Research map / Graph learning / Graph self-supervised learning
Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering
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 Medusa, a multi-scale deep multi-view clustering framework combining dual graph diffusion guided by a consensus graph, a hypergraph-based high-order contrastive loss, and sample-level adaptive fusion weights.
Problem
Deep multi-view clustering methods often fuse inter-view information simplistically, capture only low-order intra-view relations, and assign equal weights to all samples within a view during fusion.
Approach
Dual graph diffusion that propagates both graph structure and node features from a consensus graph into each view with a view-specific weighting coefficient, a hypergraph-derived contrastive loss that treats same-hyperedge samples as positives, and a self-attention plus gating module that learns sample-level fusion weights.
Key results
- Medusa achieves best ACC, NMI and Purity on all eight datasets, e.g. Citeseer ACC 65.34% vs 63.41% for GDMVC (second best).
- On MSRC-v1 Medusa reaches 94.26% ACC and 89.31% NMI, versus 93.21% ACC for GDMVC.
- Ablation shows removing dual diffusion costs about 4% on average and removing the fine-grained fusion costs about 3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dual graph diffusion convolution instead of concatenation for consensus-graph information (ablation) | Medusa · Citeseer | ACC (%): 59.74 ± 1.01 -> 65.34 ± 1.24 | Helped |
| use dual graph diffusion convolution instead of single graph diffusion convolution (ablation) | Medusa · Citeseer | ACC (%): 61.79 ± 0.96 -> 65.34 ± 1.24 | Helped |
| add adaptive weight coefficient b to graph diffusion (ablation) | Medusa · Citeseer | ACC (%): 62.54 ± 0.88 -> 65.34 ± 1.24 | Helped |
| add contrastive learning to the HOCL module (ablation) | Medusa · Citeseer | ACC (%): 58.34 ± 0.78 -> 65.34 ± 1.24 | Helped |
| add hypergraph sub-edge weights to contrastive learning (ablation) | Medusa · Citeseer | ACC (%): 61.57 ± 1.45 -> 65.34 ± 1.24 | Helped |
| use fine-grained sample-level fusion instead of averaging (ablation) | Medusa · Citeseer | ACC (%): 58.63 ± 0.87 -> 65.34 ± 1.24 | Helped |
| add gating mechanisms to sample-level fusion (ablation) | Medusa · Citeseer | ACC (%): 61.84 ± 1.25 -> 65.34 ± 1.24 | Helped |
| use Medusa rather than GDMVC | multi-view clustering · Citeseer | ACC (%): 63.41 ± 1.52 -> 65.34 ± 1.24 | Helped |
| use Medusa rather than GDMVC | multi-view clustering · MSRC-v1 | ACC (%): 93.21 ± 0.46 -> 94.26 ± 0.64 | Helped |
| use Medusa rather than GDMVC | multi-view clustering · Yale | ACC (%): 76.45 ± 0.86 -> 78.84 ± 0.61 | Helped |
| use Medusa rather than GDMVC | multi-view clustering · YoutubeFace | ACC (%): 37.96 ± 2.04 -> 39.25 ± 1.65 | Helped |
| use Medusa rather than GDMVC | multi-view clustering · AWA | ACC (%): 33.62 ± 1.65 -> 35.24 ± 1.54 | Helped |
Limitations
- Large datasets (NUSWIDEOBJ, YoutubeFace, AWA) were randomly subsampled due to graph construction cost, limiting evaluation scale.
- Hyperparameters lambda1 and lambda2 and graph neighbor count were tuned per setting, and baseline hyperparameters were tuned by the authors.
- Reported gains are around 1-2% over the second-best method, with overlapping standard deviations on several datasets.
Experimental setup
Datasets: Yale MSRC-v1 Citeseer SUNRGBD Caltech101 NUSWIDEOBJ YoutubeFace AWA
Benchmarks: Yale MSRC-v1 Citeseer YoutubeFace AWA NUSWIDEOBJ SUNRGBD Caltech101
Metrics: ACC NMI Purity
Baselines: DSMVC GCFAgg SCMVC MVD DFP-GNN AONGR GDMVC
Builds on: personalized PageRank graph diffusion graph convolutional network DEC self-trainable clustering layer contrastive learning
Closest papers by meaning
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.92)
- A Non-parametric Graph Clustering Framework for Multi-View Data (AAAI 2024 · similarity 0.91)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.90)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.90)
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.90)
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (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.