Research map / Graph learning / Graph self-supervised learning
Multi-View Multi-Label Classification via View-Label Matching 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 VAMS, which reformulates multi-view multi-label classification as selecting view-label matchings on a graph network over an object graph and a label graph.
Problem
Multi-view multi-label classification requires fusing cross-view consensus and view-specific information while modeling multi-label correlations and assigning correct labels per object; prior methods split consensus and specificity learning and do not explicitly model view-label correspondence.
Approach
Builds a unified view-label matching graph (object graph with view nodes plus KNN intra-view neighbors, fully connected label graph, and matching edges between every view and label), applies a graph network block for node and edge convolution, and scores each view-label matching, averaging across views, trained with a class-weighted cross-entropy loss.
Key results
- VAMS achieves the best Average Precision on all datasets, outperforming the second-best method by 2% to 4% on Average Precision and 2% to 14% on Micro-F1.
- Friedman test rejects the null hypothesis of equal performance at 0.05 significance; Bonferroni-Dunn test shows VAMS ranks 1st on all metrics.
- Ablation shows removing inter-view connections (VAMS-w/o InterConn) drops Micro-F1 on Emotions from 0.692 to 0.614, larger than removing label semantic correlations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intra-view K-nearest neighbors (ablation) | VAMS · Emotions | Micro-F1: 0.689±0.019 -> 0.692±0.015 | Helped |
| add intra-view K-nearest neighbors (ablation) | VAMS · Scene | Micro-F1: 0.746±0.018 -> 0.748±0.016 | Helped |
| add intra-view K-nearest neighbors (ablation) | VAMS · Corel5k | Micro-F1: 0.331±0.021 -> 0.372±0.005 | Helped |
| add inter-view connections (ablation) | VAMS · Emotions | Micro-F1: 0.614±0.058 -> 0.692±0.015 | Helped |
| add inter-view connections (ablation) | VAMS · Scene | Micro-F1: 0.713±0.022 -> 0.748±0.016 | Helped |
| add inter-view connections (ablation) | VAMS · Corel5k | Micro-F1: 0.315±0.047 -> 0.372±0.005 | Helped |
| add label semantic correlations (ablation) | VAMS · Emotions | Micro-F1: 0.682±0.016 -> 0.692±0.015 | Helped |
| add label semantic correlations (ablation) | VAMS · Scene | Micro-F1: 0.740±0.016 -> 0.748±0.016 | Helped |
| add label semantic correlations (ablation) | VAMS · Corel5k | Micro-F1: 0.340±0.035 -> 0.372±0.005 | Helped |
Surprising findings
- Removing label semantic correlations hurts less than removing intra-view KNN or inter-view connections, suggesting multi-view fusion matters more than label correlation in this setup.
Limitations
- Evaluated only on relatively small classic MVML benchmarks with features rather than modern foundation-model representations.
- Some metric comparisons show VAMS not uniformly best (e.g., Hamming Loss on some datasets, Emotions Ranking Loss against LSPC/IMvMLC).
Experimental setup
Datasets: Emotions Scene Yeast Plant Human Corel5k Pascal
Benchmarks: Emotions Scene Yeast Plant Human Corel5k Pascal
Metrics: Hamming Loss Ranking Loss Coverage Average Precision Micro-F1 Macro-F1
Baselines: LrMMC LSPC SIMM FIMAN IMvMLC ML-BVAE
Builds on: GN Block (graph network) LrMMC SIMM FIMAN
Closest papers by meaning
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.90)
- Tensorized Multi-View Multi-Label Classification via Laplace Tensor Rank (ICML 2025 · similarity 0.90)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.90)
- Partial Multi-View Multi-Label Classification via Semantic Invariance Learning and Prototype Modeling (ICML 2024 · similarity 0.90)
- Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · similarity 0.88)
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 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.