Robust Causal Graph Representation Learning against Confounding Effects

The prevailing graph neural network models have achieved significant progress in representation learning. However, this paper, we uncover an ever-overlooked phenomenon: the pre-trained learning model tested with full graphs underperforms well-pruned graphs. This observation reveals that there exist confounders graphs, which may interfere semantic information, and current methods not eliminated their influence. To tackle issue, propose Robust Causal Graph Representation Learning (RCGRL) to learn robust representations against confounding effects. RCGRL introduces active approach generate instrumental variables under unconditional moment restrictions, empowers eliminate confounders, thereby capturing discriminative information is causally related downstream predictions. We offer theorems proofs guarantee theoretical effectiveness of proposed approach. Empirically, conduct extensive experiments on a synthetic dataset multiple benchmark datasets. Experimental results demonstrate generalization ability RCGRL. Our codes are available at https://github.com/hang53/RCGRL.