Research map / Label-efficient and robust learning / Recognition and long-tail
Incomplete Multi-View Multi-Label Classification via Diffusion-Guided Redundancy Removal
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 DiffSumm, a label-guided diffusion model that recovers summary vectors for missing views in incomplete multi-view multi-label classification and iteratively removes redundant views using fidelity scores.
Problem
Incomplete multi-view multi-label classification, where samples have missing views and missing labels, and some views are redundant and hurt performance.
Approach
Trains a diffusion model conditioned on pseudo-labels to generate synthetic summary vectors for missing views, and uses these synthetic vectors together with fidelity scores (classifier precision on pseudo-labels) to adaptively identify and remove redundant views per sample, then fuses original and synthetic summaries for final prediction.
Key results
- On Corel5k, DiffSumm reaches 45.7 AP versus 38.4 for the best baseline (LMVCAT), under 50% missing views and labels.
- On Espgame, DiffSumm achieves 36.3 AP versus 29.7 for DICNet; it improves AP on all five datasets.
- Ablation: removing the redundant view identification strategy and fusion (V+L, DiffOnly) drops Corel5k AP from 45.7 to 38.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffSumm instead of the strongest comparison method | Incomplete multi-view multi-label classification · Corel5k | AP: 38.1±0.4 -> 45.7±0.1 | Helped |
| Use DiffSumm instead of the strongest comparison method | Incomplete multi-view multi-label classification · Pascal07 | AP: 53.6±0.2 -> 56.1±0.1 | Helped |
| Use DiffSumm instead of DiffOnly | Incomplete multi-view multi-label classification · Espgame | AP: 30.6±0.2 -> 36.3±0.2 | Helped |
| Use DiffSumm instead of DiffOnly | Incomplete multi-view multi-label classification · IAPRTC12 | AP: 35.2±0.2 -> 37.9±0.1 | Helped |
| Use DiffSumm instead of DiffOnly | Incomplete multi-view multi-label classification · Mirflickr | AP: 60.5±0.4 -> 64.6±0.2 | Helped |
| Replace statistical-regularity redundancy identification with DiffSumm identification (ablation) | DiffSumm redundant-view identification · Corel5k | AP: 40.7 -> 45.7 | Helped |
| Replace cosine similarity with fidelity scores for redundant-view identification (ablation) | DiffSumm redundant-view identification · Corel5k | AP: 45.0 -> 45.7 | Helped |
| Remove synthetic-summary and pseudo-label fidelity as a sample-difficulty reference (ablation) | DiffSumm redundant-view identification · Corel5k | AP: 45.7 -> 37.5 | Hurt |
| Add label-guided diffusion to view encoding (ablation) | DiffSumm · Corel5k | AP: 36.7 -> 38.2 | Helped |
| Add redundant-view removal and final prediction modules to view encoding and diffusion (ablation) | DiffSumm · Corel5k | AP: 38.2 -> 45.7 | Helped |
| Use cumulative multiplication rather than cumulative addition to set ksig (ablation) | DiffSumm parameter setting · Corel5k and Pascal07 | AP | Hurt |
| Increase missing-view ratio within the tested range (ablation) | DiffSumm · Corel5k | AP | No clear effect |
Surprising findings
- Increasing the missing view ratio within a certain range does not degrade performance, which the authors attribute to view redundancy.
- Masking some views during training may augment performance.
Limitations
- Slow generation from the label-guided diffusion module requires limiting loop rounds.
- Redundant view identification relies on hand-tuned variance multiplier schedules (ksig) and fidelity thresholds.
- Evaluated only on image-based multi-view multi-label datasets with simulated missingness at fixed 50% rates.
Experimental setup
Datasets: Corel5k Pascal07 Espgame IAPRTC12 Mirflickr
Benchmarks: Corel5k Pascal07 Espgame IAPRTC12 Mirflickr
Metrics: AP 1-RL 1-HL AUC
Baselines: lrMMC MVL-IV iMSF iMvWL NAIML DDINet DICNet LMVCAT DiffOnly
Builds on: DDPM DDIM classifier-free guidance DSMVC RMVC
Closest papers by meaning
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.90)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.89)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.89)
- View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning (CVPR 2024 · similarity 0.88)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.87)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.87)
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.