Reproduce “ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting” on OpenResearch

About this paper

Xvyuan Liu, Xiangfei Qiu, Hanyin Cheng, Xingjian Wu, Chenjuan Guo, Bin Yang, Jilin Hu

A new framework, ASTGI, directly models irregular multivariate time series by encoding observations into a learnable spatio-temporal embedding and adaptively constructing interaction graphs for dependency capture. This approach consistently achieved superior prediction accuracy, reducing Mean Squared Error by approximately 6% on average compared to the best prior methods across multiple datasets.

Key insights

• Learning an adaptive spatio-temporal coordinate space for observations is crucial, as it provides a structured metric space for dynamically identifying relevant neighbors and capturing non-linear data patterns.
• Dynamically computing relation-aware interaction weights based on current features and relative spatio-temporal positions is key to adaptively capturing complex dependencies, including synchronous, long-range, and cross-variable correlations.
• Preserving the raw, undistorted information of irregular time series through direct encoding, combined with adaptive graph interactions, is more effective than methods relying on data interpolation or rigid alignment.

Summary by alphaXiv.