Research map / Label-efficient and robust learning / Recognition and long-tail
Attention-Induced Embedding Imputation for Incomplete Multi-View Partial 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 AIMNet, which imputes missing multi-view instance embeddings via cross-view joint attention and fuses view predictions by attention-derived confidence, with a GAT-based label semantic branch, for incomplete multi-view partial multi-label classification.
Problem
Multi-view multi-label classification where both views of instances and labels are incomplete (double-missing), which hurts model performance in practice.
Approach
Impute missing instances' latent embeddings (not raw or kernel features) using a cross-view joint attention matrix built by max-aggregating view-specific instance attention, weight view predictions by attention-derived confidence in late fusion, and learn label-specific semantics with a GAT over a statistical label co-occurrence matrix.
Key results
- On Corel5k with 50% missing views and labels, AIMNet reaches AP 0.400 vs 0.381 for the next-best DICNet, and ranks first on all metrics on Corel5k, Pascal07, ESPGame and MIRFLICKR.
- Ablation on Corel5k: removing the attention-induced imputation drops AP from 0.400 to 0.390 and 1-RL from 0.902 to 0.896.
- On complete Corel5k (no missing data), AIMNet outperforms almost all baselines per the radar chart.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AIMNet instead of DICNet | iMvPMLC · Corel5k | AP: 0.3810.004 -> 0.4000.010 | Helped |
| Use AIMNet instead of DICNet | iMvPMLC · Pascal07 | AP: 0.5050.012 -> 0.5480.008 | Helped |
| Use AIMNet instead of DICNet | iMvPMLC · ESPGame | AP: 0.2970.002 -> 0.3050.004 | Helped |
| Use AIMNet instead of DICNet | iMvPMLC · IAPRTC12 | AP: 0.3230.001 -> 0.3290.005 | Helped |
| Use AIMNet instead of DICNet | iMvPMLC · MIRFLICKR | AP: 0.5890.005 -> 0.6020.004 | Helped |
| Add label-missing mask (ablation) | AIMNet · Corel5k | AP: 0.380 -> 0.400 | Helped |
| Add label-missing mask (ablation) | AIMNet · ESPGame | AP: 0.293 -> 0.305 | Helped |
| Add attention-induced view completion and confidence weighting (ablation) | AIMNet · Corel5k | AP: 0.390 -> 0.400 | Helped |
| Add attention-induced view completion and confidence weighting (ablation) | AIMNet · ESPGame | AP: 0.294 -> 0.305 | Helped |
| Add attention-induced view completion and confidence weighting (ablation) | AIMNet · Corel5k | 1-RL: 0.896 -> 0.902 | Helped |
| Add attention-induced view completion and confidence weighting (ablation) | AIMNet · ESPGame | 1-RL: 0.840 -> 0.846 | Helped |
| Add attention-induced view completion and confidence weighting (ablation) | AIMNet · Corel5k | 1-Cov: 0.764 -> 0.771 | Helped |
Limitations
- Missing-view imputation gains in ablation are modest relative to the full model's ranking claims
- Baselines not designed for missing views were adapted by mean-filling or treating unknown labels as negatives, which may disadvantage them
- Training time is longer than several baselines (244 s on Corel5k vs 16 s for CDMM)
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 average rank
Baselines: C2AE GLOCAL CDMM DM2L LVSL iMVWL NAIM3L DICNet
Builds on: graph attention network LLSF DICNet iMVWL
Closest papers by meaning
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · 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 0.90)
- Incomplete Multi-View Multi-Label Classification via Diffusion-Guided Redundancy Removal (AAAI 2025 · similarity 0.89)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 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.