LMSA-net: a lightweight multi-scale attention network for eeg-based emotion recognition

Electroencephalogram (EEG)-based emotion recognition holds great potential in affective computing, mental health assessment, and human-computer interaction. However, EEG signals are non-stationary, noisy, and composed of multiple frequency bands, making direct feature learning from raw data particularly challenging. While end-to-end models alleviate the need for manual feature engineering, advancing the performance frontier of lightweight architectures remains a crucial and complex challenge for practical deployment. To address these issues, we propose LMSA-Net (Lightweight Multi-Scale Attention Network), a lightweight, interpretable, and end-to-end model that directly learns spatio-temporal features from raw EEG signals. The architecture integrates learnable channel weighting for adaptive spatial encoding, multi-scale temporal separable convolution for rhythm-specific feature extraction, and Sim Attention Module for parameter-free saliency enhancement. Our proposed LMSA-Net is evaluated on three benchmark datasets, SEED, SEED-IV, and DEAP, under subject-dependent protocols. It achieves top performance on SEED (65.53% accuracy), competitive results on SEED-IV (48.52% accuracy), and strong performance in arousal classification on DEAP, demonstrating good generalization. Ablation studies confirm the critical role of each proposed module. Frequency analysis reveals that our multi-scale temporal kernels inherently specialize in distinct EEG rhythms, validating their neurophysiological alignment. Lightweight design is evidenced by minimal parameters (7.64K) and low latency, ideal for edge deployment. Interpretability analysis further shows the model's focus on emotion-related brain regions. LMSA-Net thus delivers an efficient, interpretable, and high-performing solution. The code is available athttps://github.com/rhr0411/LMSA-Net.git.