Table 1.

Summary of related studies

Study	Description	Model	Issues
Geng et al. (2019)	•region-level ride-hailing demand forecasting •encode the non-Euclidean correlations among regions using multiple graphs and capture correlations using multi-graph convolution •propose contextual gated RNN considering global contextual information for temporal dependencies	Spatio-temporalmulti-graph convolution network	model complexity; overfitting
Guo et al. (2019)	•traffic flow forecasting •combines the spatial-temporal attention and convolution (i.e. graph convolutions in the spatial dimension and typical convolutions in the temporal dimension)	attention based spatial-temporal graph convolutionalnetwork (ASTGCN)	model complexity; convergence
Lu and Li (2020)	•experiments on three sensor data, air quality, bike demand, and traffic flow •propose AGSTN to characterize the time-evolving spatio-temporal correlation given pre-defined graphs	attention-adjusted graph spatio-temporal network (AGSTN)	pre-defined graphs; model complexity; convergence
Cirstea et al. (2021)	•experiments on two datasets of traffic sensors and one dataset of meteorological stations •propose EnhanceNet with two plugin neural networks including distinct filter generation network and dynamic adjacency matrix generation network to capture distinct temporal dependencies among different entities	EnhanceNet	model complexity
Geng et al. (2022)	•experiments on wind speed, air quality, indoor temperature (SML2010), and traffic volume •propose graph correlated attention recurrent neural network (GCAR) to capture long-term dependencies based on the reliable interaction between historical target and external features.	graph correlated attention recurrent neural network (GCAR)	model complexity