Research map / Label-efficient and robust learning / Recognition and long-tail
A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification
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 TSIEN, a two-stage information-theoretic network (information bottleneck per view, then autoencoder-based cross-view mutual information extraction with weighted fusion) for incomplete multi-view multi-label classification with missing views and labels.
Problem
Multi-view multi-label classification when views and labels are partially missing, which breaks conventional MvMLC methods that assume complete data.
Approach
Combines the information bottleneck principle (per-view compression of task-irrelevant information via variational upper bound) with an autoencoder-based cross-view mutual information maximization stage and a missing-aware weighted fusion classifier, trained in two stages with masks for missing views and labels.
Key results
- On Corel5k, the method reaches AP 0.436 versus 0.382 for the best baseline (LMVCAT), with the paper reporting AP gains of roughly 2% to 5% across datasets.
- On ESPGame, AP reaches 0.319 versus 0.297 for the best baseline (DICNet).
- Ablation on Corel5k shows the one-stage variant (AP 0.422) underperforms the two-stage model (AP 0.436).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TSIEN instead of the strongest listed baseline | iMvMLC methods · Corel5k | AP: 0.3820.004 -> 0.4360.007 | Helped |
| use TSIEN instead of the strongest listed baseline | iMvMLC methods · Pascal07 | AP: 0.5480.008 -> 0.5810.009 | Helped |
| use TSIEN instead of the strongest listed baseline | iMvMLC methods · ESPGame | AP: 0.2970.002 -> 0.3190.004 | Helped |
| use TSIEN instead of the strongest listed baseline | iMvMLC methods · IAPRTC12 | AP: 0.3050.004 -> 0.3610.004 | Helped |
| use TSIEN instead of the strongest listed baseline | iMvMLC methods · Mirflickr | AP: 0.6120.005 -> 0.6310.003 | Helped |
| add L(k)IB to the LCE backbone (ablation) | two-stage network · Corel5k | AP: 0.375 -> 0.383 | Helped |
| add LCL1 to the LCE backbone (ablation) | two-stage network · Corel5k | AP: 0.375 -> 0.389 | Helped |
| add LCL2 to the LCE backbone (ablation) | two-stage network · Corel5k | AP: 0.375 -> 0.376 | No clear effect |
| add L(k)IB, LCL1, and LCL2 (ablation) | two-stage network · Corel5k | AP: 0.375 -> 0.436 | Helped |
| combine the two training stages into one stage (ablation) | TSIEN · Corel5k | AP: 0.436 -> 0.422 | Hurt |
Surprising findings
- Merging the two stages into one stage with a shared hyper-parameter degrades performance, which the authors attribute to imbalanced per-view information extraction.
Limitations
- Tuning the per-view hyper-parameters beta(k) becomes harder as the number of views grows, so the paper sets them equal.
- Evaluated only on five image datasets with six fixed handcrafted views and simulated missingness.
Experimental setup
Datasets: Corel5k Pascal07 ESPGame IAPRTC12 Mirflickr
Benchmarks: Corel5k Pascal07 ESPGame IAPRTC12 Mirflickr
Metrics: AP 1-HL 1-RL AUC OE Cov
Baselines: iMvWL NAIML CDMM DeepIMV LMVCAT DICNet
Builds on: information bottleneck (Alemi et al. 2017) InfoNCE supervised contrastive learning Deep Incomplete Multi-View Learning (DeepIMV)
Closest papers by meaning
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.91)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.91)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.91)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.90)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.90)
- Incomplete Multi-View Multi-Label Classification via Diffusion-Guided Redundancy Removal (AAAI 2025 · 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.