Large-scale cellular traffic prediction based on graph convolutional networks with transfer learning

Intelligent cellular traffic prediction is very important for mobile operators to achieve resource scheduling and allocation. In reality, people often need to predict very large scale of cellular traffic involving thousands of cells. This paper proposes a transfer learning strategy based on graph convolution neural network to achieve the task of large-scale traffic prediction. In this paper, we design a novel spatial-temporal graph convolutional network based on attention mechanism (STA-GCN). In order to achieve large-scale traffic prediction, this paper proposes a regional transfer learning strategy based on STA-GCN to improve knowledge reuse. The effectiveness of STA-GCN is validated through two real-world traffic datasets. The results show that STA-GCN outperforms the state-of-art baselines, and the transfer learning strategy can effectively reduce the number of epochs while training.