Efficient Mobile Cellular Traffic Forecasting using Spatial-Temporal Graph Attention Networks

Cellular traffic prediction is an essential aspect of mobile network management that uses data analytics and machine learning to forecast the volume and pattern of communication traffic generated by mobile users at a particular location and time. Graph Convolution Network (GCN) has been widely employed to model the spatial relationships between different cell towers and their neighboring counterparts. However, GCN is limited to highly regular and well-structured graphs. This paper proposes a Graph Attention Network (GAT) to capture more nuanced spatial relationships between cell towers, making it more suitable for irregular and complex graphs. Additionally, a novel graph attention mechanism is proposed that enables the creation of a dynamic graph structure, capable of capturing the evolving spatial relationships over time. Comprehensive experiments on an actual cellular traffic dataset show that the proposed technique outperforms state-of-the-art baselines on two evaluation metrics - RMSE and MAE - with a significant improvement.