Research map / Label-efficient and robust learning / Recognition and long-tail
Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding
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 DRLS, a VAE-based method for doubly incomplete multi-view multi-label classification that disentangles view-consistent and view-specific features via a mutual-information upper bound and learns label semantic embeddings with a GIN-based VGAE.
Problem
Doubly incomplete multi-view multi-label classification (DIMVMLC), where both views and labels are missing, and where view-shared and view-specific information are poorly separated.
Approach
Combine cross-view reconstruction with a PoE-fused consistent posterior, pre-train consistent encoders, then minimize a variational upper bound on mutual information between consistent and specific features; label semantic embeddings are learned with a GIN-based variational graph autoencoder over a label co-occurrence graph and used for feature selection.
Key results
- On Corel5k with 50% missing views and labels, DRLS reaches AP 0.433 and AUC 0.918 versus 0.416 AP and 0.912 AUC for the next best (SIP).
- DRLS ranks first on average across all five datasets under the 50% view and 50% label missing setting (Ave.R 1.000).
- Ablation on Corel5k shows removing the disentangling loss drops AP from 0.433 to 0.414 and removing the consistent KL term drops AP to 0.415.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DRLS instead of SIP | DIMVMLC · Corel5k with 50% missing views and labels | AP: 0.4160.015 -> 0.4330.008 | Helped |
| use DRLS instead of MTD | DIMVMLC · Pascal07 with 50% missing views and labels | AP: 0.5500.004 -> 0.5670.008 | Helped |
| use DRLS instead of SIP | DIMVMLC · Espgame with 50% missing views and labels | AP: 0.3100.004 -> 0.3260.005 | Helped |
| use DRLS instead of MTD | DIMVMLC · IAPRTC12 with 50% missing views and labels | AP: 0.3320.002 -> 0.3560.006 | Helped |
| use DRLS instead of SIP | DIMVMLC · Mirflickr with 50% missing views and labels | AP: 0.6150.004 -> 0.6300.005 | Helped |
| add disentangled specific feature extraction to consistent feature extraction (ablation) | DRLS · Corel5k | AP: 0.416 -> 0.431 | Helped |
| add label semantic embedding learning module to A+B (ablation) | DRLS · Corel5k | AP: 0.431 -> 0.433 | Helped |
| remove Lcon.kl (ablation) | DRLS · Corel5k | AP: 0.433 -> 0.415 | Hurt |
| remove Ldisent (ablation) | DRLS · Pascal07 | AP: 0.567 -> 0.557 | Hurt |
| remove Lle (ablation) | DRLS · Pascal07 | AP: 0.567 -> 0.565 | No clear effect |
Limitations
- Missing labels are treated as negatives for baselines that cannot handle them, and missingness is simulated with random 50% removal only.
- Disentanglement relies on a two-stage pre-training and then fine-tuning schedule, adding optimization complexity to tune.
Experimental setup
Datasets: Corel5k Pascal07 Espgame IAPRTC12 Mirflickr
Benchmarks: Corel5k Pascal07 Espgame IAPRTC12 Mirflickr
Metrics: AP AUC 1-RL 1-HL 1-OE 1-Cov average rank
Baselines: DM2L CDMM iMvWL NAIML DDINet DICNet MTD SIP
Builds on: variational autoencoder multi-view VAE with product-of-experts variational graph autoencoder MTD SIP
Closest papers by meaning
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.92)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.92)
- View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning (CVPR 2024 · similarity 0.91)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.91)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.90)
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (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.