PROJECT TITLE :

Graph Attention Spatial-Temporal Network with Collaborative Global-Local Learning for Citywide Mobile Traffic Prediction

ABSTRACT:

It is becoming increasingly important for proactive network service provisioning and efficient network resource allocation in Smart Cities that mobile traffic forecasting be performed in a timely manner with a high level of accuracy. This is due to the rapid development of mobile cellular technologies as well as the growing popularity of mobile and Internet of Things (IoT) devices. The majority of traditional methods for traffic forecasting are based on time series prediction techniques, which are incapable of accurately capturing the complex dynamic nature and spatial relations of mobile traffic demand. In this paper, we propose a novel Deep Learning framework for accurate citywide mobile traffic forecasting. We call it the graph attention spatial-temporal network (GASTN). This framework can capture not only local geographical dependency but also distant inter-region relationship when considering spatial factor. In particular, GASTN takes into account spatial correlation by means of the spatial relation graph that we have constructed and makes use of structural Recurrent Neural Networks in order to model the global near-far spatial relationships in addition to the temporal dependencies. Two different Attention Mechanisms are being developed within the framework of GASTN in order to integrate various effects in a more holistic manner. In addition, in order to further improve the performance of the prediction, we propose a collaborative global-local learning strategy for the training of GASTN. This strategy makes full use of the knowledge from both the global model and the local models for individual regions in order to improve the efficiency of our model. Our GASTN model has been shown to significantly outperform the methods that are considered to be state-of-the-art, as shown by extensive experiments conducted on a large-scale real-world mobile traffic dataset. In addition, it demonstrates that the utilization of the collaborative global-local learning strategy is capable of producing a sizeable improvement in the level of prediction performance achieved by GASTN.