Federated Bayesian Learning Based Content Popularity Prediction And Edge Caching In Fog Radio Access Networks

With the rapid development of smartphones,smart Io T devices,smart vehicles,and other devices,the traffic in communication networks is exploding,and the extremely high data traffic puts a great transmission burden on communication networks.One of the main problems faced in communication networks is that there is a large amount of duplicate content being transmitted in the network,resulting in a serious waste of bandwidth resources.To alleviate this situation,fog radio access networks(F-RANs),a new network architecture,reduce the pressure on communication links by storing highly popular content in advance in fog access points(F-APs).F-RANs allow F-APs to cache popular content in advance during low-peak user request times,efficiently utilizing the computing power and storage resources of F-RANs and reducing the duplicate transmission of content in the backbone network,thus effectively reducing user request latency and relieving the pressure on the communication link.Edge caching technology in F-RANs has emerged as an efficient solution.Since F-APs are equipped with limited storage capacity,they are required to predict content popularity accurately and in real time,make caching decisions efficiently,and improve cache space utilization.However,the unevenly distributed user request information and the high requirements for latency and computational complexity in communication networks pose many difficulties for popularity prediction.In this paper,we systematically investigate the content popularity prediction methods and corresponding edge caching strategies in F-RANs with respect to the above research challenges.First,a content popularity prediction policy based on Bayesian learning is studied.A system model is established,and the content popularity is approximated by the metric of content request number,and the root mean square error between the content popularity and the real popularity is used as the system performance evaluation criterion.Next,the content request model is determined and a probabilistic model is built with a Gaussian process as the core.Then,an improved model training method based on the traditional Markov Chain Monte Carlo is designed to reduce the computational complexity and accelerate the convergence by identifying the shortcomings of the traditional method of traversing the data set to obtain the gradient and combining the variance-reducing stochastic gradient into the traditional method.Finally,the popularity prediction results are obtained by point estimation based on the nature of Gaussian process and the model training results.The simulation results show that the proposed content popularity prediction policy has better convergence speed and excellent prediction accuracy.Second,a distributed popularity prediction model training method based on federated Bayesian learning is investigated.Considering the problem of insufficient data in a single region,a scheme for collaborative popularity prediction is constructed.Due to user privacy and efficient communication considerations,a federated learning framework is introduced for model training.First,each F-AP is updated with a local model based on local datasets using Bayesian learning.Then,a ”local update-global aggregation” process is performed based on the federated learning.Finally,considering that the aggregation process also requires some communications overhead to send the gradient vectors,a strategy is proposed to quantize and encode the gradients before sending them,thus reducing the waste of resources.Finally,the simulation results show that the distributed popularity prediction model training method based on federated Bayesian learning can effectively reduce the repeated computations and ensure the model accuracy while obtaining high communications resource utilization.Finally,the edge caching policy based on federated Bayesian learning is studied.The edge caching policy based on federated Bayesian learning of content popularity prediction policy is proposed by combining it with active caching.Through active tracking of user requests,the content popularity can be updated and active cache replacement decisions can be made in a timely manner based on the current number of people in the area.In addition,the internal architecture of the F-AP is designed to enable it to fulfill the communication,computation,and caching requirements to complete the proposed edge caching policy.Finally,the simulation results show that this edge caching policy can significantly improve the cache hit rate compared to the traditional caching policy.