Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-label Learning with Missing Labels Using Mixed Dependency Graphs
Baoyuan Wu, Fan Jia, Wei Liu, Bernard Ghanem, Siwei Lyu
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace class co-occurrence graph with sparse and low rank decomposition (95% missing, original init) | MLMG · MediaMill, ESP Game, IAPRTC-12 | AP relative improvement: 2.43%, 6.29%, 3.94%, 2.78%, 4.7% | Improved |
| replace class co-occurrence graph with sparse and low rank decomposition (95% missing, filled init) | MLMG · MediaMill, ESP Game, IAPRTC-12 | AP relative improvement: 6.38%, 0.23%, -0.87%, 2.76%, 1.43% | Mixed |
| use MLMG-CO instead of MLMG-SL when missing label proportion is small | MLMG · MediaMill, ESP Game, IAPRTC-12 | AP and mAP | Mixed |
| fill initial label matrix with ancestor labels from the semantic hierarchy (ablation) | MLMG · four multi-label datasets | AP and mAP | Improved |
| add semantic hierarchical constraint to MLMG-CO (ablation) | MLMG-CO · four multi-label datasets | AHL: positive (filling only) -> 0 | Improved |
| add both filling and constraint to MLMG-CO (ablation) | MLMG-CO · four multi-label datasets | AP and mAP | Improved |
| use semantic hierarchy as constraint during optimization instead of CSSAG post-processing | MLMG-CO · four multi-label datasets | mAP | Improved |
| use MLMG-CO/SL without hierarchy information compared to best competitor | MLMG · six datasets | AP and mAP margin: at least 5% (AP) or 3% (mAP) | Improved |
| compute PGD step size once every 5 iterations with damping 0.9 (ablation) | MLMG-CO · runtime | runtime: 50% -> about 40% | Improved |
| initialize missing labels randomly instead of zero (ablation) | MLMG-CO and MLMG-SL · 50% missing labels | AP | No clear effect |
| generate missing labels on all classes instead of only leaf and singleton classes (ablation) | MLMG-CO and MLMG-SL · training imputation | AP and mAP | Hurt |
| switch from traditional to CNN features (ESP Game testing AP, MLMG-CO) | MLMG-CO · ESP Game semi-supervised | AP: 0.4422 -> 0.5887 | Improved |
Closest papers by meaning
- Incomplete Label Multiple Instance Multiple Label Learning (TPAMI 2020 · similarity 0.88)
- Knowledge Amalgamation for Multi-Label Classification via Label Dependency Transfer (AAAI 2023 · similarity 0.88)
- Improving Multi-label Learning with Missing Labels by Structured Semantic Correlations (ECCV 2016 · similarity 0.88)
- Limited-Supervised Multi-Label Learning with Dependency Noise (AAAI 2024 · similarity 0.87)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.86)
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.
