Hybrid self-supervised EEG emotion representation: masked reconstruction joint with mutual information bounds

EEG signals are widely used in affective computing and brain informatics for emotion recognition due to their non-invasiveness. Deep learning and self-supervised learning (SSL) are key for EEG representation learning: masked reconstruction captures structural information, while contrastive learning extracts invariant semantic features via maximizing mutual information (MI). However, existing methods lack MI upper-bound constraints, leading to difficulty balancing structural fidelity and semantic discriminability. To address this issue, we propose HybridMI-EEG, a hybrid self-supervised framework for EEG emotion recognition. Built on a Transformer encoder-decoder, it integrates masked reconstruction and MI bound optimization into a three-component synergistic architecture. Specifically, raw EEG signals are decomposed into patch embeddings; a random continuous masking strategy on the embeddings drives reconstruction via contextual information. A dual-objective loss maximizes inter-view consistency and minimizes input-latent redundancy, while learnable augmenters generate informative contrastive views to enable synergistic optimization of local structure and global semantics. Experiments on DEAP and DREAMER datasets show that HybridMI-EEG outperforms state-of-the-art baselines in accuracy and stability, providing an effective solution to enhance EEG representation learning and emotion classification performance.