Research map / Graph learning / Graph self-supervised learning
Beyond Redundancy: Information-aware Unsupervised Multiplex Graph Structure Learning
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 InfoMGF, an unsupervised framework that learns a fused, denoised graph from a multiplex graph by refining each view and maximizing view-shared and view-unique task-relevant information via learnable graph augmentation.
Problem
Unsupervised multiplex graph learning assumes reliable graph structure and relies on contrastive learning that only captures view-shared information, missing view-unique task-relevant information and failing to remove task-irrelevant noise in edges.
Approach
Unsupervised graph structure learning for multiplex graphs: refine each view with a graph learner, maximize shared mutual information between views plus mutual information between refined graphs and learned augmented graphs (with a learnable generator that minimizes an upper bound on MI), and fuse views into one graph with a theoretically motivated objective.
Key results
- On node clustering, InfoMGF-LA reaches 76.53 NMI and 93.45 ACC on ACM and 75.18 NMI on Yelp, outperforming the best multiplex baselines (BTGF, DMG).
- On node classification, InfoMGF-LA achieves 93.42 Macro-F1 on ACM and 93.26 Macro-F1 on Yelp, exceeding supervised GSL methods such as GRCN and NodeFormer.
- Robustness analysis on ACM shows InfoMGF consistently outperforms GCN, SUBLIME and HDMI under random edge addition and deletion, especially at high perturbation rates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InfoMGF-LA instead of the strongest reported baseline | Node clustering · ACM, DBLP, Yelp; InfoMGF-LA is OOM on MAG | ACC: BTGF 90.09 -> InfoMGF-LA 93.45 (ACM); DMG 88.45 -> 91.08 (DBLP); BTGF 91.39 -> 93.26 (Yelp) | Helped |
| Use InfoMGF-RA instead of the strongest reported baseline | Node clustering · MAG | ACC: MGDCR 61.37 -> InfoMGF-RA 64.13 | Helped |
| Use InfoMGF-LA instead of the strongest reported baseline | Node classification · ACM, DBLP, Yelp; InfoMGF-LA is OOM on MAG | Macro-F1: GRCN 93.04$\pm$0.17 -> InfoMGF-LA 93.42$\pm$0.21 (ACM); SUBLIME 90.98$\pm$0.37 -> 91.28$\pm$0.31 (DBLP); BTGF 92.81$\pm$ | Helped |
| Use InfoMGF-RA instead of the strongest reported baseline | Node classification · MAG | Macro-F1: NodeFormer 77.21$\pm$0.18 -> InfoMGF-RA 77.25$\pm$0.06 | Helped |
| Add shared-information loss $\mathcal{L}_s$ (ablation) | InfoMGF · ACM, DBLP, Yelp | Macro-F1: 93.05$\pm$0.49 -> 93.42$\pm$0.21 (ACM); 90.44$\pm$0.45 -> 91.28$\pm$0.31 (DBLP); 93.15$\pm$0.12 -> 93.26$\pm$0.26 (Yelp) | Helped |
| Add task-relevant information loss $\mathcal{L}_u$ (ablation) | InfoMGF · ACM, DBLP, Yelp | Macro-F1: 92.66$\pm$0.53 -> 93.42$\pm$0.21 (ACM); 90.13$\pm$0.43 -> 91.28$\pm$0.31 (DBLP); 92.23$\pm$0.27 -> 93.26$\pm$0.26 (Yelp) | Helped |
| Add graph augmentation instead of maximizing mutual information with the original graph (ablation) | InfoMGF · ACM, DBLP, Yelp | Macro-F1: 92.84$\pm$0.17 -> 93.42$\pm$0.21 (ACM); 90.94$\pm$0.45 -> 91.28$\pm$0.31 (DBLP); 92.76$\pm$0.49 -> 93.26$\pm$0.26 (Yelp) | Helped |
| Add the reconstruction loss to the augmentation generator (ablation) | InfoMGF · ACM, DBLP, Yelp | Macro-F1: 92.91$\pm$0.53 -> 93.42$\pm$0.21 (ACM); 91.05$\pm$0.27 -> 91.28$\pm$0.31 (DBLP); 92.65$\pm$0.27 -> 93.26$\pm$0.26 (Yelp) | Helped |
Surprising findings
- The unsupervised InfoMGF outperforms supervised GSL methods (e.g., GRCN, IDGL, ProGNN) on node classification.
- Removing the unique-information loss L_u hurts more than removing the shared-information loss L_s.
Limitations
- Focuses only on the pure unsupervised setting; does not exploit partial labels.
- Theoretical guarantees rely on an optimal augmentation assumption that cannot be achieved without labels; heuristic augmentations only approximate it.
- Evaluated on four multiplex datasets with node-level tasks only.
Experimental setup
Datasets: ACM DBLP Yelp MAG
Benchmarks: ACM DBLP Yelp MAG
Metrics: NMI ARI ACC F1 Macro-F1 Micro-F1
Baselines: VGAE DGI O2MAC MvAGC MCGC HDMI MGDCR DMG BTGF GCN GAT HAN
Builds on: SUBLIME contrastive multi-view non-redundancy theory Gumbel-Max reparameterization SGC GCN
Closest papers by meaning
- Multiplex Graph Representation Learning with Homophily and Consistency (AAAI 2025 · similarity 0.90)
- GraphFLEx: Unsupervised Structure Learning $\underline{\text{F}}$ramework for $\underline{\text{L}}$arge $\underline{\text{Ex}}$panding $\underline{\text{Graph}}$s (ICML 2026 · similarity 0.89)
- Unsupervised Attributed Multiplex Network Embedding (AAAI 2020 · similarity 0.89)
- Disentangled Multiplex Graph Representation Learning (ICML 2023 · similarity 0.88)
- Hypergraph Learning for Unsupervised Graph Alignment via Optimal Transport (AAAI 2025 · similarity 0.86)
- When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors (NeurIPS 2026 · similarity 0.85)
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.