Research map / Graph learning / Graph self-supervised learning

Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Graph out-of-distribution generalization

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VIVACE, which learns invariant subgraphs for graph OOD generalization without environment labels by contrastively estimating spurious correlations on variant subgraphs with a variance regularizer and reweighting invariant predictions via inverse propensity weighting.

Problem

Graph neural networks fail to generalize under distribution shifts with severe bias, and existing graph invariant learning methods depend on predefined or generated environment labels that are unavailable or inaccurate.

Approach

Disentangle invariant and variant subgraphs with a learnable edge-mask GNN, estimate spurious correlations by self-supervised contrastive learning on variant subgraphs with a label-grouped variance regularizer, and reweight invariant-subgraph predictions with inverse propensity weights, avoiding environment labels.

Key results

• On CMNIST with bias r=0.9, VIVACE reaches 79.46% accuracy versus 78.14% for the best baseline DisC.
• On MOLSIDER VIVACE improves ROC-AUC by 1.94% over the strongest baseline (62.15 vs 60.21 for DiffPool).
• On MOLHIV VIVACE reaches 78.11 ROC-AUC, a 1.06-point gain over the strongest baseline (DIR at 77.05).

Measured effects reported in the paper

Surprising findings

• DIR collapses to near chance (about 10% accuracy) on the CMNIST, CFashion and CKuzushiji benchmarks, which the authors attribute to its reliance on non-biased data for interventions.

Limitations

• Relies on the assumption that the variance regularizer cleanly separates invariant from variant information.
• Evaluated only on graph classification benchmarks with synthetic colored-background bias and two OGB molecular datasets.

Experimental setup

Datasets: CMNIST CFashion CKuzushiji MOLSIDER MOLHIV

Benchmarks: CMNIST CFashion CKuzushiji MOLSIDER MOLHIV

Metrics: accuracy ROC-AUC

Baselines: GCN GIN FactorGCN DiffPool LDD DIR DisC

Builds on: DisC DIR LDD GCN Generalized cross entropy

Closest papers by meaning

• Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.92)
• Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.91)
• Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · similarity 0.90)
• Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.89)
• Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.89)
• Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.89)