Attention-Based Transfer Enhancement Network for Cross-Corpus EEG Emotion Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• EEG-based emotion recognition faces poor cross-dataset generalization due to domain shifts.
• Traditional methods struggle with overfitting or bridging large dataset discrepancies.
• Developing robust models for cross-corpus EEG emotion recognition is crucial for practical applications.

Purpose of the Study:

• To propose a novel framework, the Cross-corpus Attention-based Transfer Enhancement network (CATE), to address the generalization challenge in EEG emotion recognition.
• To develop a two-stage framework with a dual-view self-supervised pre-training strategy for learning domain-invariant representations.
• To significantly improve the accuracy and robustness of cross-corpus EEG emotion recognition.

Main Methods:

• Introduced a two-stage framework: CATE.
• Employed a dual-view self-supervised pre-training strategy: Noise-Enhanced Representation Modeling (NERM) and Wavelet Transform Representation Modeling (WTRM).
• Utilized attention-based mechanisms in the supervised fine-tuning stage for classification.

Main Results:

• CATE achieved state-of-the-art performance on six transfer tasks across SEED, SEED-IV, and SEED-V datasets.
• Reported accuracies ranging from 68.01% to 81.65%.
• Outperformed prior methods by up to 15.65 percentage points, demonstrating superior generalization capabilities.

Conclusions:

• The proposed CATE framework effectively learns transferable features from distinct, synergistic views.
• CATE significantly advances the practical applicability of cross-corpus EEG emotion recognition.
• The dual-view pre-training strategy enhances model resilience to domain shifts and improves recognition accuracy.