Core Concepts
Revisiting spatiotemporal time series imputation from a causal perspective, Casper introduces a novel approach to discover causal relationships and improve imputation accuracy.
Abstract
Spatiotemporal time series with missing values are common, impacting data analysis. Casper addresses confounders and non-causal correlations, introducing a Causality-Aware Spatiotemporal Graph Neural Network. By blocking confounders and using Spatiotemporal Causal Attention, Casper outperforms baselines in discovering causal relationships.
Stats
Theoretical analysis reveals that SCA discovers causal relationships based on the values of gradients.
Experimental results show that Casper significantly outperforms baselines in discovering causal relationships.