CoAdapt: Collaborative Adaptation Between Latent EEG Feature Representation and Annotation for Emotion Decoding

Electroencephalogram (EEG) data contains rich neurophysiological information that can objectively express the emotional state of human beings. However, the inherent EEG characteristics such as non-stationarity and weakness, combined with the possible limited immersion and carry-over effect of subjects during data-collection experiments, may cause that the semantic meaning of extracted EEG feature vector cannot well match its annotated emotional state, dubbed the ‘feature-label inconsistency dilemma in EEG-based emotion decoding. To this end, this paper proposes to alleviate the side effect of ‘feature-label inconsistency from both feature and label aspects. On one hand, we explore more meaningful emotion-related EEG representation by the latent low-rank representation. On the other hand, enhance the correspondence between explored EEG representation and its annotated emotional state by a label dragging strategy. As a result, a collaborative adaptation (CoAdapt) model between latent EEG feature representation and its annotation is formed for efficient emotion decoding, which is implemented within the semi-supervised framework to better capture the properties of both the labeled and unlabeled EEG data. Experimental results on three publicly available datasets, SEED-IV, SEED-V and MPED, depict that 1) CoAdapt achieves better emotion recognition performance in comparison with some related models; 2) the improvements of inter-class separability and label margin are empirically evaluated, indicating the effectiveness of the purified EEG feature representation and rectified emotion annotation; 3) some task-related results are identified from data-driven perspective, including the emotion carry-over effect and the discriminative spatial patterns in emotion decoding.