Individual traffic prediction in cellular networks based on tensor completion

Accurate individual (per-user) traffic prediction in cellular networks is considered a critical capability for the system performance improvement in terms of dynamic bandwidth allocation and network optimization. However, existing methods have limitations in capturing the characteristics of individual traffic, because the observed traffic data are usually incomplete and traffic consumption patterns significantly differ among users. In this paper, to fully exploit the inherent temporal-spatial correlations of individual traffic data, this work models the traffic data as a three-dimensional traffic tensor. Different from the "three-step" strategies (i.e., "data processing + decomposition + prediction") in the existing methods, we propose a novel tensor completion (TC)-based "two-step" strategy, that is, "data processing & decomposition + prediction" for individual traffic prediction, which can recover and decompose the traffic data simultaneously. Furthermore, an efficient algorithm based on the alternating direction method of multipliers (ADMM) framework is proposed to solve the resulting model. To the best of our knowledge, this is the first work to apply the TC-based "two-step" strategy for individual traffic prediction in cellular networks. Experiments conducted on a real cellular traffic dataset empirically validate the superiority of the proposed method.