Cross-subject emotion recognition with loop adaptive adversarial transfer network.

Emotion recognition has broad application prospects in real life. The variability across subjects in emotion-related electroencephalogram (EEG) signals is still a significant challenge for the practical use of EEG-based emotion recognition despite recent findings that suggest EEG signals are informative and beneficial for recognizing emotions. This work proposes a loop adaptive adversarial transfer network (LATN) to enhance cross-subject emotion recognition ability. LATN is an effective cross-subject classifier, which consists of Structure-aware Associative Alignment (SAA), Inner and Outer Product Combination Strategy (IOPC strategy), and a pseudo-labeling method based on semi-supervised learning. Specifically, LATN enhances the consistency of source and target embeddings through SAA to minimize differences between subjects. At the same time, the IOPC strategy is used to capture the multimodal information of the data, to address the problem of ambiguity of decision boundaries. Furthermore, in order to lessen the negative transfer issue brought on by the selection of low-quality source domain data, pseudo-labels are chosen more judiciously. We ran many experiments using the publicly accessible datasets DEAP and SEED in addition to the private dataset ECPL to assess LATN's efficacy. The experimental findings show that the three-class accuracy on the ECPL dataset is 96.33%, the binary classification accuracy for arousal and valence on the DEAP dataset is 89.21 % and 76.12 %, respectively, and the three-class accuracy on the SEED dataset is 94.34 %. This technique outperforms existing algorithms and shows state-of-the-art performance on numerous datasets.