant description of the above method embodiment and the device embodiment, the present application embodiment also provides a computer device, see Fig.12 The computer device at least includes a processor 1201 and a storage medium 1202, and the processor 1201 and the storage medium 1202 are connected via a bus or other means. In addition, the processor 1201 can load and execute one or more computer programs stored in the storage medium 1202 to implement the above-mentioned Figure 2 or Figure 6 Corresponding method steps in the method embodiment shown.

[0271] The storage medium 1202 mentioned above is a memory device in a computer device, which is used to store programs and data. It is understandable that the storage medium 1202 here can include both built-in storage media in the computer device and extended storage media supported by the computer device. The storage medium 1202 provides a storage space, which stores the operating system of the computer device. In addition, one or more computer programs suitable for being loaded and executed by the processor 1201 are also stored in the storage space. These computer programs can be one or more program codes.

[0272] It should be noted that the above storage medium can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory; optionally, it can also be at least one storage medium located away from the above processor. Processor 1201 (or CPU (Central Processing Unit)) is the computing core and control core of the computer device, which is suitable for implementing one or more computer programs, and is specifically suitable for loading and executing one or more computer programs to implement corresponding method flows or corresponding functions.

[0273] Optionally, the computer device may also be in communication connection with one or more target output devices 1203 (such as mobile terminal devices, etc.), and may implement the distribution of promotional information or the collection of samples by interacting with the target output devices 1203 .

[0274] In addition, the present application also provides a storage medium, which stores the above Figure 2 or Figure 6 The one or more computer programs corresponding to the block generation method shown can implement the description of the corresponding method in the embodiment when one or more processors load and execute the one or more computer programs, which will not be repeated here.

[0275] The present application also provides a program product or a computer program, the program product includes a computer program, the computer program is stored in a storage medium. A processor in a computer device can read the computer program from the storage medium, and then execute the computer program, thereby enabling the computer device to perform the above Figure 2 or Figure 6 Various alternatives for the steps involved in the method shown.

[0276] It can be understood by those skilled in the art that all or part of the processes in the above embodiments can be completed by instructing the relevant hardware through a computer program, and the corresponding technical effects can be seen in the above description of the embodiments. Figure 2 or Figure 6 The relevant descriptions in the , and the technical effects corresponding to the various embodiments using the same principle are not repeated here. However, it should be noted that when this embodiment is applied to specific business products or technologies, data collection can only be carried out after obtaining the permission or consent of the relevant objects, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0277] It can also be understood that what is disclosed above is only a partial embodiment of the present application, and it certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and equivalent changes made according to the claims of this application are still within the scope of the invention.

Claims

1. A data processing method for financial services, characterized in that: include: Acquire a sample set, wherein the sample set includes multiple business objects under the target financial business, the business objects have feature values ​​under the reference feature dimension, and each business object is an experimental object or a control object; According to the historical transaction amount of each business object in the sample set under the target financial business, the sample set is divided into sample buckets to obtain multiple sample buckets, and the historical transaction amounts corresponding to the business objects in different sample buckets are different; Each reference feature dimension is used as a target feature dimension respectively, and the sample set is divided into subsets according to the feature value of each business object in the sample set under the target feature dimension, so as to obtain a plurality of sample subsets divided based on the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different; Calculating the sample difference of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets; If the sample difference is greater than or equal to the preset difference, each sample subset divided based on the target feature dimension will be used as a reference sample set, and object matching will be performed based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein, the target sample set includes each sample pair that has been successfully matched, and a sample pair includes an experimental object and a control object.

2. The method according to claim 1, characterized in that Methods for determining the first quantity distribution corresponding to any sample subset include: Determine the total number of experimental objects in any of the sample subsets and the number of target experimental objects in each sample bucket; wherein the target experimental object refers to: an experimental object that exists in both the corresponding sample bucket and any of the sample subsets; The ratio of the number of target experimental objects in each sample bucket to the total number is respectively used as the proportion of experimental objects corresponding to each sample bucket, and the proportion of experimental objects corresponding to the multiple sample buckets is used as the first quantity distribution corresponding to any sample subset.

3. The method according to claim 2, characterized in that The second quantity distribution includes the proportion of control objects corresponding to each sample bucket; the calculation of the sample difference of the multiple sample subsets divided based on the target feature dimension according to the first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and the second quantity distribution of the control objects in each sample subset in the multiple sample buckets includes: Under each of the sample subsets, respectively, determine the difference between the proportion of experimental subjects and the proportion of control subjects corresponding to each sample bucket; The sum of squares of the differences corresponding to the multiple sample buckets under each sample subset is used as the sample difference degree of the corresponding sample subset; The sample differences of the multiple sample subsets divided under the target feature dimension are weightedly summed to obtain the sample differences of the multiple sample subsets divided based on the target feature dimension.

4. The method according to any one of claims 1 to 3, characterized in that: The number of reference feature dimensions is a target number; the method further includes: Obtain the sample difference corresponding to each reference feature dimension to obtain the target number of sample differences; The largest sample difference among the target number of sample differences is used as the preset difference.

5. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: The reference sample set is used as the sample set, and the step of bucketing the sample set according to the historical transaction amount of each business object in the sample set under the target financial business is triggered until the obtained reference sample set meets the preset conditions; wherein the preset conditions include that the reference sample set is a sample subset of the Mth layer or the number of business objects in the reference sample set is less than a first preset value, M is an integer greater than 1, and the target feature dimensions used by the reference sample sets of different layers obtained by division are different; The performing object matching based on the business objects in the reference sample set to obtain the target sample set includes: Object matching is performed based on the business objects in the reference sample set that meet the preset condition to obtain the target sample set.

6. The method according to claim 1, characterized in that The business processing includes gain prediction for the target object, and the gain prediction is used to predict the gain transaction amount generated by the target object after receiving the promotion information; the method also includes: Obtaining the weight of each business object in the target sample set, and selecting a preset number of target sample subsets from the target sample set; wherein the weight is the inverse of the propensity score of the business object, and the propensity score is used to indicate the probability of the business object receiving the promotion information; Based on the preset number of target sample subsets and the weights of each business object in each target sample subset, a preset number of decision trees are constructed; wherein each decision tree is used to indicate an object classification rule constructed based on a target sample subset; Based on the preset number of decision trees, the propensity score of each business object in the target sample set is updated, and the step of obtaining the weight of each business object in the target sample set is triggered, until an Nth batch of decision trees is obtained, where N is an integer greater than 1; Based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, the gain of the target object is predicted to obtain the gain transaction amount of the target object.

7. The method according to claim 6, characterized in that A leaf node of each decision tree is used to store each business object of the same category in the corresponding target sample subset; based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, the gain prediction of the target object is performed to obtain the gain transaction amount of the target object, including: According to the object classification rule indicated by each decision tree in the Nth batch of decision trees, classify the target object based on the object feature of the target object to obtain similar objects of the target object in each decision tree, where the similar objects refer to business objects belonging to the same category as the target object; Based on the weight of each business object in the target sample set and the weight of similar objects in each decision tree when constructing the Nth batch of decision trees, calculating the reference weight of each business object, wherein the reference weight is used to indicate: the degree of reference to the corresponding business object when performing gain prediction on the target object; Based on the reference weights of the various business objects and the historical transaction amounts of the various business objects, a gain prediction is performed on the target object to obtain the gain transaction amount of the target object.

8. The method according to claim 7, characterized in that The method of calculating the reference weight of any business object in the target sample set includes: Obtaining the sum of weights of similar objects in each decision tree; The ratio between the weight of any business object and the sum of the weights corresponding to each decision tree is determined, and the average of the determined ratios is used as the reference weight of any business object.

9. The method according to claim 6, characterized in that A leaf node of each decision tree is used to store each business object of the same category in the corresponding target sample subset; based on the preset number of decision trees, a method for updating the propensity score of any business object in the target sample set includes: Classifying the any business object by using the object classification rule indicated by each decision tree, and determining the category of the any business object corresponding to each decision tree; Determine similar objects of any business object under each decision tree, and use the ratio between the sum of weights of the similar objects and the sum of weights of the experimental objects in the similar objects as the propensity score of any business object under the corresponding decision tree; wherein the similar objects refer to business objects that belong to the same category as any business object; The sum of the propensity scores of any business object under the preset number of decision trees is used as a new propensity score of any business object.

10. The method according to any one of claims 6 to 9, characterized in that: The business processing also includes information promotion processing; the method also includes: Acquire the similarity between the target object and each business object in the sample set, and take the business object whose similarity is greater than or equal to the similarity threshold as the target business object; If the target business object exists in the target sample set, the confidence of the gain transaction amount of the target object is determined based on the similarity of the target business object according to the principle that confidence and similarity are positively correlated; If the target business object does not exist in the target sample set, the confidence of the target object's gain transaction amount is determined based on the similarity of the target business object according to the principle that confidence and similarity are negatively correlated; When the confidence of the gain transaction amount is greater than the confidence threshold, and the gain transaction amount is greater than the transaction amount threshold, the promotion information is sent to the target object.

11. A data processing device for financial services, characterized in that: include: A sample acquisition unit, used to acquire a sample set, wherein the sample set includes a plurality of business objects under a target financial business, wherein the business objects have feature values ​​under a reference feature dimension, and each business object is an experimental object or a control object; A sample bucketing unit, configured to bucket the sample set according to the historical transaction amount of each business object in the sample set under the target financial business, to obtain a plurality of sample buckets, wherein the business objects in different sample buckets correspond to different historical transaction amounts; A subset division unit is used to use each reference feature dimension as a target feature dimension, and to perform subset division on the sample set according to the feature value of each business object in the sample set under the target feature dimension, so as to obtain a plurality of sample subsets divided under the target feature dimension, wherein the feature values ​​corresponding to the business objects in different sample subsets are different; A difference calculation unit, configured to calculate the difference of samples after subset division based on the target feature dimension according to a first quantity distribution of the experimental objects in each sample subset in the multiple sample buckets and a second quantity distribution of the control objects in each sample subset in the multiple sample buckets; A sample screening unit is used to, if the sample difference is greater than or equal to a preset difference, use multiple sample subsets divided under the target feature dimension as reference sample sets, and perform object matching based on the business objects in the reference sample set to obtain a target sample set, so as to perform business processing under the target financial business based on the target sample set; wherein the target sample set contains each business object that has been successfully matched.

12. The device according to claim 11, characterized in that The business processing includes gain prediction of the target object, and the gain prediction is used to predict the gain transaction amount generated by the target object after receiving the promotion information; the device also includes: A weight acquisition unit, used to acquire the weight of each business object in the target sample set, and select a preset number of target sample subsets from the target sample set; wherein the weight is the inverse of the propensity score of the business object, and the propensity score is used to indicate the probability of the business object receiving the promotion information; A decision tree construction unit, based on the preset number of target sample subsets and the weights of each business object in each target sample subset, constructs a preset number of decision trees; wherein each decision tree is used to indicate an object classification rule constructed based on a target sample subset; A data updating unit, configured to update the propensity score of each business object in the target sample set based on the preset number of decision trees, and trigger the step of obtaining the weight of each business object in the target sample set until the Nth batch of decision trees is obtained; A gain prediction unit is used to perform gain prediction on the target object based on the object characteristics of the target object, the historical transaction amount of each business object in the target sample set, and the weight of each business object when constructing the Nth batch of decision trees, so as to obtain the gain transaction amount of the target object.