Discriminative possibilistic clustering promoting cross-domain emotion recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Affective Brain-Computer Interface (aBCI) systems aim to improve individual prediction accuracy using multi-subject data.
• Subject-specific EEG feature patterns and limited labeled data pose significant challenges for aBCI systems.
• Domain Adaptation (DA) is a key approach for EEG-based emotion recognition, addressing distribution mismatch between datasets.

Purpose of the Study:

• To develop a robust domain adaptation method for EEG emotion recognition that overcomes limitations of existing Maximum Mean Discrepancy (MMD) based approaches.
• To enhance the accuracy and generalization performance of aBCI systems in the presence of noisy data and inter-subject variability.
• To introduce a novel distributed distance measure, Discriminative Possibilistic Clustering (DPC), for improved domain distribution alignment.

Main Methods:

• Proposed a Discriminative Possibilistic Clustering (DPC) criterion for distributed distance measurement, focusing on shared subspace identification and fuzzy entropy regularization.
• Developed an Emotion recognition based on DPC (EDPC) domain adaptation method incorporating a graph Laplacian matrix to maintain geometric structure and enhance label propagation.
• Utilized EEG datasets (SEED and SEED-IV) for comparative experiments against established domain adaptation learning methods.

Main Results:

• The proposed DPC metric is theoretically shown to be an upper bound of the MMD metric, allowing effective optimization.
• The EDPC method demonstrated improved robustness against noisy samples and enhanced generalization performance in EEG emotion recognition.
• Comparative experiments indicated that EDPC achieved better or comparable consistent generalization performance across multiple datasets.

Conclusions:

• The novel DPC criterion effectively addresses distribution mismatch and noise in EEG data for domain adaptation.
• The EDPC method offers a significant advancement in EEG-based emotion recognition, improving aBCI system performance.
• The findings suggest that EDPC is a promising approach for robust and accurate affective brain-computer interfaces.