Research map / Learning theory and algorithms / Clustering
Ambiguous Instance-Aware Contrastive Network with Multi-Level Matching for Multi-View Document 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 AICN-MLM, a multi-view document clustering network combining multi-level matching (similarity distribution MMD alignment, cross-view prototype JS alignment) with ambiguous-instance-aware weighted contrastive learning.
Problem
Multi-view document clustering suffers from alignment based on a single perspective (features or classes) and from contrastive learning that treats all instance pairs equally, ignoring ambiguous instances.
Approach
Aligns multi-view data jointly at feature/pseudo-label (MMD between cosine similarity distribution and pseudo-label distribution) and prototype levels (JS divergence between cross-view prototypes), and reweights contrastive pairs via a weight modulation function that up-weights ambiguous pairs and down-weights clear ones.
Key results
- On KUST-ETD, AICN-MLM reaches 89.08 ACC, 86.21 NMI, 80.13 ARI, outperforming the best baseline CVCL (79.32 ACC).
- On Reuters, AICN-MLM achieves 59.03 ACC and 63.01 PUR versus DealMVC's 55.29 ACC and 62.60 PUR.
- Ablation removing the ambiguous instance-aware contrastive learning module drops KUST-ETD ACC from 89.08 to 28.36.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove similarity distribution matching (SDM) (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 83.76; 59.03 -> 58.94 | Hurt |
| remove cross-view prototype matching (CVPM) (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 84.22; 59.03 -> 57.63 | Hurt |
| remove ambiguous instance-aware contrastive learning (AICL) (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 28.36; 59.03 -> 35.29 | Hurt |
| remove ambiguous-instance weighting from AICL (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 83.15; 59.03 -> 43.15 | Hurt |
| replace MMD with KL divergence in SDM (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 83.09; 59.03 -> 53.12 | Hurt |
| replace JS divergence with KL divergence in CVPM (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 78.93; 59.03 -> 53.18 | Hurt |
| replace AICL with PSCL contrastive loss (ablation) | AICN-MLM · KUST-ETD; Reuters | ACC: 89.08 -> 58.39; 59.03 -> 47.61 | Hurt |
Surprising findings
- Removing the entire AICL module collapses performance to near-chance levels (e.g., 28.36 ACC on KUST-ETD), far below removing other components.
Limitations
- Evaluated only on self-constructed KUST datasets plus one public Reuters dataset, limiting generality.
- Hyperparameters (alpha, beta) are set differently for different dataset groups, suggesting dataset-specific tuning.
Experimental setup
Datasets: KUST-ETD KUST-CTD KUST-CVD KUST-CED KUST-CBD KUST-BTD KUST-CLD Reuters
Benchmarks: KUST-ETD KUST-CTD KUST-CVD KUST-CED KUST-CBD KUST-BTD KUST-CLD Reuters
Metrics: ACC NMI ARI PUR
Baselines: DEMVC SDMVC DSMVC MFLVC FastMICE GCFAgg DealMVC CVCL DIVIDE ICMVC MAGA
Builds on: InfoNCE MFLVC DealMVC
Closest papers by meaning
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.87)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.87)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.85)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.85)
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.85)
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.