EEG Emotion Recognition in Brain Computer Interface Through Neural Networks and Adaptive Machine Learning

Human Computer Interaction (HCI), affective computing, and mental health have experienced a sharp rise in use in EEG signal-based emotion classfication. Accurate and generalized emotion finding from Electroencephalography signal is a critical step for realizing the practical usage of functional Brain-Computer Interfaces (BCI). This work tracks a neat flowchart designed for dimensional emotion classification (Valence and Arousal) using public DEAP dataset. The primary innovation here is the use of a Multi-Task Time-Series Transformer (TST) model, which uses a Self-Attention method to dynamically filter subject specific noise and extract general neural patterns. The developed Time-Series Transformer framework predicts both Valence and Arousal simultaneously, by leveraging the intrinsic relation between two emotion factor. We used the Subject-Independent (LOSOCV) protocol combined with a non-leaking feature scaling strategy to ensure test data integrity. Experimental results confirm that the TST model’s effectiveness about 78.05% mean accuracy which was achieved for Valence classification, demonstrating state-of-the-art generalization capability across subjects. And the Arousal dimension came to 50.94% in the generalized test, experimentally confirming that Arousal-related features are highly subject-specific and thus demand individu-alized calibration. The proposed methodology establishes a new, verifiable benchmark for generalized EEG emotion recognition and automatically highlighting the potential and current limitations of BCI-based affective computing.