Search papers, labs, and topics across Lattice.
The Future Decomposition Network (FDN) is introduced as a novel model for spatiotemporal forecasting that prioritizes interpretability while maintaining competitive accuracy. By employing classification to generate interpretable predictions and uncover latent activity patterns, FDN not only enhances understanding of the underlying data but also achieves state-of-the-art performance with significantly reduced memory and runtime requirements. Comprehensive evaluations across hydrologic, traffic, and energy datasets confirm FDN's effectiveness in balancing interpretability and forecasting precision.
FDN achieves state-of-the-art forecasting accuracy while providing interpretable insights into spatiotemporal data, all with reduced computational demands.
Spatiotemporal systems comprise a collection of spatially distributed yet interdependent entities each generating unique dynamic signals. Highly sophisticated methods have been proposed in recent years delivering state-of-the-art (SOTA) forecasts but few have focused on interpretability. To address this, we propose the Future Decomposition Network (FDN), a novel forecast model capable of (a) providing interpretable predictions through classification (b) revealing latent activity patterns in the target time-series and (c) delivering forecasts competitive with SOTA methods at a fraction of their memory and runtime cost. We conduct comprehensive analyses on FDN for multiple datasets from hydrologic, traffic, and energy systems, demonstrating its improved accuracy and interpretability.
