Research map / Graph learning / GNN architectures
Deep Multi-modal Graph Clustering via Graph Transformer Network
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 DMGC-GTN, a graph-transformer-based deep multi-modal graph clustering method that combines Laplacian-smoothed features, PageRank intimacy, relative position and shortest-path embeddings, trained with reconstruction, structure and KL clustering losses.
Problem
Deep multi-modal graph clustering methods rely on GNNs and under-exploit richer graph structure (node relations, shortest paths) and complementary information across modalities.
Approach
Treats each node as a sequence of its top-t most intimate nodes (PageRank intimacy) and feeds the sum of feature, relative-position and shortest-path embeddings into a Transformer encoder, with Laplacian-smoothed features and adaptive fusion across modalities.
Key results
- On AMAP, DMGC-GTN reaches ACC 0.786, NMI 0.684, ARI 0.599, versus best baseline CMGEC ACC 0.751 on AMAP.
- On Cora, DMGC-GTN reports ACC 0.763 versus CO-GCN 0.735 and CMGEC 0.737 among the strongest baselines.
- Ablation shows feature, position and shortest-path embeddings combined give ACC 0.786 vs 0.652 with feature embeddings alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Lr only, omitting Lg and Lpq (ablation) | DMGC-GTN · AMAP | ACC: 0.768 vs 0.786 | Hurt |
| Omit Lr and use Lg + Lpq (ablation) | DMGC-GTN · AMAP | ACC: 0.663 vs 0.786 | Hurt |
| Omit Lg and use Lr + Lpq (ablation) | DMGC-GTN · AMAP | ACC: 0.763 vs 0.786 | Hurt |
| Omit Lpq and use Lr + Lg (ablation) | DMGC-GTN · AMAP | ACC: 0.772 vs 0.786 | Hurt |
| Add position embeddings to feature and shortest path embeddings (ablation) | DMGC-GTN · AMAP | ACC: 0.659 -> 0.786 | Helped |
| Add shortest path embeddings to feature and position embeddings (ablation) | DMGC-GTN · AMAP | ACC: 0.758 -> 0.786 | Helped |
| Add feature embeddings to position and shortest path embeddings (ablation) | DMGC-GTN · AMAP | ACC: 0.172 -> 0.786 | Helped |
| Add position embeddings to feature embeddings (ablation) | DMGC-GTN · AMAP | ACC: 0.652 -> 0.758 | Helped |
| Add shortest path embeddings to feature embeddings (ablation) | DMGC-GTN · AMAP | ACC: 0.652 -> 0.659 | Helped |
| Use position embeddings only instead of the full embedding combination (ablation) | DMGC-GTN · AMAP | ACC: 0.786 -> 0.145 | Hurt |
| Use shortest path embeddings only instead of the full embedding combination (ablation) | DMGC-GTN · AMAP | ACC: 0.786 -> 0.170 | Hurt |
| Use feature embeddings only instead of the full embedding combination (ablation) | DMGC-GTN · AMAP | ACC: 0.786 -> 0.652 | Hurt |
Surprising findings
- Position and shortest-path embeddings alone perform near chance (ACC about 0.15-0.17), contributing mainly when combined with feature embeddings.
- SGFormer, a graph transformer baseline, performs worse than several GNN-based baselines on these clustering benchmarks.
Limitations
- Evaluation limited to four datasets, with the second modality for three of them synthesized via FFT rather than natively multi-modal.
- Two-stage optimization and hyperparameter sensitivity (lambda1, lambda2, t, layers, smoothing steps) must be tuned per dataset.
Experimental setup
Datasets: AMAP WIKI Citeseer Cora
Benchmarks: AMAP WIKI Citeseer Cora
Metrics: ACC NMI ARI
Baselines: K-means GAE VGAE DAEGC GATE SDCN EGAE NACL DCCA DCCAE CO-GCN SGCMC
Builds on: Graph Autoencoder DAEGC SGFormer DEC clustering loss PageRank
Closest papers by meaning
- Graph Transformer Networks (NeurIPS 2019 · similarity 0.87)
- Cluster-wise Graph Transformer with Dual-granularity Kernelized Attention (NeurIPS 2024 · similarity 0.86)
- DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization (AAAI 2024 · similarity 0.86)
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.86)
- LSEnet: Lorentz Structural Entropy Neural Network for Deep Graph Clustering (ICML 2024 · similarity 0.86)
- Scalable Attribute-Missing Graph Clustering via Neighborhood Differentiation (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.