FDDGNet: An information bottleneck-inspired feature disentanglement network for cross-subject EEG-based emotion recognition

Electroencephalography (EEG)-based emotion recognition has attracted increasing interest due to its non-invasive and objective nature in measuring brain activity. However, inter-subject variability in EEG signals remains a major challenge to developing generalizable emotion recognition models. In this paper, we propose a feature disentanglement domain generalization network (FDDGNet) to tackle the cross-subject emotion recognition problem by explicitly separating emotion-relevant features from subject-specific representations. Specifically, we first employ a Long Short-Term Memory (LSTM)-based encoder to extract the latent features from the raw EEG signals. Subsequently, an orthogonal complementary project is introduced to decouple the latent features into orthogonal subspaces: emotion-relevant and subject-specific feature space, thereby enhancing the model's capacity for emotion representation learning. Inspired by the concept of sufficiency in information theory, we propose a sufficient information bottleneck module that ensures the subject-specific feature space captures sufficient domain information while suppressing emotional leakage, thereby promoting clearer feature separation. Finally, the emotion-relevant and subject-specific features are concatenated and fed into a LSTM-based decoder to reconstruct the raw EEG signals, which helps preserve information integrity during feature extraction and disentanglement. To evaluate the effectiveness of the proposed FDDGNet, cross-subject and cross-dataset experiments are conducted on two public EEG-based emotion recognition datasets, AMIGOS and DREAMER. Experimental results demonstrate that FDDGNet outperforms several state-of-the-art methods and exhibits promising generalization performance on unseen domains, highlighting its potential for robust and practical emotion recognition applications. The code is available at https://github.com/yyk90/FDDGNet.