REFS: Robust EEG feature selection with incomplete labels for multi-dimensional emotion recognition

OBJECTIVE: Compared to single-type features, multi-type electroencephalography (EEG) features provide a multi-level representation for analyzing multi-dimensional emotions. However, the high dimensionality of multi-type EEG features, combined with the relatively small number of high-quality EEG samples, poses challenges such as classifier overfitting and suboptimal real-time performance in the multi-dimensional emotion recognition task. As the number of EEG channels increases, the issue will become increasingly severe. Moreover, practical applications of affective brain-computer interface frequently encounters partial absence of multi-dimensional emotional label data due to diverse emotional self-assessment models, and ambiguity and variability in individual emotion perception. APPROACH: To address these challenges, this study proposes a novel robust EEG feature selection method for incomplete multi-dimensional emotion recognition. The method leverages adaptive orthogonal latent semantic indexing to reconstruct the multi-dimensional emotional label space through second-order and higher-order relations, which could reduce the negative impact of missing values and outliers on label reconstruction. Simultaneously, it employs least squares regression with graph-based manifold learning regularization and global feature redundancy minimization regularization to enable EEG feature subset selection despite missing information, ultimately achieving robust and fast EEG-based multi-dimensional emotion recognition. MAIN RESULTS: The proposed method was validated on three multi-dimensional emotional datasets with varying channel configurations, namely DREAMER, DEAP and HDED, under different label missing ratios. Experimental results show that the proposed method consistently outperforms thirteen state-of-the-art feature selection methods in terms of robustness for EEG emotional feature selection across various missing ratios. SIGNIFICANCE: The proposed method provides a robust and efficient solution for feature selection in multi-dimensional emotion recognition under incomplete EEG label scenarios, delivering reliable performance for practical affective brain-computer interface applications.