Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Multi-View Classification Reliability with Adaptive Rejection
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 RAR, an evidence-theory multi-view classifier that adds a distribution-free adaptive rejection mechanism (conformal-style p-values on uncertainty) into the view fusion to reject unreliable views and control rejection risk.
Problem
Evidence-based uncertainty-aware multi-view classification gives uncertainty scores whose magnitude varies with view distribution, making it hard to decide whether to trust predictions and letting low-quality views degrade fused performance.
Approach
Treat view-level rejection as a multiple-testing problem with calibration-set p-values to get a distribution-free uncertainty threshold, and fuse subjective-logic opinions only from accepted views so rejected views are masked out of the fusion.
Key results
- RAR reaches 95.65% accuracy on ANIMAL, 99.11% on HAND, 95.19% on CUB, 75.69% on SCENE, 95.00% on MRNet and 97.62% on CAL, outperforming the listed baselines on all six datasets (Table 2).
- Ablation (Table 1) shows fusion, rejection and KL regularization each add accuracy, e.g. on ANIMAL from 90.46% base to 95.65% with all components.
- With a Gaussian-noised view, RAR consistently outperforms fusion without adaptive rejection across all datasets (Figure 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RAR instead of the strongest baseline, SMDC | Multi-view classification · ANIMAL | ACC (%): 93.46±0.00 -> 95.65±0.36 | Helped |
| Use RAR instead of the strongest baseline, SMDC | Multi-view classification · HAND | ACC (%): 98.89±0.10 -> 99.11±0.27 | Helped |
| Use RAR instead of the strongest baseline, ECML | Multi-view classification · CUB | ACC (%): 92.67±0.07 -> 95.19±1.59 | Helped |
| Use RAR instead of the strongest baseline, ECML | Multi-view classification · SCENE | ACC (%): 73.84±0.19 -> 75.69±1.35 | Helped |
| Use RAR instead of the strongest baseline, TMDOA | Multi-view classification · MRNet | ACC (%): 94.01±1.22 -> 95.00±0.10 | Helped |
| Use RAR instead of the strongest baseline, SMDC | Multi-view classification · CAL | ACC (%): 96.40±0.10 -> 97.62±0.37 | Helped |
| Add multi-view fusion (ablation) | RAR · ANIMAL, HAND, CUB, SCENE, MRNet, CAL | ACC (%): 90.46±0.60 -> 94.17±0.20 (ANIMAL); 94.44±0.83 -> 98.23±0.35 (HAND); 87.28±0.46 -> 93.74±1.31 (CUB); 69.08±1.05 -> 73.95± | Helped |
| Add adaptive rejection to fusion (ablation) | RAR · ANIMAL, HAND, CUB, SCENE, MRNet, CAL | ACC (%): 94.17±0.20 -> 94.91±0.33 (ANIMAL); 98.23±0.35 -> 98.78±0.28 (HAND); 93.74±1.31 -> 94.37±0.17 (CUB); 73.95±1.15 -> 74.03± | Helped |
| Add KL-divergence regularization to the full fusion and rejection method (ablation) | RAR · ANIMAL, HAND, CUB, SCENE, MRNet, CAL | ACC (%): 94.91±0.33 -> 95.65±0.36 (ANIMAL); 98.78±0.28 -> 99.11±0.27 (HAND); 94.37±0.17 -> 95.19±1.59 (CUB); 74.03±1.60 -> 75.69± | Helped |
Limitations
- Assumes the calibration set is drawn from the same distribution as training data and the training set is clean.
- Rejection threshold is only meaningful when calibration size allows tau >= 1/(l+1); small tau requires large calibration sets.
- Gains on noise experiments depend on the synthetic Gaussian noise setting and may shrink as noise saturates.
Experimental setup
Datasets: ANIMAL HAND CUB SCENE MRNet CAL
Benchmarks: ANIMAL HAND CUB SCENE MRNet CAL
Metrics: accuracy
Baselines: DCCA DCCAE CPM-Nets DTCCA DUA-Nets MVTCAE TMC TMDOA ETMC SMDC ECML
Builds on: TMC subjective logic Dempster-Shafer theory conformal prediction p-values
Closest papers by meaning
- Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 2025 · similarity 0.87)
- Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.87)
- Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.87)
- Trusted Multi-View Classification with Expert Knowledge Constraints (ICML 2025 · similarity 0.85)
- Reliable Conflictive Multi-View Learning (AAAI 2024 · similarity 0.85)
- Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (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.