Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Compact Semantic Information for Incomplete Multi-View Missing 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 COME, a MoE-based framework for incomplete multi-view missing multi-label classification that combines incomplete multi-view contrastive learning, cross-view reconstruction-based mutual information maximization, and dual-branch soft pseudo-label cross-imputation.
Problem
Incomplete multi-view missing multi-label classification (iM3C), where some views and some labels are missing, making it hard to learn discriminative shared representations from limited supervision.
Approach
Argues contrastive learning is insufficient on incomplete multi-view data, so it adds reconstruction losses (intra- and cross-view) as variational lower bounds on mutual information between raw data and joint representation to compress task-irrelevant information, and uses two branches that exchange dynamically thresholded soft pseudo-labels for missing labels.
Key results
- On Corel5k with 50% missing views and 50% missing labels, COME reaches AP 0.432 vs 0.418 for the best baseline SIP.
- On Pascal07 under the same setting, COME achieves AP 0.590 vs 0.555 for SIP.
- COME achieves the best value on all six metrics across the five datasets in the 50%/50% missing setting, and remains competitive on complete data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use COME instead of SIP | iM3C classification · Corel5k | AP: 0.418 -> 0.432 | Helped |
| use COME instead of SIP | iM3C classification · Pascal07 | AP: 0.555 -> 0.590 | Helped |
| use COME instead of UPDGD-Net | iM3C classification · ESPGame | AP: 0.312 -> 0.319 | Helped |
| use COME instead of UPDGD-Net | iM3C classification · IAPRTC12 | AP: 0.339 -> 0.351 | Helped |
| use COME instead of SIP | iM3C classification · Mirflickr | AP: 0.614 -> 0.633 | Helped |
| add LRE1, LRE2, and LCE (ablation) | COME on Corel5k · Corel5k | AP: 0.390 -> 0.432 | Helped |
| add LRE1, LRE2, and LCE (ablation) | COME on Pascal07 · Pascal07 | AP: 0.535 -> 0.590 | Helped |
| add LCE to the model with LRE1 and LRE2 (ablation) | COME on Corel5k · Corel5k | AP: 0.390 -> 0.432 | Helped |
| add LCE to the model with LRE1 and LRE2 (ablation) | COME on Pascal07 · Pascal07 | AP: 0.538 -> 0.590 | Helped |
Surprising findings
- Contrastive learning on incomplete multi-view data yields worse classification than on complete views, motivating the shift to mutual information maximization.
- View missing has a more severe performance impact than label missing at the same rate.
Limitations
- Hard pseudo-label methods are argued to amplify errors, but the proposed method's sensitivity to threshold hyperparameters (tau_h, tau_l, decay) is not deeply analyzed in the main text.
- Evaluation limited to five image-based multi-view multi-label benchmarks with handcrafted features.
Experimental setup
Datasets: Corel5k Pascal07 ESPGame IAPRTC12 Mirflickr
Benchmarks: Corel5k Pascal07 ESPGame IAPRTC12 Mirflickr
Metrics: AP 1-HL 1-RL AUC 1-OE 1-COV
Baselines: CDMM DM2L LVSL iMVWL NAIM3L DICNet UPDGD-Net SIP
Builds on: incomplete multi-view contrastive learning (DICNet) FixMatch information bottleneck / minimal sufficient representation
Closest papers by meaning
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.93)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.92)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.91)
- View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning (CVPR 2024 · similarity 0.91)
- Incomplete Multi-View Multi-Label Classification via Diffusion-Guided Redundancy Removal (AAAI 2025 · similarity 0.90)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.90)
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.