CN-121981812-A - Risk prediction method and device

CN121981812ACN 121981812 ACN121981812 ACN 121981812ACN-121981812-A

Abstract

The application provides a risk prediction method which can be applied to the field of artificial intelligence. The risk prediction method comprises the steps of predicting an object to be predicted by using a trained target model to obtain a risk prediction result, wherein the trained target model is trained in a mode of respectively using sampling methods in a plurality of sampling-super-parameter combinations to sample a first data set, respectively training the target model based on sampling samples and super-parameters corresponding to the sampling methods to obtain a plurality of trained target models, determining a target sampling-super-parameter combination based on performance indexes of the plurality of trained target models, and retraining the target model according to the target sampling-super-parameter combination and a second data set, wherein the first data set and the second data set are unbalanced data sets, and the sample size of the first data set is smaller than that of the second data set. The application also provides a risk prediction device, equipment, a storage medium and a program product.

Inventors

WANG ZHEJIA
WEN ZHIXIANG

Assignees

中国工商银行股份有限公司

Dates

Publication Date: 20260505
Application Date: 20250620

Claims (11)

1. A risk prediction method, the method comprising: predicting an object to be predicted by using a trained target model to obtain a risk prediction result, wherein the trained target model is trained in the following manner: Sampling the first data set by using sampling methods in various sampling-super-parameter combinations, and training the target model based on sampling samples and super-parameters corresponding to the sampling methods to obtain a plurality of trained target models; Determining a target sampling-super-parameter combination based on performance indexes of each of the plurality of trained target models; Retraining the target model based on the target sample-super-parameter combination and a second data set, wherein the first data set and the second data set are both unbalanced data sets, and the first data set has a sample size that is smaller than the second data set.
2. The method of claim 1, wherein the plurality of sample-and-super-parameter combinations are obtained by: And selecting a sampling method suitable for the unbalanced data set and corresponding super parameters from a preset sampling method library by utilizing a meta-learning algorithm according to a screening strategy, wherein the screening strategy comprises at least one of a historical screening record, a preset screening strategy and a random searching strategy.
3. The method of claim 1, wherein training the target model based on the sampling samples and the super-parameters corresponding to each sampling method, respectively, comprises: And training the target model for an unbalanced data set with limited business scene data by combining cross verification and multi-fold verification based on sampling samples and super parameters corresponding to each sampling method.
4. A method according to claim 3, wherein the method of combining cross-validation with multi-fold validation comprises: Dividing sampling samples corresponding to sampling methods in the plurality of sampling-super-parameter combinations into a plurality of subsets which are not overlapped with each other, sequentially taking each subset as a verification set, taking other subsets as training sets, and training and verifying the target model for a plurality of times according to super-parameters corresponding to the same sampling sample; Calculating the average performance index of the target model after multiple times of verification; And randomly repeating the multiple training and verification operations of the target model for R times, and calculating the final average performance index of the target model, wherein R is an integer greater than or equal to 1.
5. The method of claim 1, wherein the determining a target sample-super-parameter combination based on performance metrics of each of the plurality of trained target models comprises: If the performance indexes of each of the plurality of trained target models have the performance indexes meeting the preset service requirements, taking a sampling-super-parameter combination corresponding to the performance indexes meeting the preset service requirements as the target sampling-super-parameter combination; And if the performance indexes of the plurality of trained target models do not meet the preset service requirements, adjusting a sampling strategy, acquiring a new sampling-super-parameter combination, and repeating training the target models and acquiring the performance indexes of the plurality of trained target models based on the new sampling-super-parameter combination until the performance indexes of the plurality of trained target models meet the preset service requirements.
6. The method of claim 5, wherein adjusting the sampling strategy to obtain a new sample-super-parameter combination comprises: And adjusting the sampling methods in the plurality of sampling-super-parameter combinations by utilizing a meta-learning algorithm through adjusting the sampling strategies, and selecting corresponding super-parameters according to the adjusted sampling methods to obtain a new sampling-super-parameter combination, wherein the sampling strategies comprise adjusting sampling proportions or changing the searching strategies of minority data in the unbalanced data set.
7. The method of claim 1, wherein after retraining the target model from the target sample-super-parameter combination and second data set, comprising: sequencing the performance indexes of the retrained target models according to the sequence from the big performance index to the small performance index; taking the sampling-super-parameter combination corresponding to the first performance index as a final sampling-super-parameter combination; and deploying the final sampling-super-parameter combination to a system according to a preset deployment strategy, wherein the preset deployment strategy comprises packaging or mirroring a configuration file containing the final sampling-super-parameter combination, and the system comprises a software and hardware platform for realizing the risk prediction method.
8. A risk prediction apparatus, the apparatus comprising: The risk prediction module is used for predicting the object to be predicted by using the trained target model to obtain a risk prediction result; The target model training module is used for training a target model to obtain a trained target model, wherein the trained target model is obtained by training a first data set by using sampling methods in various sampling-super-parameter combinations, training the target model based on sampling samples and super-parameters corresponding to the sampling methods to obtain a plurality of trained target models, determining a target sampling-super-parameter combination based on performance indexes of the plurality of trained target models, and retraining the target model according to the target sampling-super-parameter combination and a second data set, wherein the first data set and the second data set are unbalanced data sets, and the sample size of the first data set is smaller than that of the second data set.
9. An electronic device, comprising: One or more processors; a memory for storing one or more computer programs, Characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, realizes the steps of the method according to any one of claims 1-7.
11. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, realizes the steps of the method according to any one of claims 1-7.

Description

Risk prediction method and device Technical Field The present application relates to the field of artificial intelligence, and more particularly, to a risk prediction method, apparatus, device, medium, and program product. Background Under business scenes such as financial wind control and fraud detection, training data of the model often show serious unbalance, the number of few types of samples is far lower than that of the majority types of samples, and the serious unbalance of the training data can lead to lower prediction accuracy of the trained model. In order to solve the problem of unbalanced training data, various sampling methods are provided in the prior art, such as random undersampling, random oversampling, a synthetic minority oversampling technology, derivative methods thereof and the like, which can improve the recognition capability of a model to minority samples to a certain extent, but the effect is closely related to specific data distribution, data characteristic dimension and business scene, so that excellent performance is difficult to obtain in all scenes through a single method, sampling algorithms are obtained by means of a preset rule base in other methods, but the sampling algorithms recommended by the method are not deeply fused with the actual model training process, are screened only based on the existing rule base or historical experience, and cannot accurately evaluate the substantial influence of the screened sampling algorithms on the effect of a final model, so that the recommendation accuracy is insufficient or a large amount of manual intervention and trial-error are still needed in later stages. Disclosure of Invention In view of the foregoing, the present application provides a risk prediction method, apparatus, device, medium, and program product that improves the ability of a target model to identify minority class risks. According to the first aspect of the application, a risk prediction method is provided, which comprises the steps of predicting an object to be predicted by using a trained target model to obtain a risk prediction result, wherein the trained target model is trained in the following manner: The method comprises the steps of respectively using sampling methods in a plurality of sampling-super-parameter combinations to sample a first data set, respectively training a target model based on sampling samples and super-parameters corresponding to the sampling methods to obtain a plurality of trained target models, determining a target sampling-super-parameter combination based on performance indexes of the plurality of trained target models, and retraining the target model according to the target sampling-super-parameter combination and a second data set, wherein the first data set and the second data set are unbalanced data sets, and the sample size of the first data set is smaller than that of the second data set. According to an embodiment of the application, the plurality of sampling-super-parameter combinations are obtained by: And selecting a sampling method suitable for the unbalanced data set and corresponding super parameters from a preset sampling method library by utilizing a meta-learning algorithm according to a screening strategy, wherein the screening strategy comprises at least one of a historical screening record, a preset screening strategy and a random searching strategy. According to an embodiment of the present application, the training the target model based on the sampling samples and the super-parameters corresponding to the sampling methods includes: And training the target model for an unbalanced data set with limited business scene data by combining cross verification and multi-fold verification based on sampling samples and super parameters corresponding to each sampling method. According to an embodiment of the present application, the method for combining cross-validation with multi-fold validation includes: Dividing sampling samples corresponding to sampling methods in the plurality of sampling-super-parameter combinations into a plurality of subsets which are not overlapped with each other, sequentially taking each subset as a verification set, taking other subsets as training sets, and training and verifying the target model for a plurality of times according to super-parameters corresponding to the same sampling sample; Calculating the average performance index of the target model after multiple times of verification; And randomly repeating the multiple training and verification operations of the target model for R times, and calculating the final average performance index of the target model, wherein R is an integer greater than or equal to 1. According to an embodiment of the present application, the determining the target sampling-super-parameter combination based on the performance indexes of each of the plurality of trained target models includes: If the performance indexes of each of the plurality of trained target