BCIwiki (bciwiki.com) — Researchers affiliated with Beijing University of Technology and Anhui University, China, published REFS in the Journal of Neural Engineering on October 26, 2026. It combines label reconstruction with EEG feature selection for multidimensional emotion recognition when some labels are missing, with tests on three datasets under different missing-label ratios.
Multiple types of EEG features can represent emotion-related information at several levels. However, high feature dimensionality and relatively few high-quality samples can lead to classifier overfitting and suboptimal real-time performance. In affective brain-computer interface (BCI) applications, multi-dimensional emotional labels may also be partially absent because of differences in self-assessment models and ambiguity and variability in individual emotional perception. REFS combines label reconstruction with EEG feature selection to address these modeling challenges.
The method reconstructs incomplete emotion labels through relations among labels, reducing the effects of missing values and outliers. It also combines regression and graph-based methods to select EEG features while reducing redundancy among them.
Validation used DREAMER, DEAP and HDED, three multi-dimensional emotion datasets with different channel configurations, under varying label-missing ratios. Across the missing ratios evaluated, REFS consistently outperformed the 13 advanced feature-selection methods compared in robustness for EEG emotional feature selection. The authors describe the method as a robust and efficient approach for affective BCIs with incomplete labels; the reported validation remains limited to dataset experiments.