Semantic Knowledge Boosts Spatio-Temporal Traffic Forecasting

Graph Neural Networks (GNNs) have significantly advanced spatio-temporal traffic forecasting, but their reliance solely on sensor proximity or road-network topology limits their full potential. This research proposes a novel framework designed to integrate diverse forms of external knowledge, thereby enriching the contextual understanding of the environment at the sensor level. The framework leverages general-purpose knowledge graphs, such as Wikidata, to construct semantic subgraphs around traffic sensors. These subgraphs are then used to generate knowledge graph embeddings, which encode meaningful relationships like nearby points of interest, administrative boundaries, and the functional roles of locations. These semantic embeddings are subsequently fused with conventional traffic sensor graphs, creating additional adjacency matrices that are informed by this rich semantic context. Crucially, this study demonstrates the general impact of external knowledge on prediction accuracy rather than introducing a new GNN architecture. Experiments with established traffic forecasting methods confirm that integrating data from general-purpose knowledge graphs significantly enhances prediction accuracy, providing information that street network data alone cannot convey. This approach offers a promising pathway towards more accurate and potentially more interpretable traffic forecasting models.