Zero-Shot Learning for Multi-Class Emotion Recognition from EEG Signals Using Transformer-Based Models

Emotion recognition from EEG signals plays a significant role in affective computing, and it enables the systems to be aware of the human emotional state. This research proposes a novel methodology for zero-shot learning in multi-class emotion recognition from EEG signals, which address the issues like data scarcity, class imbalance, and generalization. A technique starting with a dataset acquisition method makes use of EEG-GANs in order to produce synthetic samples to balance class imbalance for poorly represented emotional classes. The introduction of DWC-TFS is innovatively used: the dynamic patterns in connectivity have been captured appropriately and the temporally noisy has reduced in the study. In order to extract features, DFCNs are applied in modeling time-varying relationships between EEG channels, keeping important temporal dependencies intact. For feature selection, Neuro-Symbolic Feature Selection is used by incorporating symbolic reasoning with deep learning in order to choose the most discriminative features. The classification task is done using a T-ZSC Transformer-Based Zero-Shot Classifier. The model will generalize across unseen emotional classes through the use of semantic embeddings and self-attention mechanisms. The proposed methodology achieves an accuracy of 96.2%, demonstrating robust performance in emotion recognition. This integrated approach offers a powerful solution for addressing challenges in EEG-based emotion recognition, ensuring improved class balance, generalization, and accuracy in real-world applications.