Spatio-Temporal Traffic Prediction Using Crossover Attention for Communications and Networking

This paper investigates the spatio-temporal multivariate time series prediction problem, which has important applications in various real-world tasks including network traffic modeling, network slicing, and channel estimation. To tackle this problem, many attention-based models have been proposed in the literature to predict the output of future time slots. However, we notice that the majority of attention models was designed to capture input dependency structures in only a single domain (typically the temporal domain), which limits their prediction accuracy. To solve this issue and further improve the performance of attention-based models, we propose a novel crossover attention mechanism in this paper. The crossover attention can be understood as a learnable regression kernel which prioritizes the input sequence with both spatial and temporal similarities and extracts relevant information for generating the output of future time slots. Simulation results based on realistic datasets show that when replacing the vanilla attention with the proposed crossover attention, considerable improvement on the prediction accuracy can be achieved for the existing attention-based models.