Federal learning-based sequence recommendation method and system

Abstract

The invention discloses a sequence recommendation method and system based on federal learning, and belongs to the field of personalized recommendation and the field of user privacy protection. According to the method, personal information and real-time and historical data of a user are stored locally, each client can perform state representation according to own historical data, an interaction relationship between the user and an article is captured, and a recommendation process is regarded as a serialized decision process based on deep reinforcement learning; according to the method, an attention mechanism is introduced into an aggregation algorithm of federated learning, correlation of feature information extracted by each client and data difference brought by individuation of each client are considered, a weight coefficient is formulated for each client, which is a fine-grained re-weighting means, the individuation degree of data is increased, and the robustness of the data is improved. According to the method, the recommendation accuracy is improved, the method is more suitable for the field of sequence recommendation, and personalized recommendation for the user is realized while personal data privacy is protected.

Description

technical field [0001] The present invention relates to the field of recommendation methods and user privacy protection, in particular to a method and system for sequence recommendation based on federated learning. Background technique [0002] With the continuous development of information technology, the information on the Internet has grown exponentially, and users cannot quickly find the information they want on the Internet, thus the personalized recommendation technology was born. At present, common recommendation methods assume that user preferences are a static process, that is, their preferences are basically unchanged from current historical data. Therefore, existing algorithms use technologies such as collaborative filtering, matrix factorization, and deep learning based on historical data. Modeling user preferences to independently predict the rating or ranking of each item to be recommended. Although this technical solution is highly interpretable, it can contin...

AI Technical Summary

Problems solved by technology

Among them, policy-based methods learn to generate strategies to directly generate actions. This type of scheme focuses on continuous action spaces, and there is a non-negligible gap between discrete and continuous action spaces; value-based methods evaluate the Q value of all actions in a specific state. Select the action with the largest Q value, so when the action space is large, the algorithm efficiency will become very inefficient

[0004] In addition to considering the performance of recommendations, data privacy security and real-time performance are also important research directions. The existing classic general-purpose federated learning algorithm is FedAvg and its related variants. In this way, the model parameters or gradient information participating in the aggregation are involved in the aggregation. Most of them use the average or weighted average method. Such a simple aggregation method tends to ignore the personalization of the client-side model. More importantly, it may encounter some The potential threat of model attacks, that is, some local models perform extremely or poorly, such a simple average will affect the performance of the global model, especially in the field of sequence recommendation, which requires more intelligent and personalized selection modes and aggregation methods

Images