Research map / Label-efficient and robust learning / Recognition and long-tail
Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label 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 DHLI, a double-layer hybrid-label identification method that splits observed multi-view multi-label targets into common, view-specific and noisy labels and uses them to guide feature selection with logic-operation regularization.
Problem
Multi-view multi-label feature selection where label noise and view-specific label information are not handled, so the mapping between labels and features is degraded.
Approach
Decomposes observed labels via OR-based logic into common labels, per-view specific labels and sparse noisy labels, and jointly learns feature weights for common and specific mappings with logic-based constraints, solved by multiplicative update rules.
Key results
- DHLI reaches AP 0.7951 on SCENE and 0.6635 on MIRFlickr, the best among compared feature selection methods in most cases.
- Ablation on SCENE, Corel5K and IAPRTC12 shows removing any component (alpha, beta, gamma, delta) lowers average precision, with removing the specific-label constraint causing the largest drop.
- Convergence curves on OBJECT and MIRFlickr show the optimization converges quickly.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the common-label majority regularization (alpha component) (ablation) | DHLI · SCENE, Corel5K, and IAPRTC12 | Average Precision: 0.7852 -> 0.7951; 0.2366 -> 0.2370; 0.1403 -> 0.1421 | Helped |
| add the label-noise sparsity constraint (beta component) (ablation) | DHLI · SCENE, Corel5K, and IAPRTC12 | Average Precision: 0.7851 -> 0.7951; 0.2369 -> 0.2370; 0.1401 -> 0.1421 | Helped |
| add the feature-row sparsity constraint (gamma component) (ablation) | DHLI · SCENE, Corel5K, and IAPRTC12 | Average Precision: 0.7853 -> 0.7951; 0.2364 -> 0.2370; 0.1396 -> 0.1421 | Helped |
| add the view-specific label constraints (delta component) (ablation) | DHLI · SCENE, Corel5K, and IAPRTC12 | Average Precision: 0.7842 -> 0.7951; 0.2316 -> 0.2370; 0.1390 -> 0.1421 | Helped |
Surprising findings
- ELSMML ties DHLI on several datasets' AP results (identical values on SCENE, OBJECT, MIRFlickr, Corel5K, IAPRTC12 and 3Sources), with DHLI and ELSMML sharing the same ranking in many cases.
Limitations
- Relies on the assumption that the common part accounts for the majority of observed labels.
- Hyperparameters (alpha, beta, gamma, delta) require grid search and results vary with parameter choice.
Experimental setup
Datasets: SCENE OBJECT MIRFlickr Corel5K IAPRTC12 3Sources
Benchmarks: SCENE OBJECT MIRFlickr Corel5K IAPRTC12 3Sources
Metrics: Average Precision Coverage Hamming Loss Ranking Loss
Baselines: MSFS ELSMML M2LD MoRE MRDM CLML SCMFS
Builds on: multi-label feature selection view-specific labels l2,1-norm regularization
Closest papers by meaning
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.90)
- Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · similarity 0.90)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.90)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.90)
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.89)
- Noisy Label Calibration for Multi-View Classification (AAAI 2025 · similarity 0.88)
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.