Causality-Aware Spatiotemporal Graph Neural Networks for Spatiotemporal Time Series Imputation

Area of Science:

• Data Science
• Machine Learning
• Causal Inference

Background:

• Spatiotemporal time series data often suffer from missing values due to sensor failures.
• Existing imputation methods may overfit by using non-causal correlations introduced by confounders.

Purpose of the Study:

• To propose a causality-based approach for spatiotemporal time series imputation.
• To develop a novel neural network model that accounts for causal relationships.

Main Methods:

• Revisiting spatiotemporal imputation from a causal perspective using frontdoor adjustment.
• Introducing the Causality-Aware Spatiotemporal Graph Neural Network (Casper) with a Prompt Based Decoder (PBD) and Spatiotemporal Causal Attention (SCA).

Main Results:

• Casper effectively reduces the impact of confounders and identifies sparse causal relationships.
• Theoretical analysis shows SCA discovers causal links via gradient values.
• Experimental results demonstrate Casper's superior performance over baseline methods on real-world datasets.

Conclusions:

• Casper provides an effective and robust solution for spatiotemporal time series imputation by leveraging causal inference.
• The model successfully mitigates overfitting issues caused by non-causal correlations.