Detecting network intrusions without labeled data remains challenging due to severe class imbalance, evolving traffic patterns, and computational complexity in realistic scenarios. To address these issues, we propose a fully unsupervised neuro-symbolic graph-learning pipeline that integrates symbolic reasoning into graph neural representations, enhancing interpretability and robustness. Our key contributions include a novel feature-selection strategy driven by unsupervised community graph detection, a memory-efficient line graph construction reduced via minimum-spanning trees, and a lightweight symbolic layer providing human-readable explanations of anomalies. Evaluations on IoTID20 and UNSW-NB15 benchmarks yield Matthews Correlation Coefficient (MCC) scores of 0.97 and 0.91, significantly surpassing recent unsupervised graph-based baselines demosntrating the practical viability and effectiveness of neuro-symbolic frameworks as an innovative and interpretable approach for unsupervised graph network intrusion detection.