Research map / Learning theory and algorithms / Clustering
PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced 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 PROTOCOL, a partial/unbalanced optimal-transport self-labeling and class-rebalanced contrastive framework for clustering imbalanced multi-view data.
Problem
Contrastive multi-view clustering implicitly assumes balanced class distributions, so it degrades on real-world multi-view data with class imbalance and under-represents minority samples.
Approach
Reformulates imbalanced multi-view self-labeling as a partial optimal transport problem combined with unbalanced OT (progressive mass constraint and weighted KL on class marginals), then uses the resulting POT labels for feature-level logit adjustment and class-level class-sensitive contrastive learning.
Key results
- At imbalance ratio R=0.1, PROTOCOL reaches 0.892 ACC on Hdigit versus 0.825 for the second-best method (SEM), a 6.7% gain.
- At R=0.1 on Caltech, PROTOCOL achieves 0.791 ACC and 0.679 NMI, improving over GCFAggMVC (0.743 ACC, 0.642 NMI).
- Ablation on Hdigit at R=0.1 shows the POT label assignment raises ACC from 0.722 to 0.876 over the base model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add POT label assignment to the Base method (ablation) | PROTOCOL · Caltech R=0.1 and 0.5; Hdigit R=0.1 and 0.5; Fashion R=0.1 and 0.5 | ACC: Caltech R=0.1: 0.719 -> 0.772; R=0.5: 0.737 -> 0.757; Hdigit R=0.1: 0.722 -> 0.876; R=0.5: 0.979 -> 0.980; Fashion R=0.1 | Helped |
| Add POT-guided class-rebalanced contrastive learning to the POT variant (ablation) | PROTOCOL · Caltech R=0.1 and 0.5; Hdigit R=0.1 and 0.5; Fashion R=0.1 and 0.5 | ACC: Caltech R=0.1: 0.772 -> 0.791; R=0.5: 0.757 -> 0.767; Hdigit R=0.1: 0.876 -> 0.892; R=0.5: 0.980 -> 0.983; Fashion R=0.1 | Helped |
| Use PROTOCOL instead of the strongest baseline | Imbalanced multi-view clustering · Hdigit, Fashion, Caltech, NUS-WIDE, and Cifar10, R=0.1; except NUS-WIDE R=0.5 | ACC: Hdigit: 0.825 -> 0.892; Fashion: 0.804 -> 0.846; Caltech: 0.743 -> 0.791; NUS-WIDE: 0.440 -> 0.470; Cifar10: 0.851 -> 0. | Helped |
| Use PROTOCOL instead of the strongest baseline | Imbalanced multi-view clustering · Hdigit, Fashion, Caltech, and Cifar10, R=0.5; NUS-WIDE R=0.5 | ACC: Hdigit: 0.980 -> 0.983; Fashion: 0.983 -> 0.987; Caltech: 0.761 -> 0.767; NUS-WIDE: 0.545 -> 0.542; Cifar10: 0.990 -> 0. | Mixed |
Surprising findings
- Several baselines (e.g., NUS-WIDE and Caltech) show lower performance under more severe imbalance, with the degradation most pronounced at R=0.1.
Limitations
- Imbalance is simulated by subsampling benchmark datasets rather than using naturally imbalanced multi-view data.
- Hyperparameters (temperatures, weights w_v and w_t) are chosen empirically with limited sensitivity analysis.
- Hardware and compute cost are not reported.
Experimental setup
Datasets: Hdigit Fashion NUS-WIDE Caltech Cifar10
Benchmarks: Hdigit Fashion Caltech NUS-WIDE Cifar10
Metrics: ACC NMI PUR
Baselines: CoMVC MFLVC AECoDDC GCFAggMVC SEM CSOT DSIMVC AR-DMVC AR-DMVC-AM
Builds on: Sinkhorn entropic optimal transport partial optimal transport unbalanced optimal transport SeLa P2OT GCFAggMVC
Closest papers by meaning
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.88)
- Contrastive Multi-view Subspace Clustering via Tensor Transformers Autoencoder (AAAI 2025 · similarity 0.87)
- Cluster-guided Contrastive Class-imbalanced Graph Classification (AAAI 2025 · similarity 0.87)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.87)
- Multi-view Granular-ball Contrastive Clustering (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.