<p>Accurately predicting link signs in signed graphs remains challenging due to the need to capture detailed node-level interactions and dynamic higher-order group relationships. We propose a novel dual-channel graph convolutional network that combines dynamic hypergraph learning with spectral-spatial processing. Our approach dynamically constructs signed hypergraphs using adaptive sampling, initializing hyperedges through learnable Gaussian distributions and refining them via attention mechanisms to model collaborative and adversarial group interactions. A unique spectral decomposition strategy processes positive relationships with low-frequency filtering to preserve collaborative patterns, while applying high-frequency enhancement on negative connections to amplify adversarial signals. The architecture incorporates self-gating mechanisms that automatically balance these spectral components, ensuring stable integration of multi-scale features. Extensive experiments on three real-world social networks demonstrate DSHGNN’s superior performance over state-of-the-art methods, particularly in handling class imbalance and complex group dynamics. Theoretical analysis confirms the framework’s computational efficiency with linear complexity, enabling scalability to large networks. Ablation studies confirm the critical role of dynamic hypergraph construction, which delivers the most substantial performance gains (improving F1-MA by 5.73% and AUC-L by 5.66% over the baseline) through adaptive modeling of higher-order group interactions By bridging hypergraph dynamics with spectral-spatial feature fusion, this work advances signed graph analysis and offers new insights for modeling complex relational patterns in social networks.</p>