Unsupervised domain adaptation techniques based on auto-encoder for non-stationary EEG-based emotion recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG)-based emotion recognition faces challenges due to distribution mismatches between training and testing data from different sessions or subjects.
• Existing domain adaptation methods lack the power to handle non-stationary EEG signals effectively, limiting classification accuracy.
• Conventional classifiers struggle with direct EEG pattern classification due to these distribution discrepancies.

Purpose of the Study:

• To propose a novel domain adaptation method, the subspace alignment auto-encoder (SAAE), for EEG-based emotion recognition.
• To address the limitations of existing methods by combining nonlinear transformation with a consistency constraint for improved distribution alignment.
• To enhance the performance of EEG emotion recognition systems by reducing domain discrepancy and mitigating performance degradation across subjects and sessions.

Main Methods:

• Developed a unified framework integrating an auto-encoder network with a subspace alignment solution to create the SAAE component.
• Employed nonlinear transformation and a consistency constraint to align source and target domains in a shared representation space.
• Evaluated SAAE using a public EEG dataset with positive, neutral, and negative affective states, performing subject-to-subject and session-to-session evaluations.

Main Results:

• SAAE achieved a mean accuracy of 77.88% in subject-to-subject evaluations, surpassing the state-of-the-art TCA method (73.82%).
• In session-to-session evaluations, SAAE attained an average classification accuracy of 81.81%, an improvement of 1.62% over the best baseline (TCA).
• Experimental results demonstrated SAAE's effectiveness in decreasing domain discrepancy and reducing performance degradation compared to existing methods.

Conclusions:

• SAAE is a powerful and effective tool for EEG-based emotion recognition, particularly in scenarios with distribution shifts.
• The method successfully aligns data distributions, enabling more robust and accurate emotion classification across different subjects and sessions.
• SAAE offers a significant advancement in domain adaptation techniques for non-stationary biosignal processing.