Causal discovery in telecommunication networks is challenging because alarms have irregular timing, uncertain propagation, and incomplete labeling. Existing methods often fail to ensure robustness, accuracy, and interpretability. We propose CausalGNN-Net, which integrates temporal modeling, network topology, and expert priors. A Transformer-based temporal embedding module captures timing with causal masking, a spatiotemporal graph constructor combines topology and co-occurrence with GNN message passing and adaptive edge dropout, a directional graph learner enforces acyclicity, and a prior-guided refiner aligns results with domain knowledge. Training with contrastive loss, sparsity, priors, and calibration improves stability and interpretability. CausalGNN-Net provides a unified and practical solution for causal discovery in telecom alarms.