Multi-Model Adaptation Learning With Possibilistic Clustering Assumption for EEG-Based Emotion Recognition.

Area of Science:

• Machine Learning
• Signal Processing
• Affective Computing

Background:

• Graph-based semi-supervised learning (GSSL) is effective but sensitive to noise and non-IID data.
• Electroencephalogram (EEG) data often exhibits non-IID characteristics and is susceptible to noise/outliers.
• Existing GSSL methods struggle with the inherent variability and noise in EEG signals.

Purpose of the Study:

• To propose a novel clustering method, MA-PCA, for robust EEG-based emotion recognition.
• To address the limitations of GSSL methods concerning noise, outliers, and non-IID data in EEG.
• To improve the generalization and robustness of emotion recognition systems using EEG data.

Main Methods:

• Developed a multi-model adaptation learning with possibilistic clustering assumption (MA-PCA) framework.
• Utilized fuzzy entropy regularization to mitigate the impact of noise and outliers.
• Incorporated multi-model adaptation to handle both IID and non-IID data distributions.
• Implemented the algorithm with a convergence theorem for theoretical validation.

Main Results:

• MA-PCA demonstrated superior or comparable robustness and generalization performance on DEAP and SEED datasets.
• The method effectively minimized the influence of noise/outlier samples in EEG data.
• Multi-model adaptation learning improved performance by considering both IID and non-IID data scenarios.

Conclusions:

• MA-PCA offers a robust and effective solution for EEG-based emotion recognition.
• The proposed method overcomes key limitations of traditional GSSL approaches in noisy, non-IID EEG data.
• This work advances the field of affective computing through improved signal processing techniques for EEG analysis.