Research map / Graph learning / Graph self-supervised learning
A Hubness Perspective on Representation Learning for Graph-Based 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.
Shows that hubness in high-dimensional AE embeddings degrades graph-based multi-view clustering and proposes hubREP, a lightweight hub-aware encoder plus hub-aware cross-view pairing loss that improves GMVC across eight benchmarks.
Problem
Graph-based multi-view clustering builds kNN graphs from high-dimensional autoencoder embeddings, which suffer from hubness (a few points dominate neighbor lists), harming graph structure and GNN message passing; existing methods also fail to preserve local neighborhood structure.
Approach
A parametric view-specific encoder trained with a t-SNE-style KL objective over cosine-softmax similarities on a hypersphere, where a negative-entropy term spreads embeddings uniformly to reduce hubness while preserving neighborhoods; plus a cross-view pairing loss weighted by per-sample hubness scores. It replaces autoencoder decoders and is pluggable into existing GMVC methods.
Key results
- On 3Sources, hubREP reaches 0.9290 ACC, 21.30% higher than the closest competitor, and wins in all but one case on the eight benchmarks.
- Plugging hubREP into DMCAG on BBCSport raises ACC from 0.6654 to 0.8143 while cutting parameters from 11.57M to 7.94M.
- Skewness (hubness) correlates negatively with clustering performance (r=-0.36 with ACC, r=-0.57 with NMI, p<0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hub-aware representation embedding and cross-view pairing losses rather than using conventional autoencoders (ablation) | hubREP · 3Sources | ACC: 0.7065 -> 0.9290 | Helped |
| add hub-aware representation embedding and cross-view pairing losses rather than using conventional autoencoders (ablation) | hubREP · BBCSport | ACC: 0.9246 -> 0.9632 | Helped |
| add hub-aware representation embedding and cross-view pairing losses rather than using conventional autoencoders (ablation) | hubREP · BDGP | ACC: 0.7211 -> 0.9868 | Helped |
| add hub-aware representation embedding and cross-view pairing losses rather than using conventional autoencoders (ablation) | hubREP · Reuters | ACC: 0.5102 -> 0.6533 | Helped |
| add the hub-aware representation embedding loss Ls (ablation) | hubREP · 3Sources | ACC: 0.3629 -> 0.9290 | Helped |
| add the hub-aware representation embedding loss Ls (ablation) | hubREP · BBCSport | ACC: 0.4136 -> 0.9632 | Helped |
| add the hub-aware representation embedding loss Ls (ablation) | hubREP · BDGP | ACC: 0.4778 -> 0.9868 | Helped |
| add the hub-aware representation embedding loss Ls (ablation) | hubREP · Reuters | ACC: 0.2417 -> 0.6533 | Helped |
| add the hub-aware cross-view pairing loss Lc (ablation) | hubREP · 3Sources | ACC: 0.8817 -> 0.9290 | Helped |
| add the hub-aware cross-view pairing loss Lc (ablation) | hubREP · BBCSport | ACC: 0.9400 -> 0.9632 | Helped |
| add the hub-aware cross-view pairing loss Lc (ablation) | hubREP · BDGP | ACC: 0.9348 -> 0.9868 | Helped |
| add the hub-aware cross-view pairing loss Lc (ablation) | hubREP · Reuters | ACC: 0.5322 -> 0.6533 | Helped |
Surprising findings
- Reducing embedding dimensionality to avoid hubness hurts clustering performance due to information loss.
- Hubness reduction via simple normalization (L2, ZN, CL2N) consistently improves GMVC clustering across datasets.
Limitations
- Hubness reduction trades off against information loss when dimensionality is reduced, requiring careful dimension choice.
- Some datasets (e.g., Animal, MNIST-USPS) show weaker or mixed results relative to baselines; hubREP does not win every case.
Experimental setup
Datasets: 3Sources BBCSport BDGP MNIST-USPS Hdigit Reuters Animal Noisymnist
Benchmarks: 3Sources BBCSport BDGP MNIST-USPS Hdigit Reuters Animal Noisymnist
Metrics: ACC NMI ARI PUR PRE REC F1 SK HO
Baselines: MCGC EAMC MFLVC AECoDDC GCFAgg DFP-GNN DMCAG S2MVTC MVCAN SURER
Builds on: DMCAG t-SNE nohub (transductive few-shot hubness reduction) graph autoencoder
Closest papers by meaning
- Multi-View Graph Clustering via Node-Guided Contrastive Encoding (ICML 2025 · similarity 0.88)
- A Non-parametric Graph Clustering Framework for Multi-View Data (AAAI 2024 · similarity 0.87)
- Highly Efficient Rotation-Invariant Spectral Embedding for Scalable Incomplete Multi-View Clustering (AAAI 2025 · similarity 0.87)
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.87)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.87)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (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.