FDN: Interpretable Spatiotemporal Forecasting with Future Decomposition Networks

2026-06-23

AI Summary

While existing spatiotemporal forecasting methods achieve high accuracy, they are often black-box models that lack interpretability and incur substantial computational overhead. To address these limitations, this work proposes the Future Decomposition Network (FDN), which explicitly decomposes future states into interpretable semantic categories for the first time. By integrating spatiotemporal modeling with a classification-driven mechanism, FDN enables lightweight, end-to-end interpretable forecasting. Extensive experiments on real-world datasets from hydrology, transportation, and energy domains demonstrate that FDN matches or surpasses state-of-the-art methods in prediction accuracy while significantly reducing memory consumption and runtime. Moreover, it provides clear, dynamic pattern explanations, thereby overcoming the longstanding trade-off between interpretability and efficiency inherent in conventional black-box approaches.

Abstract

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.

Problem

• spatiotemporal forecasting
• interpretability
• time-series
• latent patterns
• predictive modeling

Innovation

• Future Decomposition Network
• Interpretable Forecasting
• Spatiotemporal Modeling
• Latent Pattern Discovery
• Efficient Prediction