Abstract As a core task in the field of affective computing, emotion recognition holds significant application value in areas such as medical diagnosis and human-computer interaction. EEG (Electroencephalography) has emerged as a key technology for emotion decoding, owing to its ability to directly reflect neural activity and its non-invasive nature. This paper proposes an explainable EEG emotion recognition method based on 4D attention, which achieves accurate emotion state classification and mechanism analysis by integrating frequency-domain decomposition, spatiotemporal feature extraction, and an attention mechanism. this study, the SincNet network is employed to perform frequency-domain decomposition on raw EEG signals, utilizing learnable band-pass filters to extract key frequency band information, including δ, θ, α, β, and γ bands. A spatiotemporal feature extraction module is constructed using multi-scale 1D convolutions and spatial convolution kernels to capture the temporal dynamics of signals and the spatial topological relationships of electrodes, respectively. A frequency-band attention mechanism is introduced, and a multi-head self-attention framework is used to model the intra-band local features and inter-band interactive relationships, thereby enhancing the flexibility and discriminative power of feature representation. Experiments were conducted on the DEAP dataset and SEED dataset. The results demonstrate that the proposed method exhibits competitive advantages in emotion classification and significantly outperforms comparative methods such as SVM (Support Vector Machine) and EEGNet. Furthermore, by analyzing the frequency response of the filters learned by SincNet, it is found that different frequency bands have strong correlations with negative emotions and positive emotions, which verifies the interpretability of the model. This study provides an efficient technical solution for EEG emotion recognition and contributes to promoting the application of brain-computer interfaces in the field of emotion analysis.
Liu et al. (Fri,) studied this question.