PemsSJ: A Curated Graph Version of the PeMS Data for the San Jose Bay Area

A curated graph derived from the California Performance Measurement System (PeMS) road network (Chen et al., 2001). Nodes represent traffic sensors recording vehicle speeds; edges represent road connectivity between them. Graph statistics: 1016 nodes, 1172 edges. The topology is notably elongated and sparse (low average degree, high diameter-to-node ratio), making it a challenging benchmark even among road networks. Task: Node regression — predict vehicle speed at designated test nodes given observed speeds at training nodes. Input features are constructed by propagating training targets, analogously to label propagation. Example Usage: This dataset has been used as a benchmark for Gaussian processes on graphs (Borovitskiy et al., AISTATS 2021) and uncertainty quantification with graph neural networks (C. Vieira et al., 2026). Concrete Snapshot: The curated snapshot corresponds to the South Bay / San Jose corridor (Latitude: 37.210 to 37.450 | Longitude: -122.100 to -121.800) on 2th of January 2017 at 17:30. Files edge_index.pt — edge connectivity in COO format (2, E) edge_features.pt — edge attributes node_coordinates.pt — spatial coordinates of each sensor node (N, 2), longitude and latitude respectively valid_label_index_data.npz — node targets and split information Consult the source code to obtain the raw data and understand how the data is processed.