Attention to Traffic Forecasting: Improving Predictions with Temporal Graph Attention Networks
Dynamic traffic flow forecasting remains an open issue to this day. As other spatio-temporal problems, traffic prediction deals with both temporal and spatial nonlinear relationships, with the particularity that nearby points in the Euclidean space might be allocated in different roads, adding another layer of complexity. Traffic prediction has witnessed a revolution with the appearance of deep learning, with graph neural networks being prominently responsible for a steep increase in forecasting accuracy. In this paper, we consider the use of an automatic attention mechanism in order to improve the prediction capabilities of a traffic graph convolutional network. This model is based on the composition of gated recurrent units and graph convolution networks to model space and time simultaneously. To overcome the spatial modelling limitations of the original model, our proposal replaces the graph convolutional layer with a graph attention mechanism. Our aim is to model spatial relations in an automatic, more dynamic way. In order to prove the validity and usefulness of our proposal, we have performed a thorough experimentation over two known traffic datasets used in previous research, plus a new, complex one which we have curated and published. Our results portray a clear and statistically significant advantage with the inclusion of spatial attention, surpassing the performance of a wide set of state-of-the-art models on every tested scenario.