An Adaptive Ensemble Learning Paradigm With Spatial-Temporal Feature Extraction for Wireless Traffic Prediction

Accurately predicting traffic in a cellular network is challenging since the traffic time series integrated by various wireless services is non-stationary and reveals concealed spatial correlation among different cells. Due to that, the presence of bias in a single forecast model often hinders the ability to generalise under numerous circumstances in wireless traffic data, no particular approach stands out as clearly superior to the others. In this paper, we propose an adaptive ensemble learning paradigm that can benefit from centralizing individual forecast base models. It stacks the prediction outputs of several base learners due to the traffic dynamics characteristic. An improved convolutional neural network (CNN)-based representation learning method is designed to extract the high-order spatial-temporal features in the traffic data and obtain the adaptive weights of participating base learner models for the ensemble. The experimental results verify that the proposed ensemble approach can fully utilize spatial-temporal features and outperform individual statistical and machine-learning models regarding prediction accuracy. Furthermore, the ensemble method via stacking base models with fewer parameters is capable of generating predictions close to the large-parametric spatial-temporal transformer (ST-Tran) model produced.