GOOD: A Graph Out-of-Distribution Benchmark Supplementary Material Shurui Gui

GOOD provides 11 datasets with 17 domain selections. For covariate shift splits, given a domain selection, we sort the graphs/nodes by their domains and divide the data into a certain number of domains by specifying the split ratio. Consequently, to build a specific concept, each graph has a domain-label probability to be included in this concept. Therefore, we build each concept by scanning the whole dataset and selecting graphs to be included according to their probabilities. Similarly, in node classification tasks, we apply the screening process to nodes instead of graphs.