NuwaDynamics+: A Causality-Aware Generative Framework for Spatio-Temporal Representation Learning

Area of Science:

• Earth Sciences
• Artificial Intelligence
• Data Science

Background:

• Spatio-temporal (ST) prediction is vital for fields like meteorology and urban computing.
• Data scarcity, imbalances, and lack of causal reasoning in models hinder ST prediction accuracy and generalizability.
• Existing deep learning models often struggle with real-world data complexities and computational demands.

Purpose of the Study:

• To introduce NuwaDynamics, a novel causal framework for ST prediction.
• To enhance model generalizability and causal reasoning capabilities in ST prediction.
• To address data scarcity, imbalances, and computational complexity in ST prediction tasks.

Main Methods:

• Developed NuwaDynamics, a dual-phase causal framework (Discovery and Update) for ST prediction.
• Employed upstream self-supervision for causal region identification and targeted data interventions.
• Integrated channel multiplication and conditional generation (NuwaDynamics+) for environmental controllability and efficiency, akin to front-door adjustment.

Main Results:

• NuwaDynamics+ significantly improved model performance across ten diverse ST benchmarks.
• Demonstrated enhanced versatility in dynamic ST tasks, including extreme weather forecasting.
• Showcased improved long-temporal-step super-resolution prediction capabilities.

Conclusions:

• The NuwaDynamics+ framework offers a robust solution to challenges in ST prediction.
• Causal reasoning integration substantially boosts ST model performance and generalizability.
• NuwaDynamics+ provides a computationally efficient and versatile approach for complex ST prediction tasks.