Validation of Model-Basis Transfer Learning for a Personalized Electroencephalogram-Based Emotion-Classification...

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Affective brain-computer interfaces (aBCIs) using electroencephalogram (EEG) signals are gaining traction.
• Variability in emotional EEG data challenges the robustness of pre-trained machine learning models.
• Limited personal data hinders the development of effective personalized affective models.

Purpose of the Study:

• To propose and validate a model-basis transfer learning (TL) approach for personalized affective modeling.
• To construct personalized aBCI models using reduced emotion-annotated data.
• To address the challenge of inherent variability in EEG signals for real-life aBCI applications.

Main Methods:

• A longitudinal eight-day dataset from 10 subjects was utilized.
• A model-basis transfer learning (TL) approach was developed and tested.
• Daily reliability testing was performed to evaluate model performance.

Main Results:

• The proposed TL approach demonstrated superior performance compared to subject-dependent models.
• The TL approach achieved approximately 6% improvement in binary valence classification.
• Recycling a compact set of eight transferable models from other subjects enhanced personalization.

Conclusions:

• The model-basis TL approach effectively overcomes the limitations of inherent variability in EEG signals.
• This method facilitates the construction of personalized affective models with minimal data.
• The findings support the practical application of TL in realistic aBCI systems.