Attention-Based Spatio-Temporal Graph Convolutional Networks – A Systematic Review

Spatio-temporal attention-based Graph Convolutional Networks will be the subject of the current review. The aim is to identify the status quo of the research done so far in this area and identify knowledge gaps still to be covered.

Graph Convolutional Networks (GCN), which is one of the most powerful deep learning algorithms, is still a rather emerging field with applications in traffic modelling. GCNs are an important aspect of Intelligent Transportation System.

Since traffic flows are connected with spatial and temporal correlations of multiple factors, Spatio-temporal GCN were introduced. They were meant to address the fact that traffic flow is highly non-linear and not time- and spacially independent. The idea of the attention mechanism is that more attention and thus more weight is given to the information bits that contribute the most to the prediction precision. Later attention-based method was applied to solve temporal problems.

To identify relevant studies, we utilised the following academic search engines: TRID, Scopus, IEEE Xplore, IET Digital Library (search by titles and abstracts), Google Scholar and Science Direct (full-text search). Since the model under interest is a rather new method, the reviewed bibliographic database covers the period from 2017 to 2022. The focus was made on empirical studies with real or simulated data. Gaps were identified, research directions were classified and described.