Research map / Label-efficient and robust learning / Recognition and long-tail
Uncertainty-Aware Global-View Reconstruction for Multi-View Multi-Label Feature Selection
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 UGRFS, a sparse multi-view multi-label feature selection model that reconstructs a global view from label-derived distributions, sample confidence, and view weights, optimized with multiplicative updates.
Problem
Feature selection for multi-view multi-label learning (MVML) struggles with noisy separation of consistent vs. complementary information across views, and typically ignores sample-level uncertainty.
Approach
Reconstructs a unified global-view from the graph structure between samples, learnable per-view sample confidence, and automatically learned view weights, and maps the reconstructed view to the label matrix via a Gaussian-kernel label space, with l2,1 sparsity for feature ranking.
Key results
- UGRFS achieves AP of 0.6725 on yeast and 0.8010 on SCENE, best among compared methods on all six datasets (per the reported tables).
- Reports a 4.84% AP gap over the second-ranked method on average and up to 14.3% over the last-ranked method.
- Ablation shows removing global-view reconstruction terms (UGRFS_v3) yields the lowest AP, e.g. 0.6470 vs 0.6770 on MIRFlickr.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UGRFS instead of the strongest listed baseline | UGRFS · yeast | AP: 0.6595±0.016 -> 0.6725±0.014 | Helped |
| use UGRFS instead of the strongest listed baseline | UGRFS · SCENE | AP: 0.7929±0.010 -> 0.8010±0.009 | Helped |
| use UGRFS instead of the strongest listed baseline | UGRFS · VOC07 | AP: 0.5836±0.008 -> 0.5871±0.010 | Helped |
| use UGRFS instead of the strongest listed baseline | UGRFS · MIRFlickr | AP: 0.6442±0.007 -> 0.6770±0.007 | Helped |
| use UGRFS instead of the strongest listed baseline | UGRFS · IAPRTC12 | AP: 0.1416±0.003 -> 0.1474±0.003 | Helped |
| use UGRFS instead of the strongest listed baseline | UGRFS · 3Sources | AP: 0.4623±0.035 -> 0.4728±0.055 | Helped |
| add sample confidence (ablation) | UGRFS · yeast | AP: 0.6616 -> 0.6725 | Helped |
| add sample confidence (ablation) | UGRFS · MIRFlickr | AP: 0.6646 -> 0.6770 | Helped |
| add sample confidence (ablation) | UGRFS · IAPRTC12 | AP: 0.1471 -> 0.1474 | Helped |
| retain view variation in the initial global-view instead of ignoring it (ablation) | UGRFS · yeast | AP: 0.6603 -> 0.6725 | Helped |
| retain view variation in the initial global-view instead of ignoring it (ablation) | UGRFS · MIRFlickr | AP: 0.6729 -> 0.6770 | Helped |
| retain view variation in the initial global-view instead of ignoring it (ablation) | UGRFS · IAPRTC12 | AP: 0.1472 -> 0.1474 | Helped |
Surprising findings
- Ablation UGRFS_v2 (ignoring view variation in the initial global view) occasionally performs worse than UGRFS_v3 (no global-view reconstruction), highlighting the importance of view weighting.
Limitations
- Parameter sensitivity and fixed-range tuning (10^-3 to 10^3) for all methods; reliance on predefined sigma and neighborhood size q.
- Quadratic-to-cubic complexity in number of samples n may limit scalability.
- Some reported improvements over second-best methods are within standard deviations.
Experimental setup
Datasets: yeast SCENE VOC07 MIRFlickr IAPRTC12 3Sources
Benchmarks: yeast SCENE VOC07 MIRFlickr IAPRTC12 3Sources
Metrics: Average Precision Coverage Hamming Loss Ranking Loss
Baselines: M2LD MSFS MoRE MRDM MIFS CLML
Builds on: l2,1-norm regularization graph Laplacian smoothness multi-label feature selection
Closest papers by meaning
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.88)
- Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.88)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.86)
- Noisy Label Calibration for Multi-View Classification (AAAI 2025 · similarity 0.86)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.85)
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 2025 · 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.