Introduction
Predicting the evolution of complex spatiotemporal systems—such as urban traffic flows, power grid loads, or global climate patterns—is a cornerstone of modern data science. Traditionally, researchers have faced a trade-off between deterministic forecasting, which excels at capturing regular patterns but ignores uncertainty, and probabilistic forecasting, which models distributions but often struggles with the computational complexity of high-dimensional structural trends.
The CoST (Collaborative SpatioTemporal forecasting) framework addresses this by introducing a "divide and conquer" strategy. By decomposing the prediction task into mean and residual components, CoST achieves significant performance gains and computational efficiency over existing state-of-the-art methods.
The CoST Framework: Mean-Residual Decomposition
The core innovation of CoST is the Mean-Residual Decomposition strategy. Rather than forcing a single model to learn both massive structural trends (like seasonality) and stochastic fluctuations, CoST splits the task:
- The Mean ($\mu$): A high-performance deterministic backbone (e.g., STID or iTransformer) is pretrained to capture regular, predictable patterns.
- The Residual ($r$): A lightweight diffusion model is trained to model only the difference between the deterministic prediction and the ground truth.
Because residuals are mathematically simpler than raw data, the diffusion model can be significantly smaller and faster. To further refine this, CoST employs a Scale-aware Diffusion Mechanism. By applying a Fast Fourier Transform (FFT) to training data, the model calculates a "customized fluctuation scale" ($Q$), which serves as a spatial prior for the diffusion process, accounting for varying volatility across different locations.
Experimental Evaluation
CoST was evaluated across ten real-world datasets spanning four domains:
- Climate: SST-CESM2 and SST-ERA5.
- Energy: SolarPower station data.
- Communication: Mobile network traffic in Nanjing (MobileNJ) and Shanghai (MobileSH).
- Urban Systems: Crowd flows (CrowdBJ/BM), taxi trajectories (TaxiBJ), bike-sharing (BikeDC), and traffic speeds (Los-Speed).
Key Results and Efficiency
CoST demonstrated an average 25% performance improvement in probabilistic metrics (such as QICE and Interval Score) compared to baselines like CSDI and NPDiff.
Beyond accuracy, the efficiency gains are stark. For instance, on the MobileSH dataset, CoST completed training in just 2 minutes, whereas the baseline CSDI required 48 minutes. Inference times saw similar reductions, dropping from 38 minutes to just 50 seconds.
Case Study: Climate Forecasting
In climate applications, CoST's ability to model uncertainty provides geophysically consistent insights. In a sea surface temperature (SST) forecasting task, the model successfully identified high-uncertainty regions in the central equatorial Pacific, matching known areas of divergence in climate science models.
Insights and Future Directions
The success of CoST highlights several key principles:
- Modularity: The framework is model-agnostic, allowing any state-of-the-art deterministic model to serve as the backbone.
- Residual Advantage: Simplifying the generative objective to residuals prevents the diffusion model from "wasting" capacity on obvious periodic signals.
- Spatial Priors: Using frequency analysis to guide diffusion effectively incorporates domain knowledge into the generative process.
While highly effective, CoST's performance remains tied to the quality of its deterministic backbone. Future work may explore physics-informed extensions to ensure predictions obey fundamental laws (e.g., fluid dynamics) and adaptive decomposition methods to dynamically adjust the split between mean and residual components based on data complexity.
Summary updated