Research map / Label-efficient and robust learning / Recognition and long-tail
Biased Incomplete Multi-View 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 RIML, a learning-free imputation framework for incomplete multi-view learning under class-dependent (biased) view-missing patterns, using OT-based inter-class association to calibrate view-specific Gaussians and a category-aware focal loss.
Problem
Incomplete multi-view learning where view-missing patterns depend on class (biased) rather than uniformly random, causing data-poor classes to have few samples and data-rich classes to have few views.
Approach
Use optimal transport between meta-classes and rare classes to learn a class-association matrix that calibrates view-specific Gaussian distributions, then impute missing views by sampling from these calibrated distributions; train with a focal loss plus a category-aware marginal term.
Key results
- On Handwritten with no missing views, RIML reaches 0.9650 accuracy versus 0.8705 for the best baseline CMP-Nets.
- On Animal with complete views, RIML achieves about 9.7% improvement over the second-best model.
- RIML remains robust at maximum missing rate 0.5, e.g., 0.9290 on Handwritten and 0.5954 on Scene15, outperforming all compared methods across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RIML instead of CMP-Nets | Incomplete multi-view classification · HandWritten, η = 0.3 | accuracy: 0.8615 ± 0.01 -> 0.9430 ± 0.02 | Helped |
| Use RIML instead of DeepIMV | Incomplete multi-view classification · Scene15, η = 0.3 | accuracy: 0.4609 ± 0.02 -> 0.6425 ± 0.03 | Helped |
| Use RIML instead of CMP-Nets | Incomplete multi-view classification · Animal, η = 0.3 | accuracy: 0.6630 ± 0.03 -> 0.7670 ± 0.02 | Helped |
| Use RIML instead of CMP-Nets | Incomplete multi-view classification · YaleB, η = 0.3 | accuracy: 0.7172 ± 0.04 -> 0.8571 ± 0.01 | Helped |
| Use RIML instead of CMP-Nets | Incomplete multi-view classification · BRCA, η = 0.3 | accuracy: 0.7760 ± 0.04 -> 0.8440 ± 0.01 | Helped |
| Add inter-class association calibration to distribution-sampling imputation (ablation) | RIML · Five multi-view datasets, η = 0.3 | accuracy | Helped |
| Replace standard focal loss with category-aware focal loss (ablation) | RIML · Five multi-view datasets, η = 0.3 | accuracy | Helped |
Surprising findings
- Learning-free sampling-based imputation outperforms learned deep IMVL methods such as DeepIMV and CMP-Nets even with complete data, attributed largely to the enhanced focal loss.
Limitations
- Assumes rare classes have one sample per view when formulating the OT problem, and relies on a predefined threshold to define meta-classes.
- Test-time missing views are imputed via k-nearest neighbors rather than the learned distributions.
- Sensitivity to beta is analyzed only briefly; the sensitivity to class-association quality on datasets with less clear class associations (e.g., Scene15) reduces gains.
Experimental setup
Datasets: HandWritten Scene15 Animal YaleB BRCA
Benchmarks: HandWritten Scene15 Animal YaleB BRCA
Metrics: accuracy
Baselines: UIMC MIWAE CPSPAN DeepIMV CPM-Nets
Builds on: focal loss Sinkhorn algorithm few-shot distribution calibration
Closest papers by meaning
- Learning Compact Semantic Information for Incomplete Multi-View Missing Multi-Label Classification (ICML 2025 · similarity 0.86)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.85)
- Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs (ICLR 2025 · similarity 0.84)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.84)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.84)
- Incomplete Multi-View Multi-Label Classification via Diffusion-Guided Redundancy Removal (AAAI 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.