Bayesian Meta-Learning for Adaptive Traffic Prediction in Wireless Networks

Wireless traffic prediction is indispensable for network planning and resource management. Due to different population distributions and user behavior, there exist strong spatial-temporal variations in wireless traffic across different regions. Most of the conventional traffic prediction approaches can only tackle a particular spatial-temporal pattern and cannot capture such variations in wireless traffic. This motivates us to develop an adaptive approach which can tackle spatial-temporal variations and predict wireless traffic in different regions. In this paper, we formulate an adaptive traffic prediction problem from a probabilistic inference perspective and develop a variational spatial-temporal Bayesian meta-learning (VST-BML) algorithm. We model the traffic prediction in different regions as different prediction tasks. The proposed VST-BML algorithm can learn the common spatial-temporal features shared by all prediction tasks, and adaptively infer the task-specific parameters to tackle spatial-temporal variations. We evaluate the performance of our proposed VST-BML algorithm using a real-world traffic dataset. Experimental results show that the proposed algorithm can quickly adapt to different prediction tasks by using only a small number of data samples and provide accurate traffic prediction in different regions. When compared with five baseline methods, the proposed algorithm can reduce the root-mean-square error (RMSE) and mean absolute error (MAE) by 53.0% and 48.4%, respectively.