EEG-based emotional valence and emotion regulation classification: a data-centric and explainable approach

Area of Science:

• Artificial Intelligence
• Neuroscience
• Machine Learning

Background:

• Emotion classification using electroencephalographic (EEG) data presents significant challenges.
• Existing research often prioritizes model selection over dataset methodological optimization.
• Label noise in EEG datasets can impede accurate emotion recognition.

Purpose of the Study:

• To investigate the impact of different methodological approaches on emotion classification performance using EEG data.
• To explore the effectiveness of curriculum learning and confident learning in handling noisy EEG datasets.
• To assess the explainability of machine learning models using the Integrated Gradient technique.

Main Methods:

• Applied curriculum learning and confident learning strategies to address label noise in EEG datasets.
• Utilized a multilabel classification approach to identify emotional valence and emotion regulation strategies.
• Employed the Integrated Gradient technique for model explainability analysis.
• Compared model performance across datasets curated based on varying levels of label noise.

Main Results:

• Curriculum learning and confident learning demonstrated varying impacts on emotion classification accuracy depending on the dataset and model architecture.
• The choice of model architecture influenced feature importance patterns, revealing distinct advantages and limitations.
• Explainability analysis provided insights into how different methods affect model interpretability.

Conclusions:

• Methodological choices, particularly those addressing label noise like curriculum and confident learning, are crucial for optimizing EEG-based emotion classification.
• Model architecture significantly influences both performance and the interpretability of emotion recognition systems.
• Further research is needed to refine these methods for robust and explainable AI in affective computing.