Research map / Graph learning / GNN architectures
Subgraph Invariant Learning Towards Large-Scale Graph Node Classification
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 subgraph invariant learning for large-scale node classification, training on labeled subgraphs with an InVar loss and the IFM architecture (IRE + NRE) to generalize to other subgraphs.
Problem
Large-scale graph node classification with GNNs is computationally expensive, and graph reduction methods (sparsification, coarsening, condensation) need the full large graph during training and generalize poorly out-of-distribution.
Approach
Treat a large graph as many small dense subgraphs (small-world view), derive a lower bound on invariance log-likelihood yielding the InVar loss (InfoNCE-style alignment of node embeddings to same-label invariance representations plus an entropy term), and implement it with IFM, whose IRE compresses long-range same-label information via attention and whose NRE injects it into any base GNN via TeleA
Key results
- GCN-IFM raises GCN's weighted AUC-ROC on Reddit from 84.74 to 88.28.
- SAGE-IFM reaches 91.18 AUC-ROC on Reddit versus 89.07 for SAGE, with reduced std.
- UniMP-IFM improves COCO-SP AUC-ROC from 85.94 to 86.47 and F1 from 73.36 to 74.53.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IFM to SAGE | SAGE · Reddit | AUC-ROC: 89.07±14.55 -> 91.18±0.03 | Helped |
| Add IFM to SAGE | SAGE · Reddit | F1: 82.15±24.03 -> 82.31±23.89 | Helped |
| Add IFM to SAGE | SAGE · Amazon | AUC-ROC: 97.05±0.43 -> 97.11±0.41 | Hurt |
| Add IFM to SAGE | SAGE · Amazon | F1: 70.85±1.83 -> 71.06±1.71 | Helped |
| Add IFM to UniMP | UniMP · PascalVOC | AUC-ROC: 85.79±14.30 -> 85.89±14.30 | Helped |
| Add IFM to UniMP | UniMP · PascalVOC | F1: 72.74±20.05 -> 72.88±19.97 | Helped |
| Add IFM to UniMP | UniMP · COCO | AUC-ROC: 85.94±16.32 -> 86.47±16.15 | Helped |
| Add IFM to GATv2 | GATv2 · COCO | F1: 72.63±22.78 -> 74.19±21.65 | Helped |
| Add entropy loss LE (ablation) | GCN-IFM · Reddit | AUC-ROC: 85.65±9.93 -> 86.36±8.88 | Helped |
| Add entropy loss LE (ablation) | GCN-IFM · Reddit | F1: 71.79±29.57 -> 73.61±28.75 | Helped |
| Add entropy loss LE (ablation) | GAT-IFM · Reddit | AUC-ROC: 84.83±10.42 -> 85.28±11.12 | Helped |
| Add entropy loss LE (ablation) | GAT-IFM · Reddit | F1: 68.02±29.71 -> 71.07±29.56 | Helped |
Surprising findings
- On Amazon, IFM gains are negligible (e.g., GCN 97.08 vs GCN-IFM 97.05 AUC-ROC), with some models slightly lower.
Limitations
- Gains on Amazon are marginal or absent, and differences on several datasets are within large standard deviations.
- Efficiency claims are analytical (complexity) rather than measured wall-clock comparisons in the paper.
Experimental setup
Datasets: Reddit Amazon PascalVOC-SP COCO-SP
Benchmarks: Reddit Amazon PascalVOC-SP COCO-SP
Metrics: weighted AUC-ROC weighted macro F1
Baselines: GCN GraphSAGE SGC GAT GATv2 UniMP
Builds on: InfoNCE graph invariant learning Transformer attention GraphSAINT
Closest papers by meaning
- Translating Subgraphs to Nodes Makes Simple GNNs Strong and Efficient for Subgraph Representation Learning (ICML 2024 · similarity 0.90)
- A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening (NeurIPS 2024 · similarity 0.90)
- An Efficient Subgraph-Inferring Framework for Large-Scale Heterogeneous Graphs (AAAI 2024 · similarity 0.89)
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.89)
- Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks (NeurIPS 2024 · similarity 0.88)
- Union Subgraph Neural Networks (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.