Training method and device of service prediction system

Abstract

The embodiment of the invention provides a training method of a service prediction system. The service prediction system comprises a public model maintained by a server and a plurality of local models deployed in a plurality of user terminals. The method is executed by a server, and comprises the following steps: acquiring a plurality of common sample sets corresponding to a plurality of users, and training a common model by using the common sample sets; then, based on the trained public model, joint training including multiple rounds of iteration is carried out on the multiple local models, any round of iteration comprising the steps that for any first user, the public model is used for processing a public sample corresponding to the first user, a public prediction result is obtained, and the public prediction result is sent to a corresponding first terminal; then, receiving parameter update data of the first local model from the first terminal, which is determined based on the public prediction result and a local sample of the first terminal; and then, on the basis of a plurality of parameter update data received from the plurality of terminals, determining the parameters subjected to the joint update in the round, and providing the parameters to the plurality of terminals.

Description

technical field [0001] One or more embodiments of this specification relate to the field of computer technology, and in particular to a training method and device for a business forecasting system. Background technique [0002] With the advancement of science and technology and the development of society, a large number of service platforms have emerged to provide people with various services to meet people's needs in work and life. At the same time, machine learning technology has become a current research hotspot, and many service platforms try to provide users with better services by predicting user behavior with the help of machine learning models. For example, the service platform can decide whether to issue the discount group to a certain user by predicting whether a certain user will cancel the coupons issued to him, so that as many users as possible can truly enjoy the limited benefits issued by the service platform . [0003] However, there are some problems that ...

AI Technical Summary

Problems solved by technology

[0003] However, there are some problems that need to be improved in the current model prediction of user behavior, for example, the prediction accuracy is limited, or there is a risk of privacy leakage during the model training process, etc.

Images