Research map / Label-efficient and robust learning / Recognition and long-tail
Implicit Relative Labeling-Importance Aware Multi-Label Metric Learning
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 ILIA, a multi-label metric learning method that recovers implicit relative labeling-importance (RLI) via local linear reconstruction and learns a discriminative Mahalanobis metric aligning predictions with the recovered RLIs.
Problem
Existing multi-label metric learning assumes equal labeling-importance for all relevant labels, ignoring that the relative importance of labels for an instance differs and is not explicitly annotated.
Approach
Recover implicit relative labeling-importance matrix by transferring the local linear reconstruction structure of the feature space to the label space, then learn a kernel regression model and a Mahalanobis metric (with kNN-based pull/push terms and LogDet regularization) that align predictions with these recovered RLIs.
Key results
- BRKNN-ILIA achieves 46 wins / 13 ties / 1 loss against BRKNN-DeV1 across six metrics under pairwise t-test at 5% significance.
- ILIA is significantly superior or comparable to LM, LJE, COMMU, LIMIC and LSMM in 81.7%, 76.7%, 96.7%, 56.7% and 41.7% of cases (plus ties) when coupled with BRKNN, per the paper's win/tie summary.
- Ablation shows removing implicit RLI recovery (DeV1) or the discriminative metric (DeV2) degrades performance, with MLKNN-ILIA beating MLKNN-DeV2 in 45 win counts overall.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add implicit relative labeling-importance recovery to BRKNN metric learning | BRKNN-ILIA · 10 benchmark multi-label datasets | Average precision | Helped |
| Add implicit relative labeling-importance recovery (ablation) | BRKNN-ILIA · 10 benchmark multi-label datasets | Average precision: 8/2/0 | Helped |
| Add implicit relative labeling-importance recovery (ablation) | MLKNN-ILIA · 10 benchmark multi-label datasets | Average precision: 9/1/0 | Helped |
| Add discriminative metric M for similarity characterization (ablation) | BRKNN-ILIA · 10 benchmark multi-label datasets | Average precision: 7/3/0 | Helped |
| Add discriminative metric M for similarity characterization (ablation) | MLKNN-ILIA · 10 benchmark multi-label datasets | Average precision: 10/0/0 | Helped |
| Vary the number of nearest neighbors k (ablation) | A-ILIA · emotions and image | Hamming loss and Average precision | Mixed |
Limitations
- RLI recovery relies on kNN graph and k, with performance sensitive to neighborhood size k
- Training complexity scales cubically in n due to matrix operations and alternating optimization
- Evaluated only with KNN-based downstream learners (BRKNN, MLKNN)
Experimental setup
Datasets: CAL500 emotions medical image scene arts corel5k education health entertainment
Benchmarks: CAL500 emotions medical image scene arts corel5k education health entertainment
Metrics: Hamming loss Ranking loss Coverage Average precision Macro-F1 Macro-averaging AUC
Baselines: LM LJE COMMU LIMIC LSMM BRKNN MLKNN
Builds on: LSMM LIMIC Zadeh et al. 2016 geometric mean metric learning locally linear embedding MLKNN
Closest papers by meaning
- T-MDML: Triplet-based Multiple Distance Metric Learning for Multi-Instance Multi-Label Classification with Label Correlation (AAAI 2025 · similarity 0.88)
- Deep Semi-Supervised Metric Learning With Mixed Label Propagation (CVPR 2023 · similarity 0.87)
- Online Metric Learning for Multi-Label Classification (AAAI 2020 · similarity 0.87)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.84)
- In Pursuit of Causal Label Correlations for Multi-label Image Recognition (NeurIPS 2024 · similarity 0.84)
- Multi-Label Supervised Contrastive Learning (AAAI 2024 · similarity 0.84)
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.