Enhancing EEG-Based Emotion Recognition Using Sparse Dynamic Graph CNN With ℓ 2,1 -Norm

Electroencephalogram (EEG) sensors have become increasingly valuable for emotion recognition due to their ability to capture brain activity with high temporal resolution. However, conventional EEG-based models often suffer from high-dimensional noise and limited interpretability, particularly in understanding the contributions of different EEG channels. To address these challenges, we propose a novel graph-based deep learning framework named ℓ_2,1-norm-based Sparse Dynamic Graph Convolutional Neural Network (DGCNN). This framework incorporates structured sparsity into graph convolutional neural networks by applying ℓ_2,1-norm regularization to the adjacency matrix of the graph. This enforces structured, channel-wise sparsity in the adjacency matrix, enabling the model to emphasize meaningful neural connections associated with informative channels while suppressing noise and connections from irrelevant channels. A forward-backward splitting optimization strategy is employed to handle the non-differentiability of the regularization term. Experimental results on both the SEED and SEED-IV datasets demonstrate that the proposed method achieves consistently superior or comparable performance to existing CNN- and attention-based graph models, in both subject-dependent and subject-independent emotion classification tasks. Ablation studies confirm the effectiveness of the ℓ_2,1-norm by comparing it to ℓ₁-norm and unregularized variants, while a sparsity analysis highlights the model’s ability to structurally prune less informative EEG channels. These findings suggest that the proposed framework enhances not only classification performance but also model interpretability and efficiency, offering a promising direction for EEG-based affective computing in wearable and human-centered sensor environments.