Two-Level Domain Adaptation Neural Network for EEG-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Emotion recognition is crucial for human-computer interaction (HCI).
• Non-stationarity in electroencephalogram (EEG) signals poses a significant challenge for current emotion recognition models, leading to performance degradation over time.
• Developing robust models that can adapt to variations in EEG data across different conditions (e.g., days, subjects) is essential.

Purpose of the Study:

• To propose a novel two-level domain adaptation neural network (TDANN) for effective EEG-based emotion recognition.
• To address the challenge of non-stationarity in EEG signals and improve model generalization.
• To evaluate the domain-transfer performance of the TDANN model on both self-built and public datasets.

Main Methods:

• Extracted deep features from topological graphs of EEG signals using a deep neural network.
• Employed a two-level domain adaptation strategy: Maximum Mean Discrepancy (MMD) for feature distribution alignment and Domain Adversarial Neural Network (DANN) for class-center alignment.
• Evaluated the model's performance in cross-day and cross-subject transfer experiments using self-built and SEED datasets.

Main Results:

• The TDANN model demonstrated high accuracy in discriminating emotions across different days and subjects.
• On the self-built dataset, cross-day transfer achieved accuracies of 84% (sadness), 87.04% (anger), and 85.32% (fear).
• On the SEED dataset, cross-subject transfer achieved an average accuracy of 87.9%, outperforming WGAN-DA.

Conclusions:

• The proposed TDANN effectively handles domain transfer issues in EEG-based emotion recognition.
• The model's ability to preserve topological information and align feature distributions contributes to its robust performance.
• TDANN offers a promising solution for developing more reliable and adaptable EEG emotion recognition systems.