Can you name a few datasets commonly used for training GNNs?
Instruction: List several popular datasets used in the GNN community for research and benchmarking.
Updated
Example Answer
I would choose a graph dataset based on the task and the intended generalization setting. Cora, CiteSeer, and PubMed are familiar citation-network benchmarks for node classification, but they are small and their standard fixed-graph splits do not establish generalization to entirely new graphs.
The GraphSAGE Reddit benchmark is a graph of posts, not a generic friendship network. Its task is predicting the community a post belongs to; posts are connected when the same user comments on both. The GraphSAGE PPI benchmark contains protein-interaction graphs and uses multilabel protein-function prediction, with held-out graphs for inductive evaluation.
For a broader benchmark, I would also look at the Open Graph Benchmark and select the specific node, link, or graph task. I would follow its split and evaluator instead of reporting whichever metric looks strongest. The task definition, graph construction, and access to test-node features all matter when comparing results.
Follow-up
Would the deployed model predict unlabeled nodes in a known graph, unseen nodes arriving later, or completely new graphs? Those require different evidence.