Improving Cross-Day EEG-Based Emotion Classification Using Robust Principal Component Analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Emotion recognition using electroencephalogram (EEG) is challenging due to significant day-to-day variability in recorded signals.
• Existing predictive models struggle with this variability, limiting real-world applications of emotion-aware frameworks.
• Mitigating inter-day EEG variability is crucial for reliable cross-day emotion classification.

Purpose of the Study:

• To develop and validate a method for reducing inter-day EEG variability to improve cross-day emotion classification.
• To investigate the effectiveness of a robust principal component analysis (RPCA)-based filtering strategy for this purpose.
• To assess the neurophysiological validity and machine-learning practicability of the proposed approach.

Main Methods:

• Proposed a robust principal component analysis (RPCA)-based signal filtering strategy.
• Applied the strategy to a five-day EEG dataset from 12 subjects during a music-listening task.
• Validated the method using a binary emotion classification task (happiness vs. sadness) with an add-day-in cross-validation scheme.

Main Results:

• RPCA-decomposed sparse signals (RPCA-S) effectively filtered background EEG activity contributing to inter-day variability.
• RPCA-S predominantly captured consistent EEG oscillations related to emotional responses.
• Cross-day binary emotion classification accuracy improved from 58.31% to 64.03% using RPCA-S, with increased informative features.

Conclusions:

• The proposed RPCA-based filtering strategy effectively mitigates inter-day EEG variability.
• This approach significantly enhances cross-day emotion classification accuracy compared to original EEG features.
• The findings support the development of realistic emotion-classification frameworks that address day-to-day variability.