Emotion recognition is increasingly important for applications in mental health and personalized marketing. Traditional methods based on facial and vocal cues lack robustness due to voluntary control, motivating the use of EEG signals that capture neural dynamics with high temporal resolution. Existing EEG-based approaches using CNNs and LSTMs have improved spatial and temporal feature extraction; however, they still face critical limitations. These models struggle to represent electrode connectivity and adapt to inter-individual variability, and their architectures are typically handcrafted, requiring extensive manual tuning of hyperparameters and structural design. Such constraints hinder scalability and personalization, highlighting the need for automated architecture optimization. To address these challenges, we propose a dual-pipeline architecture that integrates frequency-domain and time-domain EEG features. The frequency-domain branch employs a Graph Convolutional Network (GCN) to model spatial relationships among electrodes, while the time-domain branch uses LSTM enhanced with Channel Attention to emphasize subject-specific informative channels. Furthermore, we introduce Differentiable Architecture Search (DARTS) to automatically discover optimal architectures tailored to individual EEG patterns, significantly reducing search cost compared to manual tuning. Experimental results demonstrate that our framework achieves competitive accuracy and high adaptability compared to state-of-the-art baselines, marking the first integration of GCN, LSTM, channel attention, and architecture search for EEG-based emotion recognition.
Tsuji et al. (Thu,) studied this question.