EEG signals face challenges such as a scarcity of training data, significant individual variability, and difficulties in designing neural network architectures, which severely limit the performance and generalization capabilities of emotion recognition models. This paper proposes an EEG-based emotion recognition method that integrates masked generative adversarial data augmentation with personalized neural architecture search. We designed a data augmentation strategy using a frequency-domain feature-based stochastic electrode mask generative adversarial network, which effectively expands the training dataset and enhances the model’s generalization ability. Subsequently, we constructed a neural architecture search space tailored for EEG signals and employed a differentiable architecture search method to automatically discover the optimal network topology. Finally, a personalized architecture search strategy is employed to independently identify the classification network best suited to each subject’s unique EEG signal characteristics. Experimental validation on the public EEG emotion dataset DEAP achieved a classification accuracy of 98.64% on the Arousal dimension and 98.59% on the Valence dimension, significantly outperforming traditional machine learning methods and deep learning models with fixed architectures.