Federated Learning Based Content Popularity Prediction In Fog Radio Access Networks

With the unprecedented rapid proliferation of intelligent devices,wireless networks are confronted with a myriad of challenges and notably data traffic pressure on the fronthaul wireless links.To cope with this issue,fog radio access networks(F-RANs)have emerged as an innovative network architecture to alleviate the traffic burden on fronthaul links by caching popular contents in fog access points(F-APs).The edge caching technology in F-RAN makes full use of the F-APs with certain computing and caching resources.By pre-caching popular contents in the F-APs,it can effectively reduce the fronthaul load and reduce the communication delay.The edge caching technology in F-RAN has become an effective solution that has received widespread attention.Due to the caching capacity constraints,F-APs need to predict future content popularity accurately in order to prefetch the most popular contents during off-peak traffic periods and improve caching efficiency.However,diverse and time-varying user requests and huge data volumes have brought many challenges to the construction and training of popularity prediction models.In this paper,we will conduct systematic research on the content popularity prediction method and distributed model training method in F-RAN based on the above problems.Firstly,a user-centric content popularity prediction policy is studied.Firstly,we give a system model and define the content popularity as the ratio of the number of content requests to the total number of requests within a region.Secondly,a user preference learning algorithm based on the user's local data is designed,then the user equipment can build a logistic regression model and learn the user preference based on the historical requests.Then,an adaptive context space partitioning algorithm based on context awareness is designed to cluster users effectively by using the context information of users.The main purpose is to reduce the amount of calculation when training popularity prediction models.Further,according to the learned user preferences and the user clustering results after adaptive context space partitioning,the model input is calculated through the mapping function and a regional content popularity prediction model is constructed.Finally,simulation results show that the proposed content popularity prediction policy has higher prediction accuracy and cache hit rate than traditional methods.Secondly,the training method of distributed popularity prediction model based on federated learning is studied.Taking the regionality of content popularity into consideration,a distributed optimization problem of learning global popularity prediction model is constructed,and a popularity prediction model integration scheme based on federated learning is designed.Firstly,a global optimization problem is constructed based on the training data with the objective of minimizing the mean square error and then transformed into a distributed optimization problem related to the local optimization problems.Subsequently,based on the basic process of federated learning,the global popularity prediction model is learned in a distributed manner through alternate iterations of ‘‘local computation-global aggregation” with the F-APs as the participants of federated learning,thereby alleviating computation burden on a single node and reducing communications resource consumption.Finally,the simulation results show that the content popularity prediction model training method based on federated learning is trained on the basis of the existing local models,which can also reduce unnecessary recalculation and accelerate the convergence speed.Finally,the performance analysis of the distributed model training method based on federated learning is studied.In order to compare the distributed model training method based on federated learning with the traditional model training method based on centralized learning,a scenario in which the model training process based on federated learning and the model training process based on centralized learning are synchronized regularly is proposed.The upper bound of the difference between the result of the model training process based on federated learning and that of the model training process based on centralized learning is derived and denoted by the expression of key parameters.The simulation results verify the correlation between the key parameters of the theoretical analysis and the convergence bound.