Research on dynamic domain adaptation selective ensemble algorithm for cross-subject EEG emotion classification

Electroencephalogram (EEG) has become an important medium for emotion recognition research due to its high temporal resolution and noninvasive nature. However, significant differences among subjects in skull structure, EEG physiological characteristics, and emotional perception lead to strong heterogeneity in cross-subject EEG signals, severely limiting the generalization ability of emotion recognition models. To address this issue, this study proposes a cross-subject emotion recognition method. This method innovatively employs dynamic classifier selection to solve the cross-subject recognition problem. By constructing a pool of base classifiers and adaptively matching the optimal classifier for different samples, it mitigates the interference caused by inter-subject heterogeneity. Additionally, a neighborhood optimization strategy based on dynamic domain adaptation is proposed. This strategy maps samples from the test subject and the validation set into a common subspace to reduce distribution differences among samples. Subsequently, classifiers are dynamically selected from the base classifier pool for prediction based on this neighborhood. The effectiveness of the proposed method was evaluated on the public datasets FACED, SEED, and SEED-IV. Experimental results show that this method outperforms some of the current state-of-the-art cross-subject EEG emotion recognition algorithms, verifying its effectiveness and feasibility in emotion recognition tasks.