Possibilistic distribution distance metric: a robust domain adaptation learning method.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Affective Brain-Computer Interface (aBCI) systems face challenges in cross-subject Electroencephalogram (EEG) pattern variability.
• Subject-specific classifiers lack sufficient labeled data, hindering performance.
• Domain Adaptation (DA) is crucial for EEG-based emotion recognition due to domain distribution inconsistencies.

Purpose of the Study:

• To develop a robust domain adaptation learning method for EEG-based emotion recognition.
• To address the limitations of existing Maximum Mean Discrepancy (MMD) methods in handling noisy EEG data.
• To improve the accuracy and robustness of aBCI systems.

Main Methods:

• Proposed a novel possibilistic distribution distance measure (P-DDM) by transforming the MMD criterion into a possibilistic clustering model to mitigate noise influence.
• Introduced a fuzzy entropy regularization term to enhance domain distribution alignment.
• Developed a robust domain adaptation classifier based on P-DDM (C-PDDM) using a Laplacian matrix for geometric consistency and maximizing source domain discriminative information.

Main Results:

• The proposed P-DDM criterion is theoretically proven to be an upper bound of the traditional MMD criterion under specific conditions.
• Experiments on SEED and SEED-IV EEG datasets demonstrated superior or comparable robustness performance, with an approximate 10% improvement in most cases.
• The C-PDDM classifier showed enhanced label propagation and generalization performance.

Conclusions:

• The novel P-DDM method offers a robust solution for domain adaptation in EEG-based emotion recognition, effectively handling noisy data.
• The proposed C-PDDM classifier improves the generalization and accuracy of aBCI systems.
• This research contributes to more reliable and effective brain-computer interfaces for emotion recognition.