GLGAT: A Graph Attention Network for Traffic Forecasting

Researchers have introduced the Global-Local Graph Attention Network (GLGAT) to enhance traffic forecasting by effectively addressing spatio-temporal correlations. Unlike conventional graph convolutional and attention networks, GLGAT differentiates vertices through pairwise encoding and utilizes an event-based adjacency matrix. It employs a global attention matrix for the entire graph while assigning local attention matrix sets to individual vertices. Tests conducted on two real-world traffic datasets reveal that GLGAT successfully captures spatio-temporal correlations and performs competitively compared to leading baselines. This study is available on arXiv in the Computer Science > Artificial Intelligence category.