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

Trusted Multi-View Classification with Expert Knowledge Constraints

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Uncertainty-aware multi-view learning

Read the paper Check 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 TMCEK, a trusted multi-view classifier for sleep staging that embeds Gabor expert-knowledge kernels for feature interpretability and a distribution-aware subjective-opinion uncertainty measure for decision-level confidence.

Problem

Trusted multi-view classification based on Dempster-Shafer/subjective logic lacks feature-level explanations and its uncertainty depends only on the total evidence magnitude, ignoring evidence distribution, leading to inaccurate confidence estimates.

Approach

Parameterized Gabor convolution kernels in the first layer inject sleep-EEG expert knowledge for interpretability, and the uncertainty mass is redefined to depend on both evidence sum and a Gini-based concentration measure, with a theoretical proof that it is more sensitive than standard subjective opinions.

Key results

On Sleep-EDF 20, TMCEK reaches 85.0% accuracy, 80.2 MF1 and 0.80 kappa, the best among compared sleep staging methods.
On Sleep-EDF 78, TMCEK reports 81.4% accuracy versus 81.3% for AttnSleep, and 84.3% accuracy on SHHS versus 84.2% for AttnSleep.
On multi-view benchmarks TMCEK reports the highest accuracy on HandWritten (98.80), Scene15 (74.84), CUB (94.67) and PIE (97.06).

Measured effects reported in the paper

What was changed	Where	Measured	Effect
Use TMCEK instead of DFSC	Sleep staging · Sleep-EDF 20	accuracy: 84.44 -> 85.0	Helped
Use TMCEK instead of AttnSleep	Sleep staging · Sleep-EDF 78	macro F1: 75.1 -> 77.5	Helped
Use TMCEK instead of AttnSleep	Sleep staging · SHHS	macro F1: 75.3 -> 78.0	Helped
Use TMCEK instead of CCML	Multi-view classification · Handwritten	accuracy: 98.75 ± 0.27 -> 98.80 ± 0.76	No clear effect
Use TMCEK instead of CCML	Multi-view classification · Scene15	accuracy: 72.60 ± 0.87 -> 74.84 ± 0.39	Helped
Use TMCEK instead of CCML	Multi-view classification · CUB	accuracy: 94.58 ± 1.30 -> 94.67 ± 1.55	No clear effect
Use TMCEK instead of CCML	Multi-view classification · PIE	accuracy: 94.56 ± 1.83 -> 97.06 ± 1.80	Helped
Add trusted learning (ablation)	TMCEK on Sleep-EDF 20 · Sleep-EDF 20 with Gaussian noise on the time-domain test view	accuracy	Helped
Account for evidence distribution in uncertainty estimation	TMCEK · Sleep-EDF 20 noisy test instances	uncertainty	Helped

Surprising findings

Under Gaussian noise, the original subjective-opinion uncertainty stays constant when evidence is redistributed, which motivates the distribution-aware formulation.

Limitations

Some learned Gabor kernels are redundant or unoptimized, so interpretability is partial; diversity regularization is left for future work.
Only evaluated on sleep staging datasets for expert-knowledge instantiation; generalization to other expert domains is untested.
Improvements over the strongest baselines on Sleep-EDF 78 and SHHS are small.

Experimental setup

Datasets: Sleep-EDF 20 Sleep-EDF 78 SHHS HandWritten Scene15 CUB PIE

Benchmarks: Sleep-EDF 20 Sleep-EDF 78 SHHS HandWritten Scene15 CUB PIE

Metrics: accuracy macro F1 Cohen's kappa

Baselines: DeepSleepNet ARNN+SVM SleepEEGNet ResNetLSTM MultiTaskCNN DFSC ResAtten AttnSleep MISC EDL DCCAE CAML

Builds on: evidential deep learning trusted multi-view classification RCML subjective logic

Closest papers by meaning

Enhancing Testing-Time Robustness for Trusted Multi-View Classification in the Wild (CVPR 2025 · similarity 0.89)
Enhancing Multi-View Classification Reliability with Adaptive Rejection (AAAI 2025 · similarity 0.85)
Navigating Conflicting Views: Harnessing Trust for Learning (ICML 2025 · similarity 0.85)
Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.85)
Deep Fuzzy Multi-view Learning for Reliable Classification (ICML 2025 · similarity 0.84)
Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.81)

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.

Open scime draft