Abstract::To address the challenge of network link quality prediction in highly dynamic environments, this paper proposes a state-clustering-guided causal spatio-temporal graph convolutional network architecture named Causal-Clustered STGCN (CC-STGCN). The work makes key breakthroughs in time-series state partitioning based on shape similarity, state-specific Granger causality graph construction, and spatio-temporal feature aggregation under causal constraints, thereby achieving adaptive perception of network operation modes and capturing latent dependencies beyond physical connectivity. The core idea is to partition continuous states into typical patterns via K-shape clustering and construct directed, weighted causal graphs within each state based on Granger causality tests. These causal graphs replace the traditional physical topology as the spatial prior for graph convolution, ensuring that feature aggregation strictly follows causal pathways. Experiments conducted on the SynthSoM dataset demonstrate that the proposed model achieves a 6.7% improvement in prediction accuracy over the strongest baseline in standard scenarios and maintains its advantage in complex scenarios.
