Electroencephalogram (EEG)-based emotion recognition faces significant generalization challenges in cross-subject and cross-session settings, primarily due to the inherent non-stationarity, individual differences, and semantic deficiency of EEG signals. To address these challenges and enhance the universality of affective brain-computer interface systems, this study proposes a novel unsupervised domain adaptation framework named Prototypical Graph-based Deep Label Propagation with Semantic Augmentation (PGDLP). PGDLP seamlessly integrates three core components—prototypical semantic augmentation, prototype-graph deep label propagation, and prototypical alignment—into an end-to-end optimized system. Specifically, class-wise multivariate normal distributions are constructed using source-domain feature statistics to augment target-domain features semantically, bridging the domain gap and mitigating semantic insufficiency. An adaptive similarity graph based on prototype-semantic distances is designed to optimize pseudo-label quality while reducing computational complexity. It is combined with linear projection and an exponential moving average (EMA) for dynamic refinement. Dual intra- and inter-domain alignment losses with an adaptive balance factor are introduced to enhance intra-class compactness and inter-domain transferability, thereby facilitating the learning of discriminative and domain-invariant features. Extensive experiments are conducted on three benchmark datasets (SEED, SEED-IV, DEAP) under four rigorous evaluation protocols (cross-subject cross-session, cross-subject single-session, within-subject cross-session, cross-database). Overall, PGDLP achieves superior recognition accuracy and generalization performance compared with most state-of-the-art methods across the majority of evaluation protocols. PGDLP also presents strong robustness against label noise, stable hyperparameter performance, and fast convergence. The results demonstrated that PGDLP outperforms state-of-the-art methods in recognition accuracy and generalization, with verified robustness to label noise, stable hyperparameters, and efficient convergence. This study provides a promising solution for unsupervised cross-domain EEG emotion recognition and offers valuable insights for domain adaptation research on other physiological signals.