An adversarial discriminative temporal convolutional network for EEG-based cross-domain emotion recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Traditional electroencephalogram (EEG) based emotion recognition struggles with domain shift issues, leading to poor performance across different subjects or datasets.
• Existing classification methods lack effective domain adaptation capabilities for EEG data.
• Addressing the distribution differences between source and target domains is crucial for robust emotion recognition.

Purpose of the Study:

• To propose a novel domain adaptation strategy for improving cross-domain emotion recognition using EEG signals.
• To develop a method that ensures feature representation invariance across different domains.
• To enhance the performance of EEG-based emotion recognition in cross-subject and cross-dataset scenarios.

Main Methods:

• Introduced adversarial discriminative-temporal convolutional networks (AD-TCNs) as a novel domain adaptation strategy.
• Utilized temporal convolutional networks (TCNs) as feature encoders, leveraging the temporal attributes of EEG data.
• Employed adversarial learning to ensure feature graph representation invariance across domains.

Main Results:

• Achieved highest accuracies for valence and arousal dimensions in cross-subject experiments on DREAMER and DEAP datasets.
• Demonstrated competitive performance in cross-dataset experiments, with two task groups reaching accuracies of 62.65% and 62.36%.
• Outperformed state-of-the-art methods in comparable protocols, validating the effectiveness of AD-TCNs.

Conclusions:

• AD-TCNs provide an effective solution for EEG-based cross-domain emotion recognition.
• The proposed method successfully addresses domain shift challenges in EEG emotion recognition.
• This approach represents a significant advancement for realizing robust and generalizable EEG emotion recognition systems.