Learning Invariant Representations of Graph Neural Networks via Cluster Generalization

In this paper, we experimentally find that the performance of GNNs drops significantly when the structure shift happens, suggesting that the learned models may be biased towards specific structure patterns.