Dynamic Modes as Time Representation for Spatiotemporal Forecasting

2025-06-01

AI Summary

Existing spatiotemporal forecasting methods rely on handcrafted temporal features or explicit timestamps to model long-range seasonal dependencies, limiting adaptability and generalizability. To address this, we propose a data-driven, dynamic time representation method that introduces Dynamic Mode Decomposition (DMD) into time embedding learning for the first time. Our approach unsupervisedly extracts multiscale periodic modes directly from raw time series, enabling timestamp-agnostic and adaptive capture of multiple periodicities. The resulting representation is model-agnostic, lightweight, and highly generalizable, and can be seamlessly integrated into mainstream spatiotemporal architectures—including graph neural networks and Transformers. Extensive experiments on urban traffic, highway, and climate datasets demonstrate that our method significantly improves long-horizon forecasting accuracy, reduces residual autocorrelation, and enhances cross-temporal generalization performance.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Data Mining & Knowledge Management: Mining of Spatial, Temporal or Spatio-Temporal Data
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

This paper introduces a data-driven time embedding method for modeling long-range seasonal dependencies in spatiotemporal forecasting tasks. The proposed approach employs Dynamic Mode Decomposition (DMD) to extract temporal modes directly from observed data, eliminating the need for explicit timestamps or hand-crafted time features. These temporal modes serve as time representations that can be seamlessly integrated into deep spatiotemporal forecasting models. Unlike conventional embeddings such as time-of-day indicators or sinusoidal functions, our method captures complex multi-scale periodicity through spectral analysis of spatiotemporal data. Extensive experiments on urban mobility, highway traffic, and climate datasets demonstrate that the DMD-based embedding consistently improves long-horizon forecasting accuracy, reduces residual correlation, and enhances temporal generalization. The method is lightweight, model-agnostic, and compatible with any architecture that incorporates time covariates.

Problem

• Modeling long-range seasonal dependencies in spatiotemporal forecasting
• Eliminating need for explicit timestamps or hand-crafted time features
• Improving long-horizon forecasting accuracy and temporal generalization

Innovation

• Dynamic Mode Decomposition for time embedding
• Data-driven spectral analysis of spatiotemporal data
• Model-agnostic lightweight temporal representation