Research map / Graph learning / Graph self-supervised learning
GLDL: Graph Label Distribution 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 GLDL, a dual-GCN framework that jointly learns node-node and label-label graphs to predict label distributions for networked data, and builds four graph LDL benchmarks from DBLP, Yelp, Yelp2 and ACM.
Problem
Label distribution learning assumes IID instances, so it ignores instance-instance dependencies and label-label correlations that arise on graph-structured data.
Approach
Induces a label-label network from class-aggregated node features (static or dynamically generated by a parametric graph generator) and co-trains two GCNs with heterogeneous message passing via induced instance-label edges, regularized by consistency and momentum losses on graph Laplacian eigenvalues.
Key results
- On DBLP, GLDL dynamic (GLDLd) achieves CHD 0.2142 and KLD 0.3449 versus GCN-KL 0.2267 and 0.3428 and SA-IIS 0.3491 (CHD), with GLDLd winning 7 of 8 metrics.
- On Yelp2, GLDLd wins 8/8 metrics with CHD 0.2874 versus 0.3021 for GAT-KL and accuracy 0.6022 versus 0.61 for GAT-KL.
- On ACM, GLDLd achieves accuracy 0.726 and weighted F1 0.7008, with CHD 0.3534 versus 0.3849 for GAT-KL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model node-label dependencies with GLDL instead of GCN-KL | GLDLd · DBLP | KLD: 0.3428 -> 0.3449 | No clear effect |
| model node-label dependencies with GLDL instead of GCN-KL | GLDLd · Yelp | KLD: 0.4611 -> 0.3843 | Helped |
| model node-label dependencies with GLDL instead of GAT-KL | GLDLd · Yelp2 | KLD: 0.4214 -> 0.3805 | Helped |
| model node-label dependencies with GLDL instead of GLDLs (ablation) | GLDLd · DBLP | KLD: 0.3455 -> 0.3449 | No clear effect |
| model node-label dependencies with GLDL instead of GLDLs (ablation) | GLDLd · Yelp | KLD: 0.4167 -> 0.3843 | Helped |
| model node-label dependencies with GLDL instead of GLDLs (ablation) | GLDLd · Yelp2 | KLD: 0.3851 -> 0.3805 | Helped |
| model node-label dependencies with GLDL instead of GLDLs (ablation) | GLDLd · ACM | KLD: 0.9093 -> 0.9132 | Hurt |
| use GLDLd instead of GCN-KL | GLDLd · Yelp | accuracy: 0.6133 -> 0.6115 | Hurt |
| use GLDLd instead of GAT-KL | GLDLd · ACM | accuracy: 0.7243 -> 0.726 | Helped |
| use GLDLd instead of GAT-KL | GLDLd · Yelp2 | weighted F1: 0.5033 -> 0.5456 | Helped |
| use GLDLd instead of GAT-KL | GLDLd · ACM | weighted F1: 0.7052 -> 0.7008 | Hurt |
Surprising findings
- Distribution-distance metrics can improve while weighted F1 and accuracy worsen on imbalanced datasets, as seen in over-smoothing experiments.
- SA-IIS achieves good accuracy and F1 on Yelp and Yelp2 but poor distribution measures, attributed to heterophily of those graphs.
Limitations
- Generalization bound depends on label dimension q, graph degree, and GCN depth/width, suggesting degradation for larger label spaces and more complex graphs.
- Results on Yelp and Yelp2 show lower accuracy and F1 than baselines for some variants, and performance gaps are often small.
- Benchmarks are constructed by converting existing heterogeneous graphs into label distributions rather than collected from human-annotated distributions.
Experimental setup
Datasets: DBLP Yelp Yelp2 ACM
Benchmarks: DBLP Yelp Yelp2 ACM
Metrics: Chebyshev distance Cosine distance Canberra distance Clark distance Intersection distance KL divergence accuracy weighted F1 number of top metrics
Baselines: SA-IIS GCN-KL GAT-KL
Builds on: GCN GAT graph condensation LDLLC graph autoencoder
Closest papers by meaning
- HGDL: Heterogeneous Graph Label Distribution Learning (NeurIPS 2024 · similarity 0.92)
- Enhancing the Influence of Labels on Unlabeled Nodes in Graph Convolutional Networks (ICML 2025 · similarity 0.87)
- Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.87)
- Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.86)
- Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.86)
- Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.86)
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.