Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Incomplete Multi-View Learning Network with Insufficient Label Information
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 DIMvLN, a deep graph-network and semi-supervised framework that jointly recovers missing views and assigns pseudo-labels to handle simultaneous view missing and label scarcity in multi-view classification.
Problem
Multi-view classification where many instances lack some view features and only a small fraction of training instances are labeled (incomplete multi-view semi-supervised classification).
Approach
A GNN-based completion module recovers missing view features via transferred k-NN graphs, while a pseudo-labeling mechanism (agreement between a deep classifier and a similarity-based classifier) supplies extra supervision that guides the completion; autoencoders with instance- and category-level contrastive losses learn cross-view consistent representations, trained end-to-end.
Key results
- On Caltech101-20 with PER=0% and LER=10%, DIMvLN reaches 91.28% ACC versus 82.94% for the best baseline CPM-Nets.
- On Animal with PER=90%, DIMvLN achieves 28.11% ACC versus 17.46% for DCP.
- When LER drops from 30% to 5% on ALOI and Out-Scene, DIMvLN's performance declines by less than 1%, while other methods decline by more than 5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DIMvLN instead of the strongest reported baseline | Caltech101-20 · Caltech101-20 (PER=90%, LER=10%) | ACC: 73.85 -> 85.23 | Helped |
| use DIMvLN instead of the strongest reported baseline | CUB · CUB (PER=90%, LER=10%) | ACC: 56.58 -> 72.58 | Helped |
| use DIMvLN instead of the strongest reported baseline | Wikipedia · Wikipedia (PER=90%, LER=10%) | ACC: 38.74 -> 44.60 | Helped |
| use DIMvLN instead of the strongest reported baseline | ALOI · ALOI (PER=90%, LER=10%) | ACC: 54.49 -> 76.96 | Helped |
| use DIMvLN instead of the strongest reported baseline | Out-Scene · Out-Scene (PER=90%, LER=10%) | ACC: 52.79 -> 71.75 | Helped |
| use DIMvLN instead of the strongest reported baseline | Animal · Animal (PER=90%, LER=10%) | ACC: 20.93 -> 28.11 | Helped |
| add GNN-based completion (ablation) | DIMvLN · ALOI (PER=60%, LER=10%) | ACC: 55.44 -> 83.74 | Helped |
| add semi-supervised pseudo-labeling (ablation) | DIMvLN · ALOI (PER=60%, LER=10%) | ACC: 82.04 -> 83.74 | Helped |
| add GNN-based completion (ablation) | DIMvLN · Animal (PER=60%, LER=10%) | ACC: 20.37 -> 33.86 | Helped |
| add semi-supervised pseudo-labeling (ablation) | DIMvLN · Animal (PER=60%, LER=10%) | ACC: 31.50 -> 33.86 | Helped |
Surprising findings
- The ablation shows the semi-supervised pseudo-labeling module contributes more to performance than the GNN completion module on ALOI and Animal.
Limitations
- Hyperparameters (lambda1, lambda2, alpha, k) are fixed by grid search rather than adaptively chosen.
- Evaluated only on six public datasets with simulated missing views and labels.
Experimental setup
Datasets: Caltech101-20 CUB Wikipedia ALOI Out-Scene Animal
Benchmarks: Caltech101-20 CUB Wikipedia ALOI Out-Scene Animal
Metrics: ACC F1 PREC AUC
Baselines: SLIM AMSC AMMSS CPM-Nets DCP TMC
Builds on: DCP EE-IMVC AMSC SLIM
Closest papers by meaning
- Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding (CVPR 2025 · similarity 0.89)
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.89)
- Attention-Induced Embedding Imputation for Incomplete Multi-View Partial Multi-Label Classification (AAAI 2024 · similarity 0.88)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 2025 · similarity 0.88)
- A Two-Stage Information Extraction Network for Incomplete Multi-View Multi-Label Classification (AAAI 2024 · similarity 0.87)
- View-Category Interactive Sharing Transformer for Incomplete Multi-View Multi-Label Learning (CVPR 2024 · similarity 0.87)
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.