A causal graph convolutional network considering missing values for spatio-temporal prediction

Abstract

Spatio-temporal prediction is one of the basic research topics ot geographic spatio-temporal big data mining. There are many attempts to predict spatio-temporal state of unknown systems using various deep learning algorithms. However, most existing prediction models are only tested on spatio-temporal data assuming no missing data entries, ignoring the impact of missing values on the prediction results. In the actual scenarios, data missing is an inevitable problem due to sensor or network transmission failures. Therefore, we propose a novel causal graph convolutional network considering missing values (Causal-GCNM) for spatio-temporal prediction. The proposed model can automatically capture missing patterns in the spatio-temporal data, enabling the Causal-GCNM model to directly complete the spatio-temporal prediction task without additional interpolation. The proposed model was validated on three real spatiotemporal datasets (traffic flow dataset, PM2.5 monitoring dataset, and temperature monitoring dataset). Experimental results show that the Causal-GCNM model has good prediction performance under four missing scenarios (20% random missing, 20% block missing, 40% random missing, 40% block missing), and outperforms ten existing baseline methods in terms of prediction accuracy and computational efficiency.