Research On Key Technologies Of Personalized Federated Learning For Recommender System

With the rapid development of science and technology,the Internet plays an increasingly important role in human society.After the development and accumulation in recent years,recommendation system plays an important role in many aspects of life.With the rapid development of Internet economy,it is increasingly important to find users’ interests on the Internet.Therefore,recommendation system is one of the most important research fields nowadays.However,with the continuous improvement of popularity and practicability of recommendation system,a series of problems need to be solved,among which the most prominent is the user’s data privacy.Federated learning has the characteristics of processing private data locally and only transferring model update information to the central server.Therefore,federated learning has the natural ability of privacy protection and is suitable for solving the data privacy problem of recommendation system.However,the recommendation system based on the classical federated learning model does not have the ability of personalized recommendation.This thesis intends to design the recommendation system based on personalized federated learning algorithm to solve the problems of privacy and accuracy of personalized recommendation.After that,the incentive mechanism is designed to improve the training quality and efficiency of the whole system.The specific research contents are as follows:Firstly,a recommendation system based on personalized federated learning is designed for privacy protection and personalized recommendation.By combining the ideas of MAML and Fed Avg,a personalized federated learning algorithm MFL-RS is proposed,which has both the data privacy protection ability and the personalized ability of the meta-learning algorithm.On this basis,by choosing the architecture of the recommender system based on this algorithm,the main part of the recommender system based on the personalized federated learning algorithm is proposed.The ability of personalized recommendation and speed of model aggregation of the recommender system are verified by simulation experiments.The privacy data protection feature of federated learning also gives the recommendation system the ability to protect users’ local privacy data.Secondly,aiming at the problem that participants who provide low quality data will reduce the speed of system model aggregation and recommendation accuracy,an incentive mechanism algorithm suitable for personalized federated learning recommendation system is proposed.In the design of incentive mechanism,accurate calculation of each participant’s contribution is the basis of the whole algorithm.Based on Shapley value theory,the contribution quantification algorithm of participants is proposed.On this basis,through incentive allocation according to the contribution of participants,high-quality participants are attracted to participate in training and low-quality participants give up participation due to low input-output ratio.In the simulation experiments,the effectiveness of the proposed incentive mechanism on shortening model training time and improving recommendation quality is verified by comparing with the recommendation system which adopts other contribution quantization algorithms.