Research map / Graph learning / Graph self-supervised learning
Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing
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 M3P-GCL, a graph contrastive learning framework that combines structural and attribute kNN views with aligned priority/supporting view encoding (macro level) and adaptive self-propagation of self-loops in an SGC encoder (micro level), evaluated on homophilous and non-homophilous node classification.
Problem
Graph contrastive learning relies on GNN encoders with a homophily assumption, which restricts the direction and process of message passing and hurts performance on non-homophilous graphs.
Approach
Uses both structural and attribute-kNN views, assigns priority/supporting roles depending on graph homophily, aligns the supporting view embeddings with a global embedding of the priority view (APS-VE), and splits self-propagation into neighbor-like and identity (self-individual) parts weighted by omega (ASP) within an SGC encoder.
Key results
- On Cornell, M3P-GCL reaches 85.03% accuracy, improving accuracy by 13.08% over the best baseline per the paper's text.
- On Cora, M3P-GCL achieves 85.62% accuracy, a 1.45% gain over the strongest baseline (ASP at 84.17%).
- Ablation removing the structural view drops Cora accuracy to 63.08%, and removing the attribute view drops Cornell accuracy to 49.78%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use M3P-GCL instead of the strongest baseline | Node classification · Cornell | accuracy (%): 71.95±2.12 -> 85.03±1.23 | Helped |
| Use M3P-GCL instead of the strongest baseline | Node classification · Texas | accuracy (%): 76.86±3.80 -> 81.08±0.00 | Helped |
| Use M3P-GCL instead of the strongest baseline | Node classification · Wisconsin | accuracy (%): 77.65±2.00 -> 82.35±0.00 | Helped |
| Use M3P-GCL instead of the strongest baseline | Node classification · Actor | accuracy (%): 35.67±0.51 -> 35.48±0.82 | Hurt |
| Use M3P-GCL instead of ASP | Node classification · Cora | accuracy (%): 84.17 -> 85.62 | Helped |
| Use M3P-GCL instead of the strongest baseline | Node classification · PubMed | accuracy (%): 80.63 -> 80.44 | Hurt |
| Add the attribute view (ablation) | M3P-GCL · Cornell | accuracy (%): 49.78 -> 85.03 | Helped |
| Add the structural view (ablation) | M3P-GCL · Wisconsin | accuracy (%): 72.97 -> 82.35 | Helped |
| Add APS-VE (ablation) | M3P-GCL · Cornell | accuracy (%): 64.65 -> 85.03 | Helped |
| Add ASP (ablation) | M3P-GCL · Cornell | accuracy (%): 58.38 -> 85.03 | Helped |
Surprising findings
- Optimal self-loop role weight omega tends to be at an extreme (0 or 1), and combining dual roles could undermine performance on non-homophilous graphs.
Limitations
- Results on Actor are only modestly better than baselines.
- Performance on non-homophilous graphs is sensitive to kNN size k and the self-loop weight omega on small datasets.
Experimental setup
Datasets: Cora CiteSeer PubMed Cornell Texas Wisconsin Actor
Benchmarks: Cora CiteSeer PubMed Cornell Texas Wisconsin Actor
Metrics: accuracy F1 score
Baselines: MLP GCN SGC H2GCN UGCN FAGCN CSGCL NCLA HSAN HomoGCL ASP GREET
Builds on: SGC InfoNCE ASP GREET HomoGCL
Closest papers by meaning
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.92)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.92)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.91)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- StructComp: Substituting propagation with Structural Compression in Training Graph Contrastive Learning (ICLR 2024 · similarity 0.90)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
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.