Embedded EEG Feature Selection for Multi-Dimension Emotion Recognition via Local and Global Label Relevance

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalogram (EEG) feature selection is crucial for emotion recognition due to limited samples and high dimensionality.
• Current methods often transform multi-dimension emotional labels to single-dimension, neglecting inter-dimensional relationships.

Purpose of the Study:

• To develop an efficient EEG feature selection method for multi-dimension emotion recognition.
• To address the limitations of existing methods by incorporating local and global label relevance.

Main Methods:

• Proposed an efficient EEG feature selection method for multi-dimension emotion recognition (EFSMDER).
• Utilized orthogonal regression to reduce EEG feature space dimensionality and capture local label correlations.
• Employed global label correlations in the original multi-dimension emotional label space to construct label information in the reduced space.

Main Results:

• EFSMDER achieved superior multi-dimension classification accuracies compared to fourteen state-of-the-art methods.
• Achieved accuracies of 86.43% (DREAMER), 84.80% (DEAP), and 97.86% (HDED).

Conclusions:

• The EFSMDER method effectively performs representational EEG feature subset selection by leveraging local and global relevance.
• The proposed method demonstrates significant improvements in multi-dimension emotion recognition using EEG data.