Inductive Subgraphs as Shortcuts: Causal Disentanglement for Heterophilic Graph Learning

Xiangmeng Wang, Qian Li, Haiyang Xia, Hao Miao, Qing Li, Guandong XuarXiv 2026 · DOI · arXiv

TL;DR

Standard Graph Neural Networks (GNNs) frequently fail on heterophilic graphs—where connected nodes belong to different classes—because they fall prey to "shortcut learning" from recurring local subgraphs. This paper introduces Causal Disentangled GNN (CD-GNN), a novel framework that leverages causal inference to isolate and block these misleading structural shortcuts, ensuring the model relies strictly on true causal features for robust node classification.

What Problem Does This Paper Address?

Most conventional GNNs operate under the assumption of homophily ("birds of a feather flock together"), meaning connected nodes are expected to share similar attributes and labels. However, many real-world networks are highly heterophilic. For example, in telecom fraud networks, scammers predominantly link with legitimate users rather than other scammers.

When standard GNNs encounter heterophilic graphs, their performance plummets. While prior research has attempted to fix this by expanding neighborhood receptive fields or redesigning message-passing architectures, the underlying reason for these failures remained poorly understood.

This paper uncovers a structural culprit: inductive subgraphs acting as malignant shortcuts. In homophilic settings, recurring local structural motifs (like a cycle or a grid) provide helpful, benign biases for classification. In heterophilic settings, however, GNNs easily overfit to these same recurring structural patterns, treating them as shortcuts to minimize training loss without learning the actual causal relationships driving the node labels.

Figure 1: A toy example shows that an inductive subgraph helps a homophilic graph infer correct labels, while harming a heterophilic graph for false label prediction.

Methods

To systematically address shortcut learning, the authors model the graph generation process using a Structural Causal Model (SCM). They identify two primary paths that leak bias into GNN predictions:

  1. A Confounding Path: Where a shortcut subgraph acts as an observed confounder between the true causal features and the final prediction.
  2. A Spillover Path: Where network interference causes shortcut features and causal features to become entangled during message passing.

To break these non-causal paths, the authors propose CD-GNN. Instead of altering the underlying GNN architecture, CD-GNN introduces generative probabilistic edge and feature masks to decompose a node's neighborhood into two distinct components: a causal subgraph ($G_c$) and a shortcut subgraph ($G_s$).

Figure 4: (a) Causal graph of existing GNN training. (b) Our proposed debiased causal graph.

The framework optimizes these subgraphs via three key mechanisms:

  • Shortcut Amplification: The shortcut branch is intentionally trained using a Generalized Cross-Entropy (GCE) loss. Because shortcuts are inherently easier to learn, this loss amplifies their gradients, quickly isolating them into the shortcut branch.
  • Causal Priority Reweighting: The causal branch uses a relative difficulty score to up-weight hard-to-learn, truly informative causal patterns, preventing the model from relying on easy shortcut cues.
  • Spillover Control: To ensure complete independence between the two branches, the authors apply a Hilbert-Schmidt Independence Criterion (HSIC) regularizer alongside counterfactual shortcut embeddings generated by randomly permuting shortcut representations across mini-batches.

Datasets

The authors evaluated their approach using both controlled synthetic benchmarks and real-world graphs:

  • Synthetic Benchmarks (with controlled heterophily levels): Tree-Cycles, Tree-Grid, BA-Shapes, and BA-Community.
  • Real-World Heterophilic Graphs: Chameleon, Squirrel, Cornell, Roman-empire, and Amazon-ratings.
  • Real-World Homophilic Graphs: Computers and Questions.

Key Insights and Ideas

  • Shortcuts are Context-Dependent: The exact same local motif (e.g., a house structure) that acts as a benign cue in a homophilic graph becomes a malignant distractor in a heterophilic graph due to shifting neighborhood label distributions.
  • Embrace the Shortcut to Eliminate It: Rather than trying to suppress shortcut learning altogether, CD-GNN dedicates a specific branch to amplify and capture the shortcut. By isolating what is "easy" to learn, the remaining network is forced to uncover the true causal drivers.
  • Causal Graph Realignment: The paper shifts the conversation around heterophily from node-level dynamics (like oversmoothing or high/low-pass filtering) to a structural, causal-inference perspective.

Key Results

CD-GNN achieved state-of-the-art performance, ranking first on six out of seven real-world benchmarks and significantly outperforming ten competitive baselines:

  • Substantial Gains under High Heterophily: On the highly heterophilic Squirrel dataset ($h_L = 0.7928$), CD-GNN achieved an accuracy of 0.5898, vastly outperforming standard GCN (0.5024) and beating the best heterophily-aware baseline, CAGNN (0.4076).
  • Robustness Across the Board: On Chameleon, CD-GNN reached an accuracy of 0.6762 (compared to GCN's 0.6031). It also maintained top-tier performance on homophilic datasets (Computers and Questions), proving that debiasing does not come at the cost of standard homophilic accuracy.

Limitations and Future Work

  • Sensitivity to Graph Scale: CD-GNN achieved runner-up status on the Cornell dataset. The authors attribute this to Cornell's very small scale (183 nodes), which provides insufficient structural variety for the generative masking model to cleanly separate causal and shortcut factors.
  • Computational Complexity: The inclusion of the full-matrix HSIC regularizer introduces an $O(N^2d)$ time complexity. While manageable via mini-batching on typical graph scales, scaling this causal disentanglement framework to massive, web-scale networks remains an open challenge for future engineering.

Summary updated

Abstract

Heterophily is a prevalent property of real-world graphs and is well known to impair the performance of homophilic Graph Neural Networks (GNNs). Prior work has attempted to adapt GNNs to heterophilic graphs through non-local neighbor extension or architecture refinement. However, the fundamental reasons behind misclassifications remain poorly understood. In this work, we take a novel perspective by examining recurring inductive subgraphs, empirically and theoretically showing that they act as spurious shortcuts that mislead GNNs and reinforce non-causal correlations in heterophilic graphs. To address this, we adopt a causal inference perspective to analyze and correct the biased learning behavior induced by shortcut inductive subgraphs. We propose a debiased causal graph that explicitly blocks confounding and spillover paths responsible for these shortcuts. Guided by this causal graph, we introduce Causal Disentangled GNN (CD-GNN), a principled framework that disentangles spurious inductive subgraphs from true causal subgraphs by explicitly blocking non-causal paths. By focusing on genuine causal signals, CD-GNN substantially improves the robustness and accuracy of node classification in heterophilic graphs. Extensive experiments on real-world datasets not only validate our theoretical findings but also demonstrate that our proposed CD-GNN outperforms state-of-the-art heterophily-aware baselines.
BibTeX
@misc{wang2026inductive,
  author = {Xiangmeng Wang and Qian Li and Haiyang Xia and Hao Miao and Qing Li and Guandong Xu},
  title = {{Inductive Subgraphs as Shortcuts: Causal Disentanglement for Heterophilic Graph Learning}},
  year = {2026},
  eprint = {2604.19186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.19186},
}

Loading related content...

Figure