TWEN: EEG Emotion Recognition Model Based on Weakly Supervised Learning Framework with Two-Phase Multitask Autoencoder

Emotions significantly influence human cognition, behavior, and social interactions, making accurate recognition essential in Human-Computer Interaction (HCI) applications. This study addresses challenges in EEG-based emotion recognition, particularly inter-subject variability and label noise, which hinder the development of robust and generalized models. We propose a robust Two-phase Weakly Supervised Emotion Network (TWEN), a novel deep learning model designed to enhance emotion recognition. TWEN incorporates a Two-phase Multitask Autoencoder to mitigate inter-subject variability and a Top-k Selection method to reduce label noise. The model captures both local and global temporal features of EEG signals through an innovative fusion of attention mechanisms, ensuring accurate classification of emotions over varying durations. Evaluations on the THU-EP dataset demonstrate that TWEN outperforms state-of-the-art models, achieving a classification accuracy of 60.8%, with a standard deviation of 4.07%.