.png){kind=icon}
.png){kind=icon}
Return
Imbalance-aware spectral classification via logit-label adjustment and multi-expert mutual learning
.png){kind=icon}
PRE
AI
W
H
T
A
S
X
DOI：10.1016/j.aca.2026.346003.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We propose a Logit-Label Adjustment (LLA) strategy that recalibrates both logits and labels to mitigate decision boundary bias toward many-shot classes and alleviate overfitting. • To enhance model robustness, we present a Cosine-Annealed Data Augmentation (CADA) strategy and concurrently introduce a Consistency Regularization (CR) technique to ensure stable gradient updates during data augmentation. • We design a Multi-Expert Mutual Learning (MEML) mechanism that enables experts to distill knowledge from one another, thereby enriching feature diversity and boosting overall recognition performance. • Extensive experiments on two bacterial spectral datasets and one cancer tissue spectral dataset across varying simulated imbalance scenarios demonstrate the effectiveness of our method and its consistent improvements over existing approaches, indicating its potential for class-imbalanced spectral analysis tasks.
Journal
IF:
6
Papers:
3.3W
Citations:
6.1W
Organization
No organization information available
Cited Papers