Data Generation for Enhancing EEG-Based Emotion Recognition: Extracting Time-Invariant and Subject-Invariant Components With Contrastive Learning

• Published in: IEEE transactions on consumer electronics Vol. 71; no. 1; pp. 1371 - 1384
• Main Authors: Wan, Zhijiang, Yu, Qianhao, Dai, Wujie, Li, Siyue, Hong, Jin
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.02.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; AIGC; Artificial intelligence; Brain modeling; contrastive learning; Data generation; Data models; Deep learning; EEG; Electroencephalography; Emotion recognition; Emotions; Invariants; Machine learning; Multitasking; Performance enhancement; Training; Workflow
• ISSN: 0098-3063; 1558-4127
• DOI: 10.1109/TCE.2024.3414154

The utilization of Artificial Intelligence for Generative Content (AIGC) has emerged as an effective and sophisticated approach for generating synthetic Electroencephalography (EEG) signals. This approach proves beneficial in augmenting EEG data and enhancing the performance of deep learning (DL) methods for emotion recognition. However, the temporal non-stationary nature and inter-subject variability of EEG signals still pose a great challenge for the practical applications of EEG-based emotion recognition. To address the challenges, we propose a novel data generation workflow that combines multi-task learning. This workflow incorporates a generative pre-trained Transformer (EEGPT) to generate time-invariant components from raw EEG data. Additionally, we introduce a model training strategy called Contrastive Learning method for Time-invariant and Subject-invariant (CLTISI) EEG data generation. This strategy aligns inter-subject data into a shared high-dimensional space, imparting subject-invariant characteristics to the generated data. Experimental results demonstrate that the generated data not only improves the generalization of DL models but also adapts effectively to novel emotional stimuli. Furthermore, the CLTISI strategy enables DL models to maintain stable performance across diverse datasets. The regions identified as crucial for emotion recognition by the EEGPT model offer valuable insights into the neural mechanisms of human emotion processing.