Graphical model full-life-cycle modeling method for federated learning

Abstract

The invention provides a graphical model full-life-cycle modeling method for federated learning, which comprises the following steps: analyzing a data sample before modeling, and determining a federated learning scheme; drawing up a modeling strategy according to a joint analysis operation result; adopting a dragging mode and using a federation feature engineering means to process the data; in thefederated model training process, after the two parties construct models respectively, exchanging an intermediate state model and a loss function which are obtained through calculation ; predicting and using the federated model; and monitoring the operation process of the federated model to obtain a model monitoring statistical index so as to realize maintenance iteration of the federated model.Personnel in each step of federated learning modeling can participate in optimization points through imaging, the personnel participation optimization difficulty is reduced, a later model applicationmaintenance way is provided, and the application scene and usability of the federated learning technology are improved.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a modeling method for the whole life cycle of a graphical model for federated learning. Background technique [0002] Federated learning is a great way to connect data silos to get high-quality models. At present, there are not many mature technologies for federated learning, and the existing technologies are mainly devoted to solving the problem of homomorphic and semi-homomorphic encryption and decryption in the modeling process. However, since the multi-party data participating in the federation needs to be kept secret from each other, the basic statistical information exchange of federated samples, federated modeling strategy formulation, federated feature engineering, federated model training, federated model predictive use, federated model maintenance iterations, etc. are involved. Both are the key to federation members being able to use high-quality models. Each ...

AI Technical Summary

Problems solved by technology

[0003] The current existing technologies are insufficient for these steps or splits, so that many steps can only use the default settings of the framework without manual participation in the optimization model; or some steps in the model construction process, such as binning adjustment, cannot be manually intervened; or for the model There are too many constraints in the training process, which makes it impossible to manually adjust and optimize the model, resulting in poor model results; or the training effect of traditional modeling frameworks such as sk-learn, which is more mature in the model training framework, is too different; or the types of models that can be realized are limited to LR and BOOST models ; or use the "debugging background" type of interaction, which requires users not only to have professional risk control modeling capabilities, but also to be very familiar with various software frameworks used in it, and to be able to read obscure debugging logs; or the later stage of the model Difficulty in application and maintenance

Images