Selecting transferrable neurophysiological features for inter-individual emotion recognition via a shared-subspace...

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Individual differences in emotions, personality, and motivation create unique neurophysiological data patterns.
• This variability poses a challenge for developing accurate emotion recognition systems using electroencephalogram (EEG) data.

Purpose of the Study:

• To propose and validate a Shared-Subspace Feature Elimination (SSFE) approach for identifying common EEG features across diverse individuals.
• To enhance the performance of emotion recognition systems by addressing inter-subject variability.

Main Methods:

• Developed the SSFE framework to create a low-dimensional subspace representing inter-emotion discrimination.
• Utilized a leave-one-subject-out validation strategy on the DEAP and MAHNOB-HCI public EEG databases.
• Compared SSFE performance against five other feature selection methods across six machine learning models.

Main Results:

• SSFE effectively identifies EEG variables with shared characteristics, reducing individual specificity.
• Achieved competitive binary classification accuracies for arousal (0.6521 DEAP, 0.6520 MAHNOB-HCI) and valence (0.6635 DEAP, 0.6537 MAHNOB-HCI).
• Evaluated the effectiveness and computational cost of SSFE across various machine learning models.

Conclusions:

• The SSFE approach offers a robust method for improving cross-subject emotion recognition from EEG data.
• SSFE demonstrates potential for building more generalizable and accurate emotion recognition systems.
• The method effectively handles inter-subject variability in neurophysiological signals.