Fig. 2
From: Quantifying the non-isomorphism of global urban road networks using GNNs and graph kernels
Graph Classification Models and Their Accuracy. (a) Street map of central London. (b) The road network, utilized for GNNs. The node attributes are latitude and longitude coordinates normalized to the [0, 1] interval. (c) GNN. The embedding representations of all nodes are processed by torch.mean(). (d) Test accuracy of graph classification. The EdgeCNN achieves an accuracy close to 85%. The WL kernel reaches nearly 80%, while the GIN exceeded 75%. (e) The topology of the road network, used for the WL kernel. The node labels are assumed identical. (f) WL kernel. The node labels (colors) are iteratively refined through color refinement. (g) Graph classification accuracy for 30 cities obtained through the EdgeCNN. Maps a, b, and e are copyrighted by OpenStreetMap contributors.
