LiteCST: Lightweight Sparse Spatial-Temporal Modeling by Leveraging Cyclic Shift for Cellular Traffic Prediction

Knowledge discovery from cellular traffic data, which serves as a crucial foundation for perceiving network dynamics and enabling proactive interventions, is advancing towards higher accuracy in its core traffic prediction tasks. Nevertheless, most high-accuracy predictive methods are associated with complicated models and intensive computational demand, which is prohibitive for resource-constrained edge devices. To this end, this paper proposes a Lightweight Cyclic shift-empowered Spatial-Temporal modeling method for efficient cellular traffic prediction (LiteCST). Concretely, long sequences are adaptively divided into patches of varying lengths at different time periods and standardized by downsampling, thereby dramatically reducing the computational complexity. To further enhance efficiency, patches are cyclically shifted across neighboring sequences to model sparse spatial-temporal correlations, which enables the simultaneous spatial-temporal feature extraction by a single One-Dimensional (1D) attention mechanism. Additionally, two cyclic shift patterns are developed, one of which is LiteCST-Fix, a fixed pattern aimed at achieving ultimate efficiency. LiteCST-Opt, as another pattern, optimizes the selection of shifted patches based on significance measurement and theoretically derives the optimal shift offsets, trading a slight increase in complexity for higher accuracy. Extensive experiments against state-of-the-art baselines are conducted on three real-world datasets: WKG, Milan and AIIA. The experimental results demonstrate that LiteCST achieves a substantial improvement in efficiency while maintaining competitive accuracy compared with the advanced methods.