Spatial Bottleneck Transformer for Cellular Traffic Prediction in the Urban City

Due to the widespread use of portable devices and the advancement of 5G technology, we have received a significant amount of mobile data, which requires prediction models for cellular traffic data. However, accurately forecasting mobile traffic data is challenging due to the complex spatial and temporal correlations, especially when the mobile data comes from a large geographical area. To tackle this challenge, we propose a new model, called ST-InducedTrans, to dynamically explore the large geographical correlations (spatial) and periodic variations (temporal). Specifically, a Spatial Bottleneck Transformer is devised to obtain spatial correlations from the most relevant grids in the geographical area at the cost of linear complexity. For the temporal blocks, we embed the elaborately selected temporal clues into a temporal Transformer to offer useful temporal prompts for cellular prediction. Finally, several spatial and temporal blocks are effectively stitched into a whole model for complementary cellular traffic prediction. We conducted comprehensive experiments on the public real-world cellular data from Milan. Results show that our model outperforms the state-of-the-art methods on three metrics (MAE, NRMSE, and R \(^2\) ) at the cost of lower time complexity.