Multitask Autoencoder-Based Two-Phase Framework Using Multilevel Feature Fusion for EEG Emotion Recognition

Emotion recognition has emerged as a active research area, gaining relevance from advancements in deep learning. This study focuses on using electroencephalogram (EEG) data for emotion recognition and addresses the challenge of subject-dependent variability in EEG-based emotion recognition by proposing a novel architecture that employs multilevel feature fusion and a multitask autoencoder-based two-phase framework. The first phase generates classspecific data, while the second phase uses these for model training. The proposed model was validated using the SEED dataset and demonstrated state-of-the art perforamnce with an accuracy of 99.4 % in a subject-independent setting.