NLBST: New Bayesian model for spatio-temporal forecasting with long-range coupling
Handles irregular time points, sparse data, and uncertainty quantification in one framework.
Deep Dive
Researchers propose NLBST, a hierarchical Bayesian model for continuous spatio-temporal dynamics with nonlocal interactions. It uses a spatial basis expansion, continuous-time ODE, and Neural ODE residuals to capture nonlinear dynamics. Kalman-style sequential updates handle missing and irregular observations. The model achieves strong forecasting and spatial generalization with well-calibrated uncertainty, outperforming baselines in partially observed regimes.
Key Points
- NLBST uses a coordinate-based spatial basis expansion with a continuous-time ODE to model nonlocal spatio-temporal dynamics.
- Kalman-style sequential updates handle missing and irregular observations without retraining for new locations.
- Achieved substantial gains over baselines in forecasting accuracy and uncertainty calibration on synthetic and real-world datasets.
Why It Matters
Enables accurate, uncertainty-aware forecasting for climate, epidemiology, and sensor networks with sparse, irregular data.