Research map / Graph learning / Graph self-supervised learning
Multiplex Graph Representation Learning with Homophily and Consistency
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 MGHC, an unsupervised multiplex graph representation learning method that restructures multi-order neighbor relations via ASGC-style correlation coefficients to handle heterophily, and uses a self-expression-matrix-based contrastive loss for node-level and class-level consistency.
Problem
Unsupervised multiplex graph representation learning ignores heterophily in graph structure and only enforces node-level cross-graph consistency, missing class-level consistency.
Approach
Restructure each graph's adjacency using learned multi-order correlation coefficients (positive for homophilic, negative for heterophilic neighbors) in the structure preservation loss, and build a contrastive loss whose positives/negatives come from a closed-form self-expression matrix so same-class nodes across graphs are pulled together.
Key results
- MGHC achieves best node classification on all four datasets, e.g. 92.3 Macro-F1 on ACM and 57.8 Macro-F1 on IMDB, versus 92.0 for the best prior CoCoMG on ACM.
- Compared with the best unsupervised baseline DMG, MGHC improves by about 0.2% on average.
- Ablation: full loss improves 17.4% over contrastive loss alone and 2.8% over structure preservation loss alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add structure preservation loss to contrastive loss (ablation) | MGHC · ACM | Macro-F1: 78.0 ± 0.4 -> 92.3 ± 0.3 | Helped |
| add structure preservation loss to contrastive loss (ablation) | MGHC · DBLP | Macro-F1: 64.7 ± 0.4 -> 93.0 ± 0.3 | Helped |
| add structure preservation loss to contrastive loss (ablation) | MGHC · IMDB | Macro-F1: 48.3 ± 0.3 -> 57.8 ± 0.3 | Helped |
| add structure preservation loss to contrastive loss (ablation) | MGHC · Freebase | Macro-F1: 43.4 ± 0.3 -> 61.2 ± 0.5 | Helped |
| add contrastive loss to structure preservation loss (ablation) | MGHC · ACM | Macro-F1: 90.4 ± 0.3 -> 92.3 ± 0.3 | Helped |
| add contrastive loss to structure preservation loss (ablation) | MGHC · DBLP | Macro-F1: 89.1 ± 0.6 -> 93.0 ± 0.3 | Helped |
| add contrastive loss to structure preservation loss (ablation) | MGHC · IMDB | Macro-F1: 53.3 ± 0.5 -> 57.8 ± 0.3 | Helped |
| add contrastive loss to structure preservation loss (ablation) | MGHC · Freebase | Macro-F1: 60.3 ± 1.1 -> 61.2 ± 0.5 | Helped |
| use MGHC instead of CoCoMG | multiplex graph node classification · ACM | Macro-F1: 92.0 ± 0.2 -> 92.3 ± 0.3 | Helped |
| use MGHC instead of DMG | multiplex graph node classification · Freebase | Micro-F1: 65.6 ± 0.7 -> 64.1 ± 0.6 | Hurt |
Surprising findings
- Average gain over the best unsupervised baseline DMG is only about 0.2%.
Limitations
- Gains over the strongest unsupervised baseline (DMG) are marginal.
- Heterophily ratios are computed from labels, and the method's benefit is evaluated only on node classification.
Experimental setup
Datasets: ACM DBLP IMDB Freebase
Benchmarks: ACM DBLP IMDB Freebase
Metrics: Macro-F1 Micro-F1
Baselines: GCN GAT DGI HAN DMGI DMGIattn HDMI HeCo CKD MGDCR DMG CoCoMG
Builds on: ASGC CoCoMG HeCo DMGI
Closest papers by meaning
- Beyond Redundancy: Information-aware Unsupervised Multiplex Graph Structure Learning (NeurIPS 2024 · similarity 0.90)
- Transitivity-Preserving Graph Representation Learning for Bridging Local Connectivity and Role-Based Similarity (AAAI 2024 · similarity 0.88)
- Homophily-Related: Adaptive Hybrid Graph Filter for Multi-View Graph Clustering (AAAI 2024 · similarity 0.87)
- PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering (AAAI 2024 · similarity 0.87)
- Learning Representations for Hierarchies with Minimal Support (NeurIPS 2024 · similarity 0.87)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.87)
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.