BCIwiki (bciwiki.com) — According to arXiv, a study uploaded on November 1, 2026, reports that Researchers at Sun Yat-sen University proposed the VARY framework, a framework for improving EEG-based emotion recognition when emotion labels are coarse. Its formal publication and peer-review status remain unknown. The reported results come from label-noise learning tests on public datasets.
The authors explain that fine-grained emotional annotation is costly, so most datasets assign one global label to each long EEG trial. For training, recordings are commonly split into short temporal windows that inherit the same label. Yet emotions may fluctuate within a trial, leaving some windows weakly aligned with, or different from, the overall emotional state and making their assigned labels unreliable.
VARY uses two-stage sample selection to identify EEG windows with reliable emotional labels. It also introduces MixAug, an EEG-specific data augmentation method intended to improve robustness when supervision is limited and unreliable. Evaluation covered two public datasets, three network architectures, and multiple label-noise settings.
The authors report maximum accuracy gains of 11.54% on DREAMER and 5.39% on SEED compared with the best-performing baseline. These maxima do not imply equal gains across architectures or label-noise settings. The percentages retain the source's wording and have not been converted to percentage points.