Abstract
Idle time windows (ITWs) consist of one critical trigger for various functions in green intelligent network management and traffic scheduling in mobile networks. In this paper, we study the ITW prediction in mobile networks based on network subscribers' demand and mobility behaviors observed by network operators. We first innovatively formulate the ITW prediction into a regression problem with an ITW presence confidence index that facilitates direct ITW detection and estimation. Feature extraction on the demand and mobility history is then proposed to capture the current trends of subscribers' demand and mobility as well as to account for the periodicity underlying subscribers' demand and mobility patterns as exogenous inputs. In light of feature engineering, a deep learning-based ITW prediction model is proposed, which consists of two components, namely the representation learning network and the output network. The representation learning network is aimed to learn effective patterns, whereas the output network is designed to produce the desired ITW presence confidence index and the ITW estimate by integrating the learned representation and exogenous inputs. In this paper, a novel temporal graph convolutional network (TGCN) for the representation learning network is proposed to effectively capture the graph-based spatiotemporal input features. The experiment results validate the proposed direct ITW prediction formulation and demonstrate the superiority of the proposed TGCN in terms of both ITW detection and ITW estimation performance, which can achieve a significant intersection-over-union (IoU) improvement compared with baselines.
Original language | English (US) |
---|---|
Article number | 8667305 |
Pages (from-to) | 1441-1454 |
Number of pages | 14 |
Journal | IEEE Journal on Selected Areas in Communications |
Volume | 37 |
Issue number | 6 |
DOIs | |
State | Published - Jun 2019 |
Externally published | Yes |
Bibliographical note
Publisher Copyright:© 1983-2012 IEEE.
Keywords
- Machine learning
- mobile communication
- predictive models