Research map / Learning theory and algorithms / Clustering
Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder
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 TTAE, a multi-view subspace clustering method combining a tensor transformer autoencoder for cross-view interaction, a self-expression layer, and an adaptive contrastive loss that excludes kNN affinity pairs from negatives.
Problem
Deep multi-view subspace clustering often uses view-specific encoders that lack cross-view complementary interaction, and contrastive multi-view clustering treats same-cluster samples as negatives, creating false negatives that conflict with clustering.
Approach
A tensor attention layer enables inter-view information exchange; the self-expression coefficient matrix defines kNN affinity pairs that are excluded from negatives in an adaptive contrastive loss, and in turn that loss feeds discriminative features back to the self-expression layer.
Key results
- TTAE reaches 65.75% ACC and 69.36% NMI on Fashion-MNIST, outperforming the best baseline (D2MVSC/DMSC-UDL/CMSC-DCCA around 65.45 ACC).
- On Noisy-MNIST TTAE reports 94.53% ACC and 90.08% NMI, above MFLVC (92.82/88.79).
- Ablation on Fashion-MNIST shows the full loss combination (Lre+Lse+Lacl) gives the best ACC of 65.75 versus 53.50 with reconstruction only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reconstruction loss Lre (ablation) | TTAE · Fashion-MNIST | ACC (%): 53.50 -> 65.75 | Helped |
| add reconstruction loss Lre (ablation) | TTAE · YTF | ACC (%): 54.45 -> 68.10 | Helped |
| add self-expression loss Lse (ablation) | TTAE · Fashion-MNIST | ACC (%): 52.81 -> 65.75 | Helped |
| add self-expression loss Lse (ablation) | TTAE · YTF | ACC (%): 60.18 -> 68.10 | Helped |
| add adaptive contrastive loss Lacl (ablation) | TTAE · Fashion-MNIST | ACC (%): 51.96 -> 65.75 | Helped |
| add adaptive contrastive loss Lacl (ablation) | TTAE · YTF | ACC (%): 61.10 -> 68.10 | Helped |
| add reconstruction loss Lre (ablation) | TTAE · Fashion-MNIST | NMI (%): 52.76 -> 69.36 | Helped |
| add reconstruction loss Lre (ablation) | TTAE · YTF | NMI (%): 71.56 -> 83.54 | Helped |
| add self-expression loss Lse (ablation) | TTAE · Fashion-MNIST | NMI (%): 58.06 -> 69.36 | Helped |
| add self-expression loss Lse (ablation) | TTAE · YTF | NMI (%): 78.32 -> 83.54 | Helped |
| add adaptive contrastive loss Lacl (ablation) | TTAE · Fashion-MNIST | NMI (%): 51.12 -> 69.36 | Helped |
| add adaptive contrastive loss Lacl (ablation) | TTAE · YTF | NMI (%): 77.23 -> 83.54 | Helped |
Surprising findings
- Table 3 reports DCCA at 85.53% ACC on Noisy-MNIST, higher than many contrastive and subspace baselines, though the authors note DCCA-based methods were suboptimal on large datasets in the text.
Limitations
- Attention adds computational overhead, stated as O(N^2 ov) complexity.
- Large datasets were randomly subsampled to make baseline comparisons feasible.
- Hyperparameters (lambda1, lambda2, xi) require tuning, with analysis only on Fashion-MNIST.
Experimental setup
Datasets: Fashion-MNIST COIL-20 FRGC Youtube-Faces (YTF) ALOI-100 Noisy-MNIST
Benchmarks: Fashion-MNIST COIL-20 FRGC Youtube-Faces (YTF) ALOI-100 Noisy-MNIST
Metrics: ACC NMI runtime
Baselines: K-means CC LALMVC SMVSC EOMSC-CA DCCA DMSC CMSC-DCCA DMJC DMSC-UDL DMIM MFLVC
Builds on: Transformer contrastive learning DMSC MFLVC
Closest papers by meaning
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Multi-view Granular-ball Contrastive Clustering (AAAI 2025 · similarity 0.88)
- Partial Multi-View Clustering via Self-Supervised Network (AAAI 2024 · similarity 0.88)
- From Dictionary to Tensor: A Scalable Multi-View Subspace Clustering Framework with Triple Information Enhancement (NeurIPS 2024 · similarity 0.88)
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.88)
- PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering (ICML 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.