CN-117196018-B - Service prediction optimization method, system, equipment and storage medium

CN117196018BCN 117196018 BCN117196018 BCN 117196018BCN-117196018-B

Abstract

The application discloses a service prediction optimization method, a device, equipment and a storage medium, wherein the method comprises the steps of performing federal training on a preset gradient lifting tree model to be trained with at least one second terminal equipment based on a first parameter value of a first super parameter set and local service sample data, and evaluating multiple targets corresponding to the first super parameter set; the multi-objective system comprises at least two objectives of system training overhead, model precision loss, model privacy leakage degree and model interpretability degree, optimizes a first super-parameter set based on an evaluation result of the multi-objective system, and/or optimizes the multi-objective system and/or performs federal iterative training, determines a target service prediction model based on the evaluation result reaching a target super-parameter set of a preset evaluation standard, and inputs service data to be processed into the target service prediction model to perform prediction processing on the service data to be processed to obtain a service prediction result. The application aims to improve the data processing effect.

Inventors

REN ZIYAO
KANG YAN
FAN LIXIN
TONG YONGXIN
Gu hanlin

Assignees

深圳前海微众银行股份有限公司
北京航空航天大学

Dates

Publication Date: 20260505
Application Date: 20230719

Claims (9)

1. A method for traffic prediction optimization, applied to a first terminal device, the method comprising: Based on a first parameter value of a first super parameter set and local service sample data, performing federal training on a preset gradient lifting tree model to be trained with at least one second terminal device, and evaluating multiple targets corresponding to the first super parameter set, wherein the multiple targets comprise at least two targets of system training overhead, model precision loss, model privacy leakage degree and model interpretability degree; Optimizing the first super-parameter set based on the multi-objective evaluation result, and/or optimizing the multi-objective and/or performing federal iterative training, and determining a target service prediction model based on the evaluation result reaching a target super-parameter set corresponding to a corresponding preset evaluation standard, wherein the target service prediction model is used for predicting the service data to be processed when the service data to be processed is received, so as to obtain a service prediction result; Before the step of performing federal training on the preset gradient lifting tree model to be trained by at least one second terminal device, the method comprises the following steps: The step of determining the target service prediction model based on the target super parameter group corresponding to the corresponding preset evaluation standard reached by the evaluation result comprises the following steps: estimating and obtaining an evaluation value of multiple targets in a target model process through a trained proxy model, wherein the multiple targets respectively correspond to one Gaussian process model GP, optimizing the corresponding Gaussian process model GP by using preset Bayesian optimization until a model for accurately predicting the corresponding multiple target evaluation value is obtained, and setting the model as a proxy model; Determining the first hyper-parameter set through the trained pareto front model and preference characteristics of the corresponding user; and if the evaluation value based on the multiple targets does not reach the corresponding preset evaluation standard, optimizing the first super parameter set to generate a new second super parameter set, wherein the step of generating the new second super parameter set comprises the following steps: Determining the commonly corresponding supersvolume based on the multi-objective evaluation value; When the super volume is smaller than a preset super volume threshold, determining that the corresponding preset evaluation standard is not reached, and if the pareto optimal solution set corresponding to the super parameter group is not converged; If the corresponding preset evaluation standard is not met, a new second super-parameter set is generated based on the first super-parameter set.
2. The method for optimizing business prediction according to claim 1, wherein the step of optimizing the first superparameter group based on the multi-objective evaluation result, performing federal iterative training, and determining the objective business prediction model based on the evaluation result reaching the objective superparameter group corresponding to the preset evaluation standard comprises: Based on the multi-objective evaluation values, if the corresponding preset evaluation standard is not met, optimizing the first super-parameter set to generate a new second super-parameter set; And iteratively executing the step of carrying out federal training on the preset gradient lifting tree model to be trained together with other second terminal equipment on the basis of the second hyper-parameter set and the local sample service data until the corresponding multi-objective evaluation value reaches the corresponding preset evaluation standard, and determining the objective service prediction model on the basis of the evaluation result reaching the objective hyper-parameter set corresponding to the corresponding preset evaluation standard.
3. The method for optimizing service prediction according to claim 2, wherein the step of iteratively executing the step of continuing federal training on the gradient boost tree model to be trained together with other second terminal devices based on the second hyper-parameter set and the local sample service data until the evaluation values of the respective multi-objective reach the corresponding preset evaluation criteria, and determining the target service prediction model based on the evaluation result reaching the target hyper-parameter set corresponding to the corresponding preset evaluation criteria, includes: Based on the second hyper-parameter group and the local sample service data, jointly carrying out federal training on a preset gradient lifting tree model to be trained with other second terminal equipment, and determining whether the node depth of a decision tree constructed in the federal training process reaches a preset maximum depth; if not, determining whether the target purity of the node which does not reach the preset maximum depth is greater than a preset purity threshold, wherein the purity is used for representing the ratio of the number of samples of a preset category in the node to the total number of samples of the node; If the purity is larger than or equal to a preset purity threshold, calculating statistical information based on the first gradient corresponding to the local sample, wherein if the purity is smaller than the preset purity threshold, calculating the statistical information together with the second gradients corresponding to the corresponding samples of other second terminal equipment; Determining a splitting point for splitting the node which does not reach the preset maximum depth based on the statistical information, and returning to the step of determining whether the node depth of the decision tree constructed in the federal training process reaches the preset maximum depth or not until the node depths of the decision tree reach the preset maximum depth; if the depths of the decision tree nodes reach the preset maximum depth, determining weights of the corresponding leaf nodes, determining whether the evaluation values of the corresponding multi-targets reach the corresponding preset evaluation standards based on the gradient lifting tree model after the weights of the leaf nodes are determined, iteratively executing the step of continuously performing federal training on the gradient lifting tree model to be trained together with other second terminal equipment until the evaluation values of the corresponding multi-targets reach the corresponding preset evaluation standards, and determining a target service prediction model based on the evaluation results to reach target super-parameter groups corresponding to the corresponding preset evaluation standards.
4. The method for optimizing service prediction according to claim 3, wherein the step of performing federal training on the gradient boost tree model to be trained together with other second terminal devices based on the second hyper-parameter set and local sample service data, and determining whether the node depth of the decision tree constructed in the federal training process reaches the preset maximum depth comprises: Training a local preset gradient lifting tree model to be trained based on the second hyper-parameter set and local sample service data so as to reduce residual errors between a corresponding predicted value of a local sample and a sample label; If the local training meets the preset local training completion condition, carrying out federal training on a preset gradient lifting tree model to be trained together with other second terminal equipment, and determining whether the node depth of a decision tree constructed in the federal training process reaches a preset maximum depth.
5. The traffic prediction optimization method according to claim 1, wherein the multi-objective includes a plurality of objects to be optimized, and the manner of determining the evaluation value of the multi-objective includes at least one of: if the target to be optimized is training expenditure, determining the time occupied by the operation of the preset homomorphic encryption in the model training process and the corresponding operation times, and determining an evaluation value of the training expenditure in the target model process based on the time and the corresponding operation times; If the target to be optimized is model precision loss, determining and obtaining an evaluation value of model precision loss in the process of the target model based on the target model and a corresponding test data set; If the target to be optimized is the model privacy disclosure degree, clustering samples based on the sample similarity of the target model, and determining and obtaining an evaluation value of the model privacy disclosure degree in the target model process based on the label of the sample inferred after clustering and the label of the corresponding sample; And if the target to be optimized is the model interpretable degree, determining an evaluation value of the model interpretable degree in the process of obtaining the target model based on the number of leaf nodes of the decision tree in the target model.
6. The traffic prediction optimization method according to claim 2, wherein the step of optimizing the first super parameter set to generate a new second super parameter set includes: determining the dominance relation of each first super-parameter group, and non-dominance sorting the first super-parameter groups based on the dominance relation; Determining the distance between each first super-parameter set and the adjacent first super-parameter set; Selecting a next generation first superparameter group based on the non-dominant ordering and the distance; And performing preset crossover and mutation operation on the next generation first super parameter set to obtain a new second super parameter set.
7. A traffic prediction optimization device, applied to a first terminal device, the device comprising: the training module is used for carrying out federal training on a preset gradient lifting tree model to be trained with at least one second terminal device based on a first parameter value of a first super parameter set and local service sample data, and evaluating multiple targets corresponding to the first super parameter set, wherein the multiple targets comprise at least two targets of system training overhead, model precision loss, model privacy leakage degree and model interpretability degree; The optimizing module is used for optimizing the first super-parameter set based on the multi-objective evaluation result, optimizing the multi-objective and/or performing federal iteration training, and determining a target service prediction model based on the evaluation result reaching a target super-parameter set corresponding to a corresponding preset evaluation standard, wherein the target service prediction model is used for predicting the service data to be processed when the service data to be processed is received, so as to obtain a service prediction result; The service prediction optimization device is used for realizing: Determining the first hyper-parameter set through the trained pareto front model and preference characteristics of the corresponding user; estimating and obtaining an evaluation value of multiple targets in a target model process through a trained proxy model, wherein the multiple targets respectively correspond to one Gaussian process model GP, optimizing the corresponding Gaussian process model GP by using preset Bayesian optimization until a model for accurately predicting the corresponding multiple target evaluation value is obtained, and setting the model as a proxy model; Determining the commonly corresponding supersvolume based on the multi-objective evaluation value; When the super volume is smaller than a preset super volume threshold, determining that the corresponding preset evaluation standard is not reached, and if the pareto optimal solution set corresponding to the super parameter group is not converged; If the corresponding preset evaluation standard is not met, a new second super-parameter set is generated based on the first super-parameter set.
8. A traffic prediction optimization device comprising a memory, a processor and a traffic prediction optimization program stored on the memory and executable on the processor, the processor implementing the steps of the traffic prediction optimization method of any one of claims 1 to 6 when the processor executes the traffic prediction optimization program.
9. A storage medium having stored thereon a traffic prediction optimization program which, when executed by a processor, implements the steps of the traffic prediction optimization method according to any one of claims 1 to 6.

Description

Service prediction optimization method, system, equipment and storage medium Technical Field The present application relates to the field of communications computers, and in particular, to a method, a system, an apparatus, and a storage medium for optimizing service prediction. Background In the model training process, how to protect personal data, especially in sensitive industries such as finance, medical treatment and the like, so that people can enjoy convenience and protect privacy, and the method is particularly important. Federal learning can solve the defects of traditional centralized machine learning or deep learning in protecting data privacy, and simultaneously can train a global model by using distributed data. At present, when a model is used for processing data in sensitive industries such as finance, medical treatment and the like, the data processing effect is poor due to poor performance of the model. Disclosure of Invention In view of the above, the present application provides a method, a system, a device and a storage medium for optimizing service prediction, which aim to solve the technical problem of poor data processing effect in the prior art. The embodiment of the application provides a service prediction optimization method, which is applied to first terminal equipment, and comprises the following steps: Based on a first parameter value of a first super parameter set and local service sample data, performing federal training on a preset gradient lifting tree model to be trained with at least one second terminal device, and evaluating multiple targets corresponding to the first super parameter set, wherein the multiple targets comprise at least two targets of system training overhead, model precision loss, model privacy leakage degree and model interpretability degree; Optimizing the first super-parameter set based on the multi-objective evaluation result, and/or optimizing the multi-objective and/or performing federal iterative training, and determining a target service prediction model based on the evaluation result reaching a target super-parameter set corresponding to a corresponding preset evaluation standard, wherein the target service prediction model is used for predicting the service data to be processed when the service data to be processed is received, so as to obtain a service prediction result. In a possible implementation manner of the present application, the step of optimizing the first superparameter set based on the multi-objective evaluation result, performing federal iterative training, and determining a objective business prediction model based on the evaluation result reaching a objective superparameter set corresponding to a preset evaluation standard includes: Based on the multi-objective evaluation values, if the corresponding preset evaluation standard is not met, optimizing the first super-parameter set to generate a new second super-parameter set; And iteratively executing the step of carrying out federal training on the preset gradient lifting tree model to be trained together with other second terminal equipment on the basis of the second hyper-parameter set and the local sample service data until the corresponding multi-objective evaluation value reaches the corresponding preset evaluation standard, and determining the objective service prediction model on the basis of the evaluation result reaching the objective hyper-parameter set corresponding to the corresponding preset evaluation standard. In a possible implementation manner of the present application, the step of iteratively executing, together with other second terminal devices, the step of continuing federal training on the gradient tree model to be trained together until the evaluation value of the corresponding multi-objective reaches the corresponding preset evaluation criterion, and determining the objective business prediction model based on the evaluation result reaching the objective super-parameter set corresponding to the corresponding preset evaluation criterion, where the objective business prediction model includes: Based on the second hyper-parameter group and the local sample service data, jointly carrying out federal training on a preset gradient lifting tree model to be trained with other second terminal equipment, and determining whether the node depth of a decision tree constructed in the federal training process reaches a preset maximum depth; if not, determining whether the target purity of the node which does not reach the preset maximum depth is greater than a preset purity threshold, wherein the purity is used for representing the ratio of the number of samples of a preset category in the node to the total number of samples of the node; If the purity is larger than or equal to a preset purity threshold, calculating statistical information based on the first gradient corresponding to the local sample, wherein if the purity is smaller than the preset purity threshold, calc