Risk control strategy model retraining triggering method and device, equipment and medium

Abstract

The invention provides a risk control strategy model retraining triggering method and device, equipment and a medium. The method comprises the following steps: acquiring first feature distribution of first feature data and second feature distribution of second feature data; and triggering the risk control strategy model to perform retraining by using the third feature data under the condition that distribution drift occurs between the second feature distribution and the first feature distribution. Wherein the first feature data is sample data when the risk control strategy model is trained. The second feature data is the current feature data. The first feature data includes a plurality of feature dimensions, and the second feature data also includes a plurality of feature dimensions. The first feature data and the second feature data have the same dimension. At least one feature dimension in the third feature data and the first feature data is different.

Description

Technical Field

[0001] The present application relates to artificial intelligence, and in particular to a risk control strategy model retraining triggering method, device, equipment and medium. Background Art

[0002] Fraud is a particularly dangerous behavior in transactions. Once the fraud is not identified, it often causes irreparable losses. Therefore, the identification of fraud attacks covers every link of the entire risk control business. However, the current risk control strategy for identifying fraud attacks is mainly set up based on the experience of experts. It is difficult to set a high-quality risk control strategy (referred to as: risk control strategy) directly based on the experience of experts, and it is also difficult to maintain and optimize the risk control strategy in the later stage. Summary of the invention

[0003] The present application provides a risk control strategy model retraining trigger method, device, equipment and medium, which can automatically trigger the training of the risk control strategy model when distribution drift occurs, so as to facilitate the maintenance and optimization of the risk control strategy in the later stage.

[0004] In a first aspect, a method for triggering retraining of a risk control strategy model is provided. A computing device obtains a first feature distribution of first feature data and a second feature distribution of second feature data. When a distribution drift occurs between the second feature distribution and the first feature distribution, the risk control strategy model is triggered to use the third feature data for retraining, or when a distribution drift does not occur between the second feature distribution and the first feature distribution, the second feature data is input into the risk control strategy model for processing. The first feature data is sample data during the training of the risk control strategy model. The second feature data is current feature data. The first feature data includes multiple feature dimensions, the second feature data also includes multiple feature dimensions, and the first feature data and the second feature data have the same dimension. At least one feature dimension of the third feature data and the first feature data is different. The first feature data is feature data obtained by processing transaction data. For example, taking the transaction amount as an example, the transaction amount can be averaged to obtain a feature. When the transaction amount is processed more, for example, variance processing, median processing, maximum processing, minimum processing, etc., more features of different dimensions can be obtained. The same is true for the second feature data and the third feature data.

[0005] In the above scheme, it is possible to automatically detect whether the feature distribution of the current second feature data has generated distribution drift relative to the feature distribution of the first feature data used to train the risk control strategy model, and determine whether to use the risk control strategy model to generate a risk control strategy for the input second feature data, and whether to trigger retraining of the risk control strategy model, based on whether drift has occurred. This can prevent the law learned by the risk control strategy model on the first feature data from no longer being applicable to the second feature data, resulting in a decrease in the performance of the risk control strategy model.

[0006] In some possible designs, the parameters of the risk control strategy model include feature thresholds and feature weights. The feature thresholds and feature weights are obtained by training the risk control strategy model using the first feature data, and the feature thresholds include multiple dimension thresholds of the same number as the dimension of the first feature data. The feature weights include multiple dimension weights of the same number as the dimension of the first feature data. That is, when the dimension of the first feature data is n, the dimension of the feature thresholds and the feature weights is also n. n is an integer greater than zero.

[0007] The process of generating a risk control strategy for the second feature data through the risk control strategy model can be as follows: first, determine multiple differences between multiple dimensional features of the second feature data and multiple dimensional thresholds. Then, determine the coefficients of multiple dimensional weights according to the multiple differences. Specifically, when the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is larger, the coefficient of the dimensional weight of the dimension is determined to be closer to the first value; when the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is smaller, the coefficient of the dimensional weight of the dimension is determined to be closer to the second value. Afterwards, determine whether the multiple dimensional weights are selected according to the coefficients of the multiple dimensional weights. Taking one of the dimensions as an example, when the coefficient of the dimensional weight of one of the dimensions is greater than the selected threshold, determine that the dimensional weight of the dimension is selected; when the coefficient of the dimensional weight of one of the dimensions is less than or equal to the selected threshold, determine that the dimensional weight of the dimension is not selected. Similarly, the same applies to other dimensions. Finally, determine the output value of the risk control strategy model according to the sum of the selected dimensional weights. Specifically, when the sum of the selected dimension weights is larger, it is determined that the output value of the risk control strategy model is closer to the first value; when the sum of the selected dimension weights is smaller, it is determined that the output value of the risk control strategy model is closer to the second value. The increment of the coefficient of the dimension weight of one dimension of the second feature data caused by the increment of the difference between the dimensional feature of the dimension and the dimensional threshold of the dimension is α times the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimension weights, where α is a real number greater than 1.

[0008] In the above scheme, the increment of the coefficient of the dimension weight of one dimension caused by the increment of the difference between the dimensional feature of one dimension of the second feature data and the dimensional threshold of the dimension is greater than the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimensional weights, which can make the coefficient of the dimension weight of the dimension tend to the first value or the second value more quickly, thereby achieving the effect of selecting and not selecting the dimension weight.

[0009] In some possible designs, before training the risk control strategy model, a first filter condition input by a user may be received, and the first original data may be filtered using the first filter condition to obtain the first characteristic data. Therefore, when the risk control strategy model is used to generate a risk control strategy, the second original data may still be filtered using the first filter condition to obtain the second characteristic data. After triggering the risk control strategy model to be retrained, a second filter condition input by a user is received, and the third original data may be filtered using the second filter condition to obtain the third characteristic data. Among them, at least one characteristic dimension of the first characteristic data obtained by filtering using the first filter condition and the third characteristic data obtained by filtering using the second filter condition is different.

[0010] In some possible designs, the first filtering condition includes one or more of a population deviation index (PSI), a variance check, an information value (IV) value, and a feature correlation analysis.

[0011] In some possible designs, one or more of the following may be displayed on the interface:

[0012] Displaying one or more feature distributions in each feature dimension of the first feature data, and / or displaying one or more feature distributions in each feature dimension of the second feature data;

[0013] Display one or more of the feature threshold, the feature weight, the maximum value and the minimum value in the first feature data;

[0014] Display one or more of the precision-recall PR curve of the risk control strategy model, the precision curve under different score thresholds, the recall curve under different score thresholds, and the harmonic mean of the precision and recall under different score thresholds.

[0015] Among them, the score threshold is used to compare with the output of the risk control strategy model to distinguish whether the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is greater than the score threshold, the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is less than or equal to the score threshold, the input of the risk control strategy model is not a fraud attack.

[0016] In a second aspect, a risk control strategy model retraining trigger device is provided. The device includes an acquisition unit, a trigger unit, and a processing unit. The acquisition unit is used to acquire a first feature distribution of first feature data and a second feature distribution of second feature data. The trigger unit is used to trigger the risk control strategy model to retrain using third feature data when a distribution drift occurs between the second feature distribution and the first feature distribution. The processing unit is used to input the second feature data into the risk control strategy model for processing when no distribution drift occurs between the second feature distribution and the first feature distribution. The first feature data is sample data during the training of the risk control strategy model. The second feature data is current feature data. The first feature data includes multiple feature dimensions, the second feature data includes multiple feature dimensions, and the first feature data and the second feature data have the same dimension. The third feature data dimension includes multiple feature dimensions, and at least one feature dimension of the third feature data and the first feature data is different. The first feature data is feature data obtained by processing transaction data. For example, taking the transaction amount as an example, the transaction amount can be averaged to obtain a feature. When more processing is performed on the transaction amount, for example, variance processing, median processing, maximum processing, minimum processing, etc., more features of different dimensions can be obtained. The same is true for the second feature data and the third feature data.

[0017] In some possible designs, the parameters of the risk control strategy model include feature thresholds and feature weights. The feature thresholds and feature weights are obtained by training the risk control strategy model using the first feature data. The feature thresholds include multiple dimensional thresholds with the same number as the dimension of the first feature data. The feature weights include multiple dimensional weights with the same number as the dimension of the first feature data. That is, when the dimension of the first feature data is n, the dimension of the feature thresholds and the feature weights is also n. n is an integer greater than zero.

[0018] The processing unit is also used to determine multiple differences between multiple dimensional features of the second feature data and multiple dimensional thresholds. Then, the processing unit determines the coefficients of multiple dimensional weights according to the multiple differences. Specifically, when the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is larger, the coefficient of the dimensional weight of the dimension determined by the processing unit is closer to the first value; when the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is smaller, the coefficient of the dimensional weight of the dimension is closer to the second value. Afterwards, the processing unit determines whether the multiple dimensional weights are selected according to the coefficients of the multiple dimensional weights. Taking one of the dimensions as an example, when the coefficient of the dimensional weight of one of the dimensions is greater than the selected threshold, the processing unit determines that the dimensional weight of the dimension is selected; when the coefficient of the dimensional weight of one of the dimensions is less than or equal to the selected threshold, the processing unit determines that the dimensional weight of the dimension is not selected. Similarly, the same applies to other dimensions. Finally, the processing unit determines the output value of the risk control strategy model according to the sum of the selected dimensional weights. Specifically, when the sum of the selected dimension weights is larger, the processing unit determines that the output value of the risk control strategy model is closer to the first value; when the sum of the selected dimension weights is smaller, the processing unit determines that the output value of the risk control strategy model is closer to the second value. The increment of the coefficient of the dimension weight of one dimension of the second feature data caused by the increment of the difference between the dimension feature of the dimension of the second feature data and the dimension threshold of the dimension is α times the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimension weights, where α is a real number greater than 1.

[0019] In some possible designs, before training the risk control strategy model, the receiving unit may receive a first filtering condition input by the user, and the filtering unit uses the first filtering condition to filter the first original data to obtain the first characteristic data. Therefore, when using the risk control strategy model to generate a risk control strategy, the filtering unit still uses the first filtering condition to filter the second original data to obtain the second characteristic data. After triggering the risk control strategy model to be retrained, the receiving unit may receive a second filtering condition input by the user, and the filtering unit uses the second filtering condition to filter the third original data to obtain the third characteristic data. Among them, at least one characteristic dimension of the first characteristic data obtained by filtering using the first filtering condition is different from that of the third characteristic data obtained by filtering using the second filtering condition.

[0020] In some possible designs, the first filtering condition includes one or more of a population deviation index (PSI), a variance check, an information value (IV) value, and a feature correlation analysis.

[0021] In some possible designs, the device further includes a display unit, which is used to display one or more feature distributions in each feature dimension of the first feature data, and / or one or more feature distributions in each feature dimension of the second feature data. The display unit is used to display one or more of the feature threshold, the feature weight, the maximum value and the minimum value in the first feature data. The display unit is used to display one or more of the precision-recall PR curve of the risk control strategy model, the precision curve under different score thresholds, the recall curve under different score thresholds, and the harmonic mean of the precision and recall under different score thresholds. Among them, the score threshold is used to compare with the output of the risk control strategy model to distinguish whether the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is greater than the score threshold, the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is less than or equal to the score threshold, the input of the risk control strategy model is not a fraud attack.

[0022] In a third aspect, a computing device is provided, comprising a processor and a storage device, wherein the storage device stores instructions, and when the instructions are executed, the method described in any one of the first aspects is executed.

[0023] In a fourth aspect, a computing cluster is provided, comprising a plurality of computing devices, each computing device comprising a processor and a storage device, wherein the storage device stores instructions, and when the instructions are executed, the method described in any one of the first aspects is executed.

[0024] In a fifth aspect, a computer-readable storage medium is provided, comprising a plurality of instructions, wherein when the instructions are executed, the method as described in any one of the first aspect is performed.

[0025] In a sixth aspect, a computer program product is provided, comprising a plurality of instructions, wherein when the instructions are executed, the method as described in any one of the first aspect is performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background technology, the drawings required for use in the embodiments of the present application or the background technology will be described below.

[0027] Figure 1 It is a flowchart of a risk control strategy model training method provided by this application;

[0028] Figure 2 is a schematic diagram of an input interface provided by the present application;

[0029] Figure 3 It is a characteristic distribution diagram of a dimensional characteristic of a first characteristic data in one dimension provided by the present application;

[0030] Figure 4 It is a schematic diagram of the dimension weight, dimension threshold, maximum value of the dimension feature, and minimum value of the dimension feature of each dimension of the first feature data provided by the present application;

[0031] Figure 5 It is a schematic diagram of a precision-recall PR curve of a risk control strategy model provided in this application;

[0032] Figure 6 It is a flowchart of a risk control strategy model retraining triggering method provided by this application;

[0033] Figure 7 A schematic diagram of the structure of a risk control strategy model retraining trigger device provided in this application;

[0034] Figure 8 It is a structural schematic diagram of a computing device provided by this application. DETAILED DESCRIPTION

[0035] In a transaction scenario, a large amount of data is often generated, such as one or more of the following: transaction type, transaction amount, transaction time, transaction location, transaction participants, transaction goods / services, transaction status, transaction-related identification information, transaction fees, etc. Among them, the transaction type may be the type of transaction, such as purchase, sale, transfer, etc. The transaction amount may be the monetary amount involved in the transaction, such as the price of the purchased goods, the amount of the transfer, etc. The transaction time may be the specific time when the transaction occurs, including the date and time. The transaction location may be the specific location where the transaction occurs, such as an online platform, a physical store, an automated teller machine (ATM), etc. The transaction participants may be the parties involved in the transaction, such as the buyer, the seller, the transfer parties, etc. The transaction goods / services may be the specific goods or services involved in the transaction, such as the product name, the service item, etc. The transaction status may be the status information of the transaction, such as success, failure, pending, etc. The transaction fee may be the fee incurred during the transaction, such as handling fee, transportation fee, etc. The transaction-related identification information may be a unique identifier used to identify the transaction, such as a transaction number, an order number, etc. Transaction verification information can be relevant information used to verify the legitimacy of a transaction, such as identity authentication, payment password, etc. By processing the data, a large amount of feature data can be obtained. For example, taking the transaction amount as an example, the transaction amount can be averaged to obtain a feature. When the transaction amount is processed more, such as variance processing, median processing, maximum processing, minimum processing, etc., more features of different dimensions can be obtained. Similarly, other data, such as transaction type, transaction time, transaction location, transaction participants, transaction goods / services, transaction status, transaction-related identification information, transaction fees, etc., can be processed to obtain many features. Features of multiple different dimensions are combined together and can be called raw feature data. In other words, raw feature data includes features of multiple different feature dimensions. Users can collect some raw feature data in non-fraud attacks, label these raw feature data as normal samples; collect some raw feature data in fraud attacks, label these raw feature data as abnormal samples. Here, normal samples and target samples constitute the sample set.

[0036] See also Figure 1 , Figure 1 1 is a flow chart of a risk control strategy model training method provided by the present application. The risk control strategy training method is applied to a risk control strategy model training device. Figure 1 As shown, the risk control strategy model training method of this embodiment includes:

[0037] S101: Acquire a first filtering condition input by a user in an input interface, and filter first original feature data and corresponding tags according to the first filtering condition to obtain first feature data and corresponding tags.

[0038] In some possible embodiments, Figure 2 As shown, the input interface may include a basic information column and a sample time column. The basic information column is used to input information related to the training task. The sample time column is used to input relevant information of the sample data required for the training task. The basic information column may include: a task name label and a task name input box, a business label and a business input box. The task name input box is used to input the name of the task, the task name label is used to prompt the user to input the name of the task in the task name input box, the business input box is used to input the business to which the task belongs, and the business label is used to prompt the user to input the business to which the task belongs in the business input box. The sample time column includes a normal sample time label and a normal sample time input box, a target sample time label and a target sample time input box, a filter condition label and a filter condition input box, a label type label and a label type input box. The normal sample time input box is used to input the time period of the normal sample to obtain the normal sample in the time period, and the normal sample time label is used to prompt the user to input the time period of the normal sample in the normal sample time input box. The target sample time input box is used to input the time period of the target sample to obtain the target sample in the time period, and the target sample time label is used to prompt the user to input the time period of the target sample in the target sample time input box. The filter condition input box is used to input the filter condition of the sample, and the filter condition label is used to prompt the user to input the filter condition of the sample in the filter condition input box. The label type input box is used to input the type of label, and the label type label is used to prompt the user to input the type of label in the label type input box. In the above example, the filter condition, normal sample time, target sample time, and label type are set separately. In actual applications, one or more of the normal sample time, target sample time, and label type can also be used as part of the filter condition, which is not specifically limited here.

[0039] In some possible embodiments, the user can import the sample set into the risk control strategy model training device in advance. Then, the user can Figure 2In the corresponding input box of the input interface, the first task name, the first business, the first normal sample time, the first target sample time, the first filtering condition and the first label type are input. Among them, the first task name can be a combination of one or more of any characters, numbers, and words, and the first task name is unique and can distinguish different tasks. The first business can be selected from multiple businesses, and the multiple businesses can be pre-configured. The first normal sample time can be a continuous or discontinuous time period. The first target sample time can be a continuous or discontinuous time period. The first normal sample time and the first target sample time can be the same time period or different time periods. The first filtering condition can be used to filter normal samples and target samples in the first normal sample time and the first target sample time. For example, the first filtering condition can include one or more of the population stability index (PSI), variance check, information value (IV), and feature correlation analysis. Here, the first normal sample time and the first target sample time can also be included in the first filtering condition, which is not specifically limited here. The first label type can be any one of fraud attack and non-fraud attack, high-risk transaction and low-risk transaction, etc. Here, when the label type is fraudulent attack and non-fraudulent attack, the requirements are often higher than when the label is high-risk transaction and low-risk transaction. The former requires clear identification of fraudulent attack and non-fraudulent attack, while the latter only needs to judge the level of transaction risk. After obtaining the first filter condition entered by the user in the input interface, the sample set is filtered according to the first filter condition to obtain the first sample. In fact, the sample set includes a normal sample set and a target sample set. Here, the normal sample set and the target sample set are not distinguished. It is uniformly believed that the sample set includes the first original feature data and the label of the first original feature data. Therefore, the first sample includes the first feature data obtained by filtering the first original feature data, and the label of the first feature data.

[0040] S102: Input the first feature data and the corresponding label into the risk control strategy model for training to obtain a trained risk control strategy model.

[0041] In some possible embodiments, the parameters of the risk control strategy model may include feature thresholds and feature weights. Among them, the feature thresholds and feature weights are obtained by training the risk control strategy model using the first feature data, the feature thresholds include multiple dimension thresholds of the same number as the dimension of the first feature data, and the feature weights include multiple dimension weights of the same number as the dimension of the first feature data. For example, assuming that the first feature data is data of 5 feature dimensions {dimension feature 1, dimension feature 2, dimension feature 3, dimension feature 4, dimension feature 5}, then the feature threshold is also 5 dimensions {dimension threshold 1, dimension threshold 2, dimension threshold 3, dimension threshold 4, dimension threshold 5}, and the dimension of the feature weight is also 5 dimensions {dimension weight 1, dimension weight 2, dimension weight 3, dimension weight 4, dimension weight 5}. When the input of the risk control strategy model is the first feature data, the output of the risk control strategy model can be a value between the first value and the second value. For example, the first value can be 1 and the second value can be 0. In addition, a score threshold can be set for the output of the risk control strategy model. The score threshold is used to compare with the output of the risk control strategy model to distinguish whether the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is greater than the score threshold, the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is less than or equal to the score threshold, the input of the risk control strategy model is not a fraud attack.

[0042] The risk control strategy model needs to meet the following conditions: the risk control strategy model is differentiable everywhere to ensure that gradient updates can be performed. The greater the difference between the dimension feature of one dimension of the input of the risk control strategy model and the dimension threshold of the dimension, the closer the coefficient of the dimension weight of the dimension is to the first value; the smaller the difference between the value of one dimension of the input of the risk control strategy model and the dimension threshold of the dimension, the closer the coefficient of the dimension weight of the dimension is to the second value. When the coefficient of the dimension weight of one dimension is greater than the selected threshold, the dimension weight of the dimension is determined to be selected; when the coefficient of the dimension weight of the dimension is less than or equal to the selected threshold, the dimension weight of the dimension is determined not to be selected. When the sum of the selected dimension weights is greater, the output value of the risk control strategy model is determined to be closer to the first value; when the sum of the selected dimension weights is smaller, the output value of the risk control strategy model is determined to be closer to the second value. The increment of the coefficient of the dimension weight of the dimension caused by the increment of the difference between one of the dimensional features of the input of the risk control strategy model and the dimension threshold of the dimension is α times the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimension weights, where α is a real number greater than 1. By setting α, the coefficient of the dimension weight can be made to tend to the first value or the second value faster than the output of the risk control strategy model. In an ideal state, when the dimensional feature of one of the dimensions of the input of the risk control strategy model is greater than the corresponding dimensional threshold in the feature threshold, the coefficient of the dimension weight of the dimension in the feature weight is 1, and conversely, when the dimensional feature of one of the dimensions of the input of the risk control strategy model is less than or equal to the corresponding dimensional threshold in the feature threshold, the coefficient of the dimension weight of the dimension in the feature weight is 0. Therefore, when the coefficient of the dimension weight is 1, it means that the dimension weight is selected, and when the coefficient of the dimension weight is 0, it means that the dimension weight is not selected.

[0043] In a more specific embodiment, the risk control strategy model can be expressed as:

[0044] Y=σ2(σ1(X-Alpha)*W+b)

[0045] Here, Y is the output of the risk control strategy model, σ2() is the first differentiable function, and the first differentiable function is differentiable everywhere. For example, the activation function σ2 = 1 / (1+e -x ), σ1() is the second differentiable function, which is differentiable everywhere. For example, the variant of the activation function σ1=1 / (1+e -x*α ), α is a real number greater than 1, X is the input of the risk control strategy model, Alpha is the feature threshold, W is the feature weight, and b is the bias value.

[0046] In some possible embodiments, the risk control strategy model is trained by using the first feature data and the label corresponding to the first feature data until the convergence condition is met, and the feature threshold and feature weight can be determined. After the feature threshold and feature weight are determined, a trained risk control strategy model can be obtained. The process of training the risk control strategy model may include: initializing the risk control strategy model according to the structure and parameters of the risk control strategy model. The initialization of the risk control strategy model can be performed randomly or according to prior knowledge. The initial state of the risk control strategy model will affect the subsequent training process. Generally speaking, the convergence speed of initialization based on prior knowledge will be faster than the convergence speed of random initialization. After the initialization is completed, the first feature data is input into the risk control strategy model, and the prediction result is obtained by the forward propagation calculation of the risk control strategy model. Then, the prediction result of the risk control strategy model is compared with the true label to calculate the loss function. Among them, the loss function measures the gap between the prediction result of the risk control strategy model and the actual situation. Thereafter, according to the loss function, the gradient of the parameters of the risk control strategy model is calculated by the back propagation algorithm. Among them, the gradient represents the sensitivity of the loss function to the model parameters. Finally, use an optimization algorithm (such as gradient descent) to update the model's parameters according to the gradient of the parameters. The gradient of the parameter update will affect the convergence speed and performance of the risk control strategy model. Repeat the above steps until the preset stop condition is reached, such as reaching the maximum number of iterations or the loss function converges.

[0047] After the risk control strategy model training is completed, the user can view the training results of the risk control strategy model, for example, display one or more feature distributions of each dimension feature of the first feature data; display one or more of the feature threshold, feature weight, maximum and minimum values ​​of the dimension features of each dimension in the first feature data; display one or more of the precision-recall rate PR curve of the risk control strategy model, the precision curve under different score thresholds, the recall curve under different score thresholds, the harmonic mean of the precision and recall rate under different score thresholds, etc. The following will combine Figures 3 to 5 Provide a detailed description.

[0048] Figure 3 The feature distribution of the dimensional features of the first feature data in one dimension is shown. Figure 3As shown in the figure, the value of the dimension feature of this dimension is between 0 and 70, wherein the number of first feature data between the interval 0 and 10 is 300, the number of first feature data between the interval 10 and 20 is 425, the number of first feature data between the interval 20 and 30 is 160, the number of first feature data between the interval 30 and 40 is 342, the number of first feature data between the interval 40 and 50 is 278, the number of first feature data between the interval 50 and 60 is 193, and the number of first feature data between the interval 60 and 70 is 158. Here, Figure 3 Only the feature distribution of dimensional features on one dimension is shown. In actual applications, the feature distribution of dimensional features on other dimensions can also be displayed. When displaying, the feature distribution of dimensional features on multiple dimensions can be displayed in the same graph or in different graphs. In addition to using bar graphs, feature distribution can also be displayed in line graphs, scatter graphs, pie charts, box plots, heat maps, tree diagrams, hash charts, bubble charts, etc.

[0049] Figure 4 The dimension weight, dimension threshold, maximum value of the dimension feature, and minimum value of the dimension feature of each dimension of the first feature data are displayed. Among them, the dimension weight of each dimension can reflect the importance of each dimension feature. The larger the dimension weight, the greater the importance of the dimension feature. The smaller the dimension weight, the smaller the importance of the dimension feature. When the value of a certain dimension feature is greater than the feature threshold, the dimension feature is selected. When the value of a certain dimension feature is less than or equal to the feature threshold, the dimension feature is not selected. The maximum value of a certain dimension feature is the largest value of the first feature data in the dimension feature of this dimension. The minimum value of a certain dimension feature is the smallest value of the first feature data in the dimension feature of this dimension. Both the maximum value and the minimum value of the dimension feature can reflect the distribution characteristics of the dimensional features of the first feature data in this dimension to a certain extent. Assuming that the first feature data has a total of 5 dimensional features from dimensional feature 1 to dimensional feature 5, then we can get the following Figure 4As shown in the table, the dimension weight corresponding to the dimension feature 1 of the first feature data is 3.68649, the dimension threshold corresponding to the dimension feature 1 of the first feature data is 27.98126, the maximum value on the dimension feature 1 of the first feature data is 87.98737, and the maximum value on the dimension feature 1 of the first feature data is 2.77642; the dimension weight corresponding to the dimension feature 2 of the first feature data is 17.97647, the dimension threshold corresponding to the dimension feature 2 of the first feature data is 3.76307, the maximum value of the dimension feature on the dimension feature 2 of the first feature data is 13.27853, and the maximum value of the dimension feature on the dimension feature 2 of the first feature data is 0.65314; the dimension weight corresponding to the dimension feature 3 of the first feature data is 1.86539, the dimension threshold corresponding to the dimension feature 3 of the first feature data is 98.56745, and the dimension weight corresponding to the dimension feature 3 of the first feature data is 1.86539. The maximum value of the dimensional feature on the dimension feature 3 is 216.95632, and the maximum value of the dimensional feature on the dimension feature 3 of the first feature data is 16.45878; the dimensional weight corresponding to the dimensional feature 4 of the first feature data is 0.12417, the dimensional threshold corresponding to the dimensional feature 4 of the first feature data is 1621.79868, the maximum value of the dimensional feature on the dimensional feature 4 of the first feature data is 3228.09642, and the maximum value of the dimensional feature on the dimensional feature 4 of the first feature data is 586.74567; the dimensional weight corresponding to the dimensional feature 5 of the first feature data is 78.6438, the dimensional threshold corresponding to the dimensional feature 5 of the first feature data is 12.78753, the maximum value of the dimensional feature on the dimensional feature 5 of the first feature data is 30.86431, and the minimum value of the dimensional feature on the dimensional feature 5 of the first feature data is 2.03698. In the above method, only the feature weight, feature threshold, maximum value of the dimensional feature, and minimum value of the dimensional feature are used as examples for explanation. In actual applications, fewer or more aspects can be selected for display.

[0050] Figure 5 The precision-recall PR curve of the risk control strategy model is shown. The horizontal axis of the curve represents the precision, and the vertical axis represents the recall. In addition, the precision curves under different score thresholds, the recall curves under different score thresholds, the harmonic mean of the precision and recall under different score thresholds, etc. can also be displayed, which are not specifically limited here.

[0051] After the risk control strategy model is trained, if the user is not satisfied with the risk control strategy model, the risk control strategy model can be retrained. Figure 2 The input interface shown is used to re-enter the filtering condition, filter the first original feature data, obtain new feature data for re-training, and at least one dimension of the new feature data is different from that of the first feature data.

[0052] See also Figure 6 , Figure 6 1 is a flow chart of a risk control strategy model retraining triggering method provided by the present application. The risk control strategy training method is applied to a risk control strategy model using device. Figure 6 As shown, the risk control strategy model retraining method includes the following steps:

[0053] S201: Filter the second original feature data according to the first filtering condition to obtain second feature data.

[0054] In some possible embodiments, the second original feature data may be the current original feature data. Here, the first filtering condition in the risk control strategy model retraining trigger method may be Figure 1 The first filtering condition used in the risk control strategy model training method shown is the same. Figure 1 When the first filtering condition used in the risk control strategy model training method is the same, the first feature data used to train the risk control strategy model and the second feature data used for prediction using the risk control strategy model have the same feature dimension.

[0055] S202: Obtain a first characteristic distribution of the first characteristic data and a second characteristic distribution of the second characteristic data.

[0056] In some possible embodiments, the first characteristic distribution of the first characteristic data can help understand the distribution of the first characteristic data, such as central tendency, dispersion degree and morphology. When performing characteristic distribution analysis, some statistical indicators can be used to describe the distribution, such as mean, median, standard deviation, skewness and kurtosis. The mean may refer to the average of the values ​​of the first characteristic data, the median refers to the middle value after the values ​​of the first characteristic data are sorted by size, the standard deviation is an indicator to measure the degree of dispersion of the first characteristic data, the skewness is a measure of the degree of distribution skewness of the first characteristic data, and the kurtosis is an indicator to reflect the peak state of the distribution of the first characteristic data. The second characteristic distribution of the second characteristic data is similar to the first characteristic distribution of the first characteristic data, however, the statistical indicators used in the first characteristic distribution and the statistical indicators used in the second characteristic distribution can be the same, that is, when the first characteristic distribution uses the mean and standard deviation, the second characteristic distribution can also use the mean and standard deviation.

[0057] In some possible embodiments, the risk control strategy model may store the first feature data or the first feature distribution in the device. When the risk control strategy model stores the first feature data in the device, the first feature distribution may be calculated based on the first feature data stored in the device. When the risk control strategy model stores the first feature distribution, the first feature distribution may be calculated and stored in the device when the risk control strategy model is trained. The second feature distribution may be calculated based on the current second feature data.

[0058] S203: Determine whether the second characteristic distribution has generated distribution drift relative to the first characteristic distribution. If yes, proceed to step S204; if no, proceed to step S205.

[0059] In some possible embodiments, distribution drift can be used to judge the performance of the risk control strategy model. When the second feature distribution has a distribution drift relative to the first feature distribution, the performance of the risk control strategy model may decline because the rules learned by the risk control strategy model on the first feature data may no longer apply to the second feature data. When the second feature distribution has not had a distribution drift relative to the first feature distribution, the performance of the risk control strategy model may remain unchanged because the rules learned by the risk control strategy model on the first feature data are still applicable to the second feature data. Distribution drift may be caused by a variety of factors, such as changes in data collection methods, changes in data sources, changes in the environment, etc.

[0060] In some possible embodiments, it can be determined in the following ways whether the second feature distribution has generated a distribution drift relative to the first feature distribution. For example, by visualizing the first feature distribution of the first feature data and the second feature distribution of the second feature data, it can be intuitively observed whether there are obvious differences. For example, a scatter plot, a box plot, a histogram, and other visualization tools can be used to compare the distribution of the data; a statistical method can be used to test whether the first feature distribution of the first feature data and the second feature distribution of the second feature data are consistent, for example, a Kolmogorov-Smirnov test, a chi-square test, a t-test, etc.; the difference in the behavior of the classifier on the first feature data and the second feature data can be observed to determine whether a distribution drift has occurred. For example, the decision boundary, the error sample distribution, the classification probability distribution, etc. of the classifier can be compared. If the performance of the classifier on the second feature data is poor or inconsistent with the behavior on the first feature data, it may mean that the second feature distribution has generated a distribution drift relative to the first feature distribution; specific indicators or performance metrics can be monitored to determine whether a distribution drift has occurred. For example, the changes in the mean, variance, accuracy, recall and other indicators of the first feature data and the second feature data can be compared. If these indicators change significantly, it may mean that the second feature distribution has drifted relative to the first feature distribution.

[0061] S204: Triggering the risk control strategy model to be retrained using the third feature data.

[0062] In some possible embodiments, the risk control strategy model can be Figure 1 The risk control strategy model is obtained by training using the risk control strategy model training method shown.

[0063] In some possible embodiments, the third feature data may be obtained by filtering the third original feature data using the second filtering condition. The second filtering condition and the first filtering condition may be different. For example, the first filtering condition may include a PSI indicator, a variance check, etc., and the second filtering condition may include a variance check, an IV value, etc. Therefore, at least one feature dimension of the first feature data obtained by the first filtering condition and the third feature data obtained by the second filtering condition is different. The first original feature data and the third original feature data may be the same or different. When the first original feature data and the third original feature data are the same, because the first filtering condition and the second filtering condition adopted are different, the first feature data obtained by the first filtering condition and the third feature data obtained by the second filtering condition are still different in at least one feature dimension.

[0064] In some possible embodiments, the risk control strategy model training device and the risk control strategy model using device are different devices, that is, the training and use of the risk control strategy model are completed in different devices. At this time, the risk control strategy model using device needs to send a notification message to the risk control strategy model training device to trigger the risk control strategy model to be retrained using the third feature data. The risk control strategy model training device and the risk control strategy model using device can also be the same device, that is, the training and use of the risk control strategy model are completed in the same device. At this time, the risk control strategy model using device can directly trigger the risk control strategy model training, and there is no need to send a notification message to trigger the risk control strategy model to be retrained using the third feature data.

[0065] S205: Input the second characteristic data into the risk control strategy model for processing.

[0066] In some possible embodiments, the second feature data can be calculated through the forward propagation of the risk control strategy model to obtain an output result. When performing the forward propagation calculation, it is necessary to ensure that the second feature data meets the input requirements of the model, such as size, data type, etc., and perform preprocessing operations such as data scaling, normalization, and standardization when necessary to enable the second feature data to adapt to the risk control strategy model. When the amount of second feature data is relatively large, batch calculation can be considered, that is, processing multiple data at the same time to improve calculation efficiency.

[0067] In some possible embodiments, the process of inputting the second feature data into the risk control strategy model for processing includes: the risk control strategy model uses a device to determine multiple differences between multiple dimensional features of the second feature data and multiple dimensional thresholds. For example, assuming that the first feature data is data of 5 feature dimensions {dimensional feature 1, dimensional feature 2, dimensional feature 3, dimensional feature 4, dimensional feature 5}, and the feature threshold is also 5 dimensions {dimensional threshold 1, dimensional threshold 2, dimensional threshold 3, dimensional threshold 4, dimensional threshold 5}, then the multiple differences are (dimensional feature 1-dimensional threshold 1), (dimensional feature 2-dimensional threshold 2), (dimensional feature 3-dimensional threshold 3), (dimensional feature 4-dimensional threshold 4), (dimensional feature 5-dimensional threshold 5).

[0068] Then, the risk control strategy model uses a device to determine the coefficients of the multiple dimensional weights according to the multiple differences. Continuing with the above example, the coefficients of the multiple dimensional weights are: σ1 (dimensional feature 1-dimensional threshold 1), σ1 (dimensional feature 2-dimensional threshold 2), σ1 (dimensional feature 3-dimensional threshold 3), σ1 (dimensional feature 4-dimensional threshold 4), σ1 (dimensional feature 5-dimensional threshold 5). When the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is larger, the coefficient of the dimensional weight of the dimension is determined to be closer to the first value; when the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is smaller, the coefficient of the dimensional weight of the dimension is determined to be closer to the second value. That is, when the difference between dimensional feature 1 and dimensional threshold 1 is larger, σ1 (dimensional feature 1-dimensional threshold 1) is determined to be closer to the first value; when the difference between dimensional feature 1 and dimensional threshold 1 is smaller, σ1 (dimensional feature 1-dimensional threshold 1) is determined to be closer to the second value. Similarly, the same is true for σ1(dimensional feature 2-dimensional threshold 2), σ1(dimensional feature 3-dimensional threshold 3), σ1(dimensional feature 4-dimensional threshold 4), and σ1(dimensional feature 5-dimensional threshold 5). Here, σ1(dimensional feature 1-dimensional threshold 1), σ1(dimensional feature 2-dimensional threshold 2), σ1(dimensional feature 3-dimensional threshold 3), and σ1(dimensional feature 4-dimensional threshold 4) are independent of each other, that is, whether σ1(dimensional feature 1-dimensional threshold 1) tends to the first value or the second value does not affect any of σ1(dimensional feature 2-dimensional threshold 2), σ1(dimensional feature 3-dimensional threshold 3), σ1(dimensional feature 4-dimensional threshold 4), and σ1(dimensional feature 5-dimensional threshold 5).

[0069] Afterwards, the risk control strategy model uses the device to determine whether the multiple dimension weights are selected according to the coefficients of the multiple dimension weights. When the coefficient of the dimension weight of one dimension is greater than the selected threshold, it is determined that the dimension weight of the dimension is selected; when the coefficient of the dimension weight of one dimension is less than or equal to the selected threshold, it is determined that the dimension weight of the dimension is not selected. Continuing with the above example, the dimension weight corresponding to dimension feature 1 is dimension weight 1, the dimension weight corresponding to dimension feature 2 is dimension weight 2, the dimension weight corresponding to dimension feature 3 is dimension weight 3, the dimension weight corresponding to dimension feature 4 is dimension weight 4, and the dimension weight corresponding to dimension feature 5 is dimension weight 5. When the value of σ1 (dimensional feature 1-dimensional threshold 1) is greater than the selected threshold, it is determined that the dimension weight 1 of the dimension is selected; when the value of σ1 (dimensional feature 1-dimensional threshold 1) is less than or equal to the selected threshold, it is determined that the dimension weight 1 of the dimension is not selected. Similarly, whether dimension weights 2 to 5 are selected can be determined based on σ1 (dimensional feature 2-dimensional threshold 2), σ1 (dimensional feature 3-dimensional threshold 3), σ1 (dimensional feature 4-dimensional threshold 4), and σ1 (dimensional feature 5-dimensional threshold 5).

[0070] Finally, the risk control strategy model uses a device to determine the output value of the risk control strategy model based on the sum of the selected dimension weights. Here, when the sum of the selected dimension weights is larger, the output value of the risk control strategy model is determined to be closer to the first value; when the sum of the selected dimension weights is smaller, the output value of the risk control strategy model is determined to be closer to the second value. Continuing with the above example, assuming that dimension weight 2, dimension weight 3, and dimension weight 5 are selected, then the output value of the risk control strategy model can be σ2 (dimension weight 2 + dimension weight 3 + dimension weight 5 + b).

[0071] In some possible embodiments, the risk control strategy model may change during use. That is, the risk control strategy model may be used normally while continuing to be trained. Specifically, after the second feature data passes through the risk control strategy model, the corresponding label is obtained. Then, the risk control strategy model may continue to be trained using the second feature data and the label corresponding to the second feature data to improve the accuracy of the risk control strategy model.

[0072] In some possible embodiments, the risk control strategy model may not change during use, that is, the risk control strategy model will not continue to be trained during use, which can reduce the requirements on the computing power and storage capacity of the device using the risk control strategy model.

[0073] In addition, one or more feature distributions of each dimensional feature of the second feature data can be displayed separately, or one or more feature distributions of each dimensional feature of the first feature data and one or more feature distributions of each dimensional feature of the second feature data can be displayed simultaneously, so as to facilitate the user to compare the feature distribution of the first feature data with the feature distribution of the second feature data.

[0074] See also Figure 7 , Figure 7 Schematic diagram of a risk control strategy model retraining trigger device provided by this application. Figure 7 As shown, the risk control strategy model retraining triggering device of the present application can be applied to a risk control strategy model using device, including: an acquisition unit 110, a triggering unit 120 and a processing unit 130.

[0075] The acquisition unit 110 is used to acquire a first feature distribution of first feature data and a second feature distribution of second feature data, wherein the first feature data is sample data during risk control strategy model training, the second feature data is current feature data, the first feature data includes multiple feature dimensions, and the first feature data and the second feature data have the same dimension.

[0076] The trigger unit 120 is used to trigger the risk control strategy model to be retrained using third feature data when a distribution drift occurs between the second feature distribution and the first feature distribution, wherein at least one feature dimension of the third feature data is different from that of the first feature data.

[0077] The processing unit 130 is used for inputting the second characteristic data into the risk control strategy model for processing when there is no distribution drift between the second characteristic distribution and the first characteristic distribution.

[0078] The acquisition unit 110 may execute Figure 6 In the steps S201 and S202 of the risk control strategy model retraining triggering method shown in FIG. 1 , the trigger unit 120 may execute Figure 6 In step S204 of the risk control strategy model retraining triggering method shown in FIG. 1 , the processing unit 130 may execute Figure 6 Step S205 in the risk control strategy model retraining triggering method shown in FIG. 10A . When the risk control strategy model training device and the risk control strategy model using device are the same device, the acquisition unit 110 can also be used to perform Figure 1 In step S101 of the risk control strategy model training method shown in FIG. 1 , the processing unit 130 may be used to execute Figure 1Step S102 in the risk control strategy model training method shown.

[0079] See also Figure 8 , Figure 8 is a schematic diagram of the structure of a computing device provided by this application. Figure 8 As shown, the computing device may be the above-mentioned risk control strategy model training device or risk control strategy model using device, or a combination of the risk control strategy model training device and the risk control strategy model using device. The computing device 700 may include:

[0080] Processor 701, storage unit 702, storage medium 703 and communication interface 704, wherein processor 701, storage unit 702, storage medium 703 and communication interface 704 communicate via bus 705, and also communicate via other means such as wireless transmission.

[0081] The processor 701 is composed of a plurality of general-purpose processors, such as a CPU. The hardware chip is an application-specific integrated circuit (ASIC), a programmable logic device (PLD) or a combination thereof. The PLD is a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), a data processing unit (DPU), a system on chip (SoC) or any combination thereof. The processor 701 executes various types of digital storage instructions, such as software or firmware programs stored in the storage unit 702, which enables the computing device 700 to provide a wide variety of services.

[0082] In a specific implementation, as an embodiment, the processor 701 includes one or more CPUs, such as Figure 8 CPU0 and CPU1 are shown in the figure.

[0083] In a specific implementation, as an embodiment, the computing device 700 also includes multiple processors, such as Figure 81 and 10. Each of these processors may be a single-CPU or a multi-CPU. A processor herein refers to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0084] The storage unit 702 is used to store program codes, and is controlled by the processor 701 to execute the above Figure 1-Figure 5 The processing steps of the risk control strategy model training method in any embodiment, or, Figure 6 The processing steps of the risk control strategy model retraining triggering method shown in the figure. The program code includes one or more software units.

[0085] In the case where the computing device 700 is a risk control strategy model using device, the risk control strategy model using device may include Figure 7 The risk control strategy model retraining trigger device shown in the figure includes: an acquisition unit 110, a trigger unit 120 and a processing unit 130. The acquisition unit 110 can execute Figure 6 In the steps S201 and S202 of the risk control strategy model retraining triggering method shown in FIG. 1 , the trigger unit 120 may execute Figure 6 In step S204 of the risk control strategy model retraining triggering method shown in FIG. 1 , the processing unit 130 may execute Figure 6 Step S205 in the risk control strategy model retraining triggering method shown.

[0086] In the case where the computing device 700 is a unified body of a risk control strategy model using device and a risk control strategy model training device, the risk control strategy model using device may include Figure 7 The risk control strategy model retraining trigger device shown in the figure includes: an acquisition unit 110, a trigger unit 120 and a processing unit 130. The acquisition unit 110 can execute Figure 6 Steps S201 and S202 in the risk control strategy model retraining triggering method shown in FIG. Figure 1 In step S101 of the risk control strategy model training method shown in FIG. 1 , the trigger unit 120 may execute Figure 6 In step S204 of the risk control strategy model retraining triggering method shown in FIG. 1 , the processing unit 130 may execute Figure 6 Step S205 in the risk control strategy model retraining triggering method shown in FIG. Figure 1 Step S102 in the risk control strategy model training method shown.

[0087] The storage unit 702 includes a read-only memory and a random access memory, and provides instructions and data to the processor 701. The storage unit 702 also includes a non-volatile random access memory. The storage unit 702 is a volatile memory or a non-volatile memory, or includes both volatile and non-volatile memories. Among them, the non-volatile memory is a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory is a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are used, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM) and direct rambus RAM (DR RAM). It can also be a hard disk, a universal serial bus (USB), a flash memory, a secure digital memory card (SD card), a memory stick, etc. The hard disk is a hard disk drive (HDD), a solid state disk (SSD), a mechanical hard disk (HDD), etc., and this application does not make specific limitations.

[0088] The storage medium 703 is a carrier for storing data, such as a hard disk, a USB flash drive (universal serialbus), a flash memory, a secure digital memory card (SD card), a memory stick, etc. The hard disk can be a hard disk drive (HDD), a solid state disk (SSD), a mechanical hard disk (HDD), etc., and this application does not make specific limitations.

[0089] The communication interface 704 is a wired interface (such as an Ethernet interface), an internal interface (such as a high-speed serial computer expansion bus (Peripheral Component Interconnect express, PCIe) bus interface), a wired interface (such as an Ethernet interface) or a wireless interface (such as a cellular network interface or a wireless local area network interface) for communicating with other servers or units.

[0090] The bus 705 is a Peripheral Component Interconnect Express (PCIe) bus, an extended industry standard architecture (EISA) bus, a unified bus (Ubus or UB), a compute express link (CXL), a cache coherent interconnect for accelerators (CCIX), etc. The bus 705 is divided into an address bus, a data bus, a control bus, etc.

[0091] The bus 705 includes not only a data bus but also a power bus, a control bus, and a status signal bus, etc. However, for the sake of clarity, various buses are labeled as bus 705 in the figure.

[0092] Need to explain, Figure 7 This is only a possible implementation of the embodiment of the present application. In practical applications, the computing device 700 may also include more or fewer components, which is not limited here. For the contents not shown or described in the embodiment of the present application, please refer to the aforementioned Figure 1-Figure 6 The relevant explanations in the embodiments will not be repeated here.

[0093] The embodiment of the present application further provides a computing device cluster, which includes at least one computing device 700. The storage unit 702 in one or more computing devices 700 in the computing device cluster may store the same or different instructions for executing the risk control strategy model training method and / or the risk control strategy model retraining triggering method.

[0094] The embodiment of the present application also provides a computer program product including instructions. The computer program product may be a software or program product including instructions that can be run on a computing device or stored in any available medium. When the computer program product is run on at least one computing device, the at least one computing device executes the risk control strategy model training method and / or the risk control strategy model retraining triggering method.

[0095] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available medium can be a magnetic medium, (e.g., a floppy disk, a storage disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium.

Claims

1. A risk control strategy model retraining triggering method, characterized in that: include: Obtaining a first feature distribution of first feature data and a second feature distribution of second feature data, wherein the first feature data is sample data during risk control strategy model training, the second feature data is current feature data, the first feature data includes multiple feature dimensions, and the first feature data and the second feature data have the same dimension; In the event that a distribution drift occurs between the second feature distribution and the first feature distribution, the risk control strategy model is triggered to be retrained using third feature data, wherein at least one feature dimension of the third feature data is different from that of the first feature data.

2. The method according to claim 1, characterized in that In a case where there is no distribution drift between the second characteristic distribution and the first characteristic distribution, the second characteristic data is input into the risk control strategy model for processing.

3. The method according to claim 2, characterized in that The parameters of the risk control strategy model include a feature threshold and a feature weight, wherein the feature threshold and the feature weight are obtained by training the risk control strategy model using the first feature data, the feature threshold includes a plurality of dimension thresholds having the same number as the dimension of the first feature data, and the feature weight includes a plurality of dimension weights having the same number as the dimension of the first feature data, and the method further includes: Determine a plurality of differences between a plurality of dimensional features of the second feature data and a plurality of dimensional thresholds respectively; Determine coefficients of multiple dimension weights respectively according to the multiple differences; Determine whether the multiple dimensional weights are selected according to the coefficients of the multiple dimensional weights; The output value of the risk control strategy model is determined according to the sum of the selected dimension weights.

4. The method according to claim 3, characterized in that: Determining coefficients of multiple dimensional weights respectively according to the multiple differences includes: When the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension is larger, the coefficient of determining the dimensional weight of the dimension is closer to the first value; The smaller the difference between one of the dimensional features of the second feature data and the dimensional threshold of the dimension, the closer the coefficient for determining the dimensional weight of the dimension is to the second value.

5. The method according to claim 3 or 4, characterized in that: Determining whether the plurality of dimensional weights are selected according to the coefficients of the plurality of dimensional weights respectively includes: When the coefficient of the dimension weight of one dimension is greater than the selection threshold, it is determined that the dimension weight of the dimension is selected; When the coefficient of the dimension weight of one of the dimensions is less than or equal to the selection threshold, it is determined that the dimension weight of the dimension is not selected.

6. The method according to any one of claims 3 to 5, characterized in that: The output value of the risk control strategy model is determined according to the sum of the selected dimension weights, including: When the sum of the selected dimension weights is larger, it is determined that the output value of the risk control strategy model is closer to the first value; When the sum of the selected dimension weights is smaller, it is determined that the output value of the risk control strategy model is closer to the second value.

7. The method according to any one of claims 3 to 6, characterized in that: The increment of the coefficient of the dimension weight of one dimension of the second feature data caused by the increment of the difference between the dimensional feature of the dimension and the dimensional threshold of the dimension is α times the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimensional weights, where α is a real number greater than 1.

8. The method according to any one of claims 1 to 7, characterized in that: The first characteristic data is obtained by filtering the first original data using the first filtering condition, and the second characteristic data is obtained by filtering the second original data using the first filtering condition; Before triggering the risk control strategy model to be retrained using the third feature data, the method includes: receiving a second filtering condition input by a user; The third original data is filtered using the second filtering condition to obtain third characteristic data, wherein at least one characteristic dimension of the first characteristic data obtained by filtering using the first filtering condition and the third characteristic data obtained by filtering using the second filtering condition is different.

9. The method according to claim 8, characterized in that The first filtering condition includes one or more of a population deviation index PSI index, a variance check, an information value IV value, and a feature correlation analysis.

10. The method according to any one of claims 1 to 9, characterized in that: The method further comprises: Display one or more feature distributions in each feature dimension of the first feature data, and / or display one or more feature distributions in each feature dimension of the second feature data.

11. The method according to any one of claims 1 to 10, characterized in that: The method further comprises: One or more of the feature threshold, the feature weight, the maximum value and the minimum value in the first feature data are displayed.

12. The method according to any one of claims 1 to 11, characterized in that: The method further comprises: Display one or more of the precision-recall PR curve of the risk control strategy model, the precision curve under different score thresholds, the recall curve under different score thresholds, and the harmonic mean of the precision and recall under different score thresholds, wherein the score threshold is used to compare with the output of the risk control strategy model to distinguish whether the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is greater than the score threshold, the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is less than or equal to the score threshold, the input of the risk control strategy model is not a fraud attack.

13. A risk control strategy model retraining trigger device, characterized in that: include: an acquisition unit, configured to acquire a first feature distribution of first feature data and a second feature distribution of second feature data, wherein the first feature data is sample data during risk control strategy model training, the second feature data is current feature data, the first feature data includes multiple feature dimensions, and the first feature data and the second feature data have the same dimension; A trigger unit is used to trigger the risk control strategy model to retrain using third feature data when a distribution drift occurs between the second feature distribution and the first feature distribution, wherein at least one feature dimension of the third feature data is different from that of the first feature data.

14. The device according to claim 13, characterized in that The device also includes a processing unit, The processing unit is used for inputting the second characteristic data into the risk control strategy model for processing when no distribution drift occurs between the second characteristic distribution and the first characteristic distribution.

15. The device according to claim 14, characterized in that The parameters of the risk control strategy model include feature thresholds and feature weights, wherein the feature thresholds and the feature weights are obtained by training the risk control strategy model using the first feature data, the feature thresholds include multiple dimension thresholds with the same number as the dimension of the first feature data, and the feature weights include multiple dimension weights with the same number as the dimension of the first feature data, The processing unit is also used to determine multiple differences between multiple dimensional features of the second feature data and multiple dimensional thresholds, determine coefficients of multiple dimensional weights according to the multiple differences, determine whether the multiple dimensional weights are selected according to the coefficients of the multiple dimensional weights, and determine the output value of the risk control strategy model according to the sum of the selected dimensional weights.

16. The device according to claim 15, characterized in that The processing unit is also used to determine that the coefficient of the dimension weight of the dimension is closer to the first value when the difference between one of the dimensional features of the second feature data and the dimension threshold of the dimension is larger, and to determine that the coefficient of the dimension weight of the dimension is closer to the second value when the difference between one of the dimensional features of the second feature data and the dimension threshold of the dimension is smaller.

17. The device according to claim 14 or 15, characterized in that The processing unit is also used to determine that the dimension weight of one of the dimensions is selected when the coefficient of the dimension weight of the dimension is greater than the selection threshold; or to determine that the dimension weight of one of the dimensions is not selected when the coefficient of the dimension weight of the dimension is less than or equal to the selection threshold.

18. The device according to any one of claims 14 to 17, characterized in that The processing unit is also used to determine that the output value of the risk control strategy model is closer to the first value when the sum of the selected dimension weights is larger; or, when the sum of the selected dimension weights is smaller, the output value of the risk control strategy model is closer to the second value.

19. The device according to any one of claims 14 to 18, characterized in that The increment of the coefficient of the dimension weight of one dimension of the second feature data caused by the increment of the difference between the dimensional feature of the dimension and the dimensional threshold of the dimension is α times the increment of the output of the risk control strategy model caused by the increment of the sum of the selected dimensional weights, where α is a real number greater than 1.

20. The device according to any one of claims 13 to 19, characterized in that The first characteristic data is obtained by filtering the first original data using the first filtering condition, and the second characteristic data is obtained by filtering the second original data using the first filtering condition; the device further includes: A receiving unit, configured to receive a second filtering condition input by a user; A filtering unit is used to filter the third original data using the second filtering condition to obtain third feature data, wherein at least one feature dimension of the first feature data obtained by filtering using the first filtering condition is different from that of the third feature data obtained by filtering using the second filtering condition.

21. The device according to claim 20, characterized in that The first filtering condition includes one or more of a population deviation index PSI index, a variance check, an information value IV value, and a feature correlation analysis.

22. The device according to any one of claims 13 to 21, characterized in that The device further comprises a display unit, The display unit is used to display one or more feature distributions in each feature dimension of the first feature data, and / or to display one or more feature distributions in each feature dimension of the second feature data.

23. The device according to any one of claims 13 to 22, characterized in that The device further comprises a display unit, One or more of the feature threshold, the feature weight, the maximum value and the minimum value in the first feature data are displayed.

24. The device according to any one of claims 13 to 23, characterized in that The device further comprises a display unit, Display one or more of the precision-recall PR curve of the risk control strategy model, the precision curve under different score thresholds, the recall curve under different score thresholds, and the harmonic mean of the precision and recall under different score thresholds, wherein the score threshold is used to compare with the output of the risk control strategy model to distinguish whether the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is greater than the score threshold, the input of the risk control strategy model is a fraud attack. When the output of the risk control strategy model is less than or equal to the score threshold, the input of the risk control strategy model is not a fraud attack.

25. A computing device, characterized in that The method comprises a processor and a storage device, wherein the storage device stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 11 is executed.

26. A computing cluster, characterized in that: The method comprises a plurality of computing devices, each of which comprises a processor and a storage device, wherein the storage device stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 11 is executed.

27. A computer-readable storage medium, characterized in that: The method comprises a plurality of instructions, and when the instructions are executed, the method according to any one of claims 1 to 11 is performed.

28. A computer program product, characterized in that The method comprises a plurality of instructions, and when the instructions are executed, the method according to any one of claims 1 to 11 is performed.

Images