Accurate spatiotemporal prediction of landslide displacement is crucial for early disaster warning and risk mitigation. However, existing deep learning-based models typically focus on either temporal or spatial features independently, without adequately capturing the interactions and dependencies between them, leading to reduced prediction accuracy. In this study, we propose a novel InSAR-based landslide displacement prediction method that integrates spatial and temporal features through a graph neural network (GNN) framework. Taking the Heifangtai landslide area in the Chinese Loess Plateau as a case study, we first utilize SBAS-InSAR techniques on multi-orbit Sentinel-1A data to derive surface deformation time series, and apply a three-dimensional decomposition method to extract east-west, vertical, and north-south displacement components. Four typical landslide regions in the Heifangtai area are selected as experimental scenarios, and their corresponding spatial nodes are defined through a grid-based approach to construct an adjacency matrix, forming a spatial graph. We then develop a hybrid model, GCN-MHSA-TCN (GMTC), which combines Graph Convolutional Networks (GCN), Multi-Head Self-Attention (MHSA), and Temporal Convolutional Networks (TCN) to jointly learn spatiotemporal dependencies for accurate landslide displacement forecasting. Experimental results demonstrate that the proposed method effectively captures both local and long-range spatiotemporal correlations, achieving improved prediction accuracy and robustness compared to existing approaches. Moreover, across the four typical landslide regions, the GMTC model achieves an average root mean square error (RMSE) of 0.13 mm, representing the best performance among all evaluated prediction models. The predicted displacement time series exhibit high consistency with observed values and demonstrate greater temporal stability. This model offers a promising tool for supporting landslide monitoring, hazard assessment, and emergency evacuation planning based on InSAR time-series data.