UniSTOK: A Framework for Inductive Spatio-Temporal Kriging with Incomplete Data
Researchers have introduced UniSTOK, a novel plug-and-play framework designed to enhance inductive spatio-temporal kriging, a technique used to infer signals at unobserved locations from sensor data. Real-world sensor observations often suffer from block-wise missingness due to failures or maintenance, causing issues for traditional impute-then-krige pipelines where reconstruction errors propagate bias. UniSTOK addresses these challenges through two key innovations: Reliability-guided Signal Regulation (RSR) and Residual Bias Calibration (RBC). RSR estimates entry-wise reliability based on temporal continuity and spatial support, emphasizing reliable observations while suppressing those with long gaps or weak support before spatial propagation. RBC further refines predictions by estimating value-conditioned residual prototypes to adaptively calibrate systematic over- or under-estimations after the main predictor converges. Extensive experiments on real-world datasets demonstrate that UniSTOK consistently improves the performance of multiple kriging backbones, offering a robust solution for handling incomplete observational data in spatial-temporal analysis.
Wire timeline
UniSTOK: A Framework for Inductive Spatio-Temporal Kriging with Incomplete Data
Researchers have introduced UniSTOK, a novel plug-and-play framework designed to enhance inductive spatio-temporal kriging, a technique used to infer signals at unobserved locations from sensor data. Real-world sensor observations often suffer from block-wise missingness due to failures or maintenance, causing issues for traditional impute-then-krige pipelines where reconstruction errors propagate bias. UniSTOK addresses these challenges through two key innovations: Reliability-guided Signal Regulation (RSR) and Residual Bias Calibration (RBC). RSR estimates entry-wise reliability based on temporal continuity and spatial support, emphasizing reliable observations while suppressing those with long gaps or weak support before spatial propagation. RBC further refines predictions by estimating value-conditioned residual prototypes to adaptively calibrate systematic over- or under-estimations after the main predictor converges. Extensive experiments on real-world datasets demonstrate that UniSTOK consistently improves the performance of multiple kriging backbones, offering a robust solution for handling incomplete observational data in spatial-temporal analysis.
cs.AI updates on arXiv.org