基于个性化联邦学习的无线通信流量预测

Citywide Wireless Traffic Prediction Based on Personalized Federated Learning

提出一种基于分布式的城市全域通信流量预测算法Fed-DenseNet,各个边缘计算服务器在中心服务器的协调下进行协同训练,中心服务器利用KL散度挑选出流量分布相似的区域流量模型,并采用联邦平均算法对具有相似流量分布的区域流量模型的参数进行融合,以较低的复杂度和通信开销实现城市全域流量预测。此外,城市范围内不同地区流量具有高度差异化的特征,为此,在Fed-DenseNet算法基础上,提出基于合作博弈的个性化联邦学习算法p-Fed-DenseNet,将本地区的各个区域性数据特征作为合作博弈的参与者,通过合作博弈的超可加性准则,进行本地区特征的筛选,从而达到既能提高模型的泛化性,又能够保持对本地流量精准刻画的目的。

Wireless communication network traffic prediction is of great significance to operators in network construction,base station wireless resource management and user experience improvement.However,the existing centralized algorithm models face the problems of complexity and timeliness,thus difficult to meet the traffic prediction of the whole city scale.A distributed urban global traffic prediction algorithm Fed-DenseNet is proposed in this paper.Each edge computing server of the algorithm performs collaborative training under the coordination of the central server,and the central server uses KL(Kullback-Leibler)divergence to select regional traffic models with similar traffic distribution and uses the federated average algorithm to fuse the parameters of these regional traffic models.In this way,the urban global traffic prediction can be realized with lower complexity and communication cost.In addition,the traffic in different areas within the city is highly differentiated,so how to improve the accuracy of model prediction is also facing challenges.Based on Fed-Densenet algorithm,a personalized federated learning algorithm p-Fed-DenseNet based on cooperative game is proposed.Each regional data feature in the region is taken as a participant of cooperative game,and local features are screened by the superadditivity criterion of cooperative game,so as to achieve the purpose of both improving the generalization of the model and maintaining the accurate description of local traffic.