Research map / Label-efficient and robust learning / Noisy labels
Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels
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 DMMIvNL, a deep multi-view multi-label method that handles both missing views and multi-label noise via an information-bottleneck feature extractor and a volume-minimization transition-matrix network with cycle-consistency.
Problem
Multi-view multi-label classification where views are missing for some instances and labels are noisy (both over-labeled and missing labels), which prior work handled only separately.
Approach
Combines an information-bottleneck-derived mutual information objective (shared vs. specific features, with variational upper bound) with a class-wise transition-matrix volume minimization and a cycle-consistency loss between forward and backward noise estimation, plus label-correlation and noisy-label regularization.
Key results
- On Espgame at PER=50%, rho=30%, DMMIvNL reports an AP improvement of up to 43% over compared methods (per the paper's text).
- Ablation on Espgame: full model AP 0.1934 / AUC 0.7307 versus 0.1750 / 0.6846 with only the base setup (no LIB, LID, or Lr terms in the first row).
- Consistently outperforms eight baselines across noise rates 10% to 40% in AUC on six datasets; ranked best (Ave.R 1.00) on Corel 5k and MIR FLICKR among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DMMIvNL instead of the strongest listed baseline | multi-view multi-label classification · Yeast, Corel 5k, VOC 2007, Espgame, IAPR TC-12, MIR FLICKR | AP: 65.61±1.34 -> 72.08±0.34; 18.49±0.21 -> 19.18±0.45; 41.41±0.68 -> 43.86±1.15; 13.38±0.29 -> 19.13±0.58; 17.37±0.89 -> 19 | Helped |
| use DMMIvNL instead of the strongest listed baseline | multi-view multi-label classification · Yeast, Corel 5k, VOC 2007, Espgame, IAPR TC-12, MIR FLICKR | AUC: 78.36±0.54 -> 81.03±0.55; 70.37±0.37 -> 72.42±0.47; 69.76±0.68 -> 74.65±1.01; 66.64±0.79 -> 69.86±0.54; 70.87±0.60 -> 72 | Helped |
| add information-bottleneck representation extraction (LIB) (ablation) | DMMIvNL · Espgame, Corel 5k, VOC 2007 | AP: 0.1499 -> 0.1934; 0.1543 -> 0.1997; 0.3612 -> 0.4622 | Helped |
| add cycle-consistency estimation (LID) (ablation) | DMMIvNL · Espgame, Corel 5k, VOC 2007 | AP: 0.1838 -> 0.1934; 0.1850 -> 0.1997; 0.4305 -> 0.4622 | Helped |
| add label-correlation regularization (Lr) (ablation) | DMMIvNL · Espgame, Corel 5k, VOC 2007 | AP: 0.1834 -> 0.1934; 0.1800 -> 0.1997; 0.4359 -> 0.4622 | Helped |
| replace cycle-consistency loss LID with empirical volume loss Lvol (ablation) | DMMIvNL · Espgame, Corel 5k, VOC 2007 | AP: 0.1934 -> 0.1750; 0.1997 -> 0.1713; 0.4622 -> 0.4145 | Hurt |
Surprising findings
- Fixed-noise-rate baselines such as NTMLCS and DIMC degrade sharply at high noise levels, while the proposed method stays stable.
Limitations
- Assumes class-dependent, instance-independent noise transition matrices and sufficiently scattered class posteriors.
- Noise is simulated by uniform flipping of labels rather than real annotation noise.
Experimental setup
Datasets: Yeast Corel 5k VOC 2007 Espgame IAPR TC-12 MIR FLICKR
Benchmarks: Yeast Corel 5k VOC 2007 Espgame IAPR TC-12 MIR FLICKR
Metrics: 1-HL 1-RL 1-OE 1-Cov AP AUC
Baselines: iMVWL TM3L AIMNet DICNet DIMC MTD SIP ENTMLC
Builds on: information bottleneck volume minimization for transition matrix estimation ENTMLC
Closest papers by meaning
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.91)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.90)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.90)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.89)
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.89)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.89)
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.