<p>Non-line-of-sight (NLOS) signals caused by building obstruction severely degrade Global Navigation Satellite System (GNSS) positioning accuracy. Existing deep learning-based recognition methods rely heavily on large labeled datasets, yet NLOS signal labeling is labor-intensive and costly. To address this, a semi-supervised graph contrastive learning (GCL)-based framework is proposed for GNSS NLOS recognition with minimal labeled samples. First, irregular GNSS measurements are modeled as a sky satellite graph (nodes: satellites; edges: spatial correlations), with signal observation features effectively integrated with satellite geometric information. Second, a hybrid loss function is adopted, which combines unsupervised contrastive learning on abundant unlabeled data with supervised guidance from limited labeled samples, reducing reliance on labeled data while ensuring discriminative feature learning. Third, an asymmetric graph transformer encoder with stochastic propagation depth and multi-head attention is developed, enhancing representation learning of complex environmental features while preserving the integrity of the graph topology. Evaluations on real-world datasets show the method achieves 93% recognition accuracy with only 15% labeled samples, outperforming state-of-the-art supervised methods by 13.5% (static) and 26.2% (dynamic). It also exhibits strong cross-domain generalization, maintaining over 80% accuracy in unseen urban canyons.</p>