Inductive Subgraphs as Shortcuts: Causal Disentanglement for Heterophilic Graph Learning

The core contribution here is diagnostic as much as corrective: the paper argues that GNNs don't just struggle with heterophilic graphs due to architecture mismatch, but because they actively learn the wrong patterns, recurring substructures that correlate with labels in training data but mislead on out-of-distribution nodes. That reframing matters for how you'd even evaluate a fix.

The GNN embedding benchmarking work covered here in mid-April ('How Embeddings Shape Graph Neural Networks') tested classical versus quantum-oriented node representations but held architecture constant, which means it was measuring a different variable entirely. Neither study speaks directly to the other, but together they sketch a picture of the field stress-testing GNNs from multiple angles simultaneously: representation quality on one side, structural bias on the other. The causal debiasing framing in this paper is relatively unusual in the GNN literature and sits closer to the fairness and robustness research tradition than to standard architecture tuning.

The real test is whether this causal framework holds on benchmark graphs with known heterophily ratios, like Actor or Cornell, under distribution shift conditions that weren't in the training split. If the debiasing gains collapse when subgraph patterns shift between train and test, the shortcut problem is being managed rather than solved.