<p>Electroencephalography (EEG) plays a crucial role in emotion recognition, however, its real-world deployment is hindered by generalization limitations due to subject-specific variability and distortions in EEG signals. While self-supervised contrastive learning offers the capacity to leverage unlabeled EEG data, existing methods critically depend on manually crafted negative sample pairs, introducing semantic ambiguity and high design costs. To overcome these challenges, we propose EmoDNCL+, the first negative-sample-free contrastive learning framework designed for EEG emotion recognition, completely eliminating the reliance on negative sample pairs. Dual-stream Negative-sample-free Contrastive Learning (DNCL) serves as the core of EmoDNCL+. It constructs local pairs from the student network’s outputs by applying strong and weak augmentations. Together with the student–teacher global pairs, the local pairs form a dual-stream optimization mechanism, achieving feature alignment at both local and global levels. Furthermore, to guarantee the reliability of positive samples, we propose Neighborhood Information Node Adjustment (NINA), grounded in the spatial continuity of EEG signals and accounting for subject-specific variability and distortions. Moreover, we dynamically construct adjacency matrices using cosine similarity to improve the capability of Graph Neural Networks in extracting cross-subject emotional features. Cross-subject emotion recognition experiments on SEED and SEED-IV, conducted under a leave-one-subject-out protocol, achieve state-of-the-art performance with accuracy rates of 93.60% and 79.32%, respectively. With half the training data, accuracy remains high at 91.88% on SEED and 74.87% on SEED-IV.</p>