Emotion Recognition Using Multi-Scale Auto-Encoders with Cross Session Adoption

Emotion recognition from EEG (electroencephalography) signals is a challenging yet promising area of research, with applications ranging from mental health monitoring to adaptive human-computer interactions. Traditional approaches, such as those using Random Forest algorithms, have shown potential but often fall short in effectively capturing the complex temporal and spatial patterns inherent in EEG data. In this study, we propose a novel framework employing Multi-Scale Masked Autoencoders (MSMAE) combined with Convolutional Neural Networks (CNNs) for cross-session emotion recognition. Utilizing the Seed IV EEG dataset, our method leverages the multi-scale feature extraction capabilities of MSMAE to handle varying signal frequencies and the powerful pattern recognition abilities of CNNs to enhance classification accuracy. The MSMAE framework pre-trains the CNN by reconstructing the masked EEG signals at different scales, enabling it to learn robust and generalized features across different sessions. Comparative evaluations demonstrate that our proposed MSMAE-CNN model significantly outperforms the existing Random Forest algorithm, providing a more reliable and effective solution for emotion recognition in diverse and dynamic environments. This advancement not only highlights the potential of deep learning models in EEG-based emotion recognition but also sets a new benchmark for future research in this field.