EEGMatch: Learning With Incomplete Labels for Semisupervised EEG-Based Cross-Subject Emotion Recognition.

Area of Science:

• Neuroscience and Artificial Intelligence
• Machine Learning for Affective Computing

Background:

• Electroencephalography (EEG) offers objective emotion recognition but faces limitations due to scarce labeled data.
• Label scarcity hinders the widespread application of EEG-based emotion recognition systems.

Purpose of the Study:

• To propose a novel semisupervised transfer learning framework, EEGMatch, to overcome the label scarcity problem in EEG emotion recognition.
• To effectively leverage both labeled and unlabeled EEG data for enhanced model training and performance.

Main Methods:

• EEG-Mixup-based data augmentation to generate additional valid samples.
• A semisupervised two-step pairwise learning approach integrating prototypewise and instancewise learning.
• Semisupervised multidomain adaptation to align data representations across different domains, mitigating distribution mismatch.

Main Results:

• EEGMatch demonstrated superior performance compared to state-of-the-art methods across three benchmark EEG databases (SEED, SEED-IV, SEED-V).
• Significant improvements observed under various incomplete label conditions, highlighting the framework's robustness.
• Achieved 5.89% improvement on SEED, 0.93% on SEED-IV, and 0.28% on SEED-V datasets.

Conclusions:

• The proposed EEGMatch framework effectively addresses the label scarcity challenge in EEG-based emotion recognition.
• The integration of data augmentation, pairwise learning, and domain adaptation proves effective for leveraging unlabeled data.
• EEGMatch offers a promising solution for developing more practical and widely applicable EEG emotion recognition systems.