ComS2T: A Complementary Spatiotemporal Learning System for Data-Adaptive Model Evolution.

Area of Science:

• Artificial Intelligence
• Urban Computing
• Machine Learning

Background:

• Spatiotemporal (ST) learning is vital for smart cities but struggles with fluctuating urban data distributions.
• Existing ST learning models lack generalization and data adaptation capabilities for new observations, requiring inefficient retraining.

Purpose of the Study:

• To introduce ComS2T, a prompt-based complementary spatiotemporal learning framework.
• To enhance model adaptability to evolving urban data and city structures.
• To enable efficient model fine-tuning for out-of-distribution scenarios.

Main Methods:

• Disentangled neural architecture into stable (neocortex) and dynamic (hippocampus) components.
• Trained dynamic spatial and temporal prompts adaptive to new data distributions.
• Employed a two-stage training process with prompt-conditioned fine-tuning.

Main Results:

• ComS2T demonstrated significant efficacy in adapting to various spatiotemporal out-of-distribution scenarios.
• The prompt-based mechanism enabled efficient data adaptation during testing.
• Maintained effective inference capabilities after adaptation.

Conclusions:

• ComS2T offers an efficient and effective solution for spatiotemporal learning in dynamic urban environments.
• The proposed method addresses the generalization and data adaptation limitations of existing ST learning models.
• ComS2T facilitates sustainable urban development through improved smart city technologies.