FreAML: A Frequency-Domain Adaptive Meta-Learning Framework for EEG-Based Emotion Recognition

Emotion recognition technology, especially methods based on electroencephalogram (EEG) signals, plays a critical role in revealing deep human emotions and enhancing human-computer interaction experiences. However, the high-frequency characteristics of EEG signals lead to significant individual differences, causing domain distribution shift issues. To address this, we propose a frequency-domain adaptive meta-learning probabilistic inference framework, named FreAML, leveraging the advantages of frequency-domain signals in preserving comprehensive views and learning global dependencies. Specifically, we first introduce an adaptive meta-task construction strategy based on a clustering-matching pattern. Then, through a dual-stream amortization network, we learn task-specific frequency-domain real and imaginary parameters from the support set. Finally, using these task-specific parameters, we achieve rapid generalization learning of the query set through a frequency-domain linear mapping. Extensive experiments on two public datasets validate the effectiveness and superiority of the proposed method.