Improve the generalization of emotional classifiers across time by using training samples from different days

Area of Science:

• Neuroscience
• Computer Science
• Human-Computer Interaction

Background:

• Electroencephalographic (EEG)-based emotion recognition is crucial for human-computer interaction (HCI).
• A significant challenge is generalizing emotion models across time due to day-to-day variations in brain activity.
• Existing models struggle to maintain accuracy as emotional states can manifest differently over time.

Purpose of the Study:

• To investigate the feasibility of enhancing emotion classifier generalization by including multi-day EEG data in the training set.
• To determine if incorporating data from different days improves the accuracy of classifying neutral, positive, and negative emotional states.
• To analyze feature selection stability and relevance across different days.

Main Methods:

• Eight subjects watched movie clips to elicit neutral, positive, or negative emotions over five daily sessions.
• Electroencephalographic (EEG) signals were recorded during these sessions.
• A Support Vector Machine (SVM) classifier was trained using data from 1, 2, 3, or 4 days, with testing on remaining days.

Main Results:

• Average classification accuracies increased with the number of training days: 64.9% (1-day), 68.7% (2-day), 70.9% (3-day), and 73.0% (4-day).
• The 4-day training condition showed an approximate 10% improvement over the 1-day condition, with a peak accuracy of 81.2%.
• Feature analysis indicated stable distributions and selection of emotion-relevant, time-invariant features.

Conclusions:

• Including EEG data from multiple days in the training set significantly improves the generalization of emotion classifiers.
• This multi-day training approach effectively addresses the challenge of temporal variability in brain activity for emotion recognition.
• The findings suggest a practical method for developing more robust and reliable EEG-based emotion recognition systems for HCI applications.