Research map / Label-efficient and robust learning / Recognition and long-tail

Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Uncertainty-aware multi-view learning · 21 citations (OpenAlex)

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 TUNED, an evidential multi-view classifier that combines per-view local feature-neighborhood graphs (GCN with adaptive CAN adjacency), a shared consensus evidence extractor, and a selective MRF for fusing view evidence, evaluated on eight multi-view datasets under normal and synthetic conflict settings.

Problem

Multi-view classification suffers from cross-view conflicts and uncertainty; existing evidential deep learning (EDL) methods fuse views with the Dempster-Shafer rule, which is sensitive to conflicting evidence and ignores neighborhood structure.

Approach

Replace Dempster-Shafer fusion with a selective Markov random field that weights and links views by similarity (thresholded edges), combined with local per-view GCN neighborhood features and a sampled global consensus evidence conditioned on local structures, trained with an evidential loss plus a consistency loss.

Key results

• On normal test sets, TUNED reaches 96.83% on PIE and 99.00% on CUB versus best baselines of 94.71% and 98.50%.
• On conflictive test sets, TUNED improves Animal accuracy to 90.90% from a best baseline of 66.00% (+37.73%) and Scene15 to 67.22% (+18%).
• Ablation on HandWritten shows S-MRF fusion (99.20% normal) outperforms arithmetic averaging (96.25%) and DST (95.75%).

Measured effects reported in the paper

Surprising findings

• On the Animal dataset, conflictive views yield higher accuracy than normal test sets for several methods, which the authors attribute to parameter sensitivity amplified by fewer views.
• On Scene15 under conflict, the DST variant scores higher than the S-MRF-free ablations on ACC in some settings, and the Scene15 normal-set result of TUNED is below ECML and TMDL-OA on accuracy for some datasets (HandWritten, Scene15 among others).

Limitations

• Not evaluated on incomplete multi-view data with missing views.
• Conflictive test sets are synthetically constructed by the authors, so real-world conflict may differ.
• Performance is sensitive to hyperparameters such as the threshold tau and the fusion functions.

Experimental setup

Datasets: PIE HandWritten Scene15 Caltech101 CUB Animal ALOI NUS-WIDE-OBJECT

Benchmarks: PIE HandWritten Scene15 Caltech101 CUB Animal ALOI NUS-WIDE-OBJECT

Metrics: accuracy F1-score AUC

Baselines: EDL DCCAE CPM-Nets DUA-Nets TMC TMDL-OA ECML RMVC

Builds on: Evidential Deep Learning Trusted Multi-view Classification (TMC) Evidential Conflictive Multi-view Learning (ECML) Clustering-with-Adaptive-Neighbors (CAN) Dempster-Shafer theory

Closest papers by meaning

• Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 2025 · similarity 0.86)
• Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.85)
• Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.85)
• Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.85)
• Uncertainty-Aware Global-View Reconstruction for Multi-View Multi-Label Feature Selection (AAAI 2025 · similarity 0.85)
• Hierarchical Consensus Network for Multiview Feature Learning (AAAI 2025 · similarity 0.85)