Abstract
Graph Neural Networks (GNNs) have emerged as powerful tools for modeling digital twins of fifth-generation (5G) core networks, which consist of highly interconnected and data-driven components. However, their performance in multiclass failure diagnosis remains challenging due to the severe class imbalance caused by the rarity of specific failure events. To address this problem, we propose the Classical-Fourier Graph Neural Network (CF-GNN), a novel spectral learning framework that employs class-specific filters to enhance feature discrimination under imbalanced conditions. Unlike conventional approaches that employ a single global spectral filter g ( Λ ) , CF-GNN learns distinct spectral filters g c ( Λ ) for each class c , enabling the transformation X c ′ = U g c ( Λ ) U ⊤ X , which adapts to both eigenvalue and eigenvector-based structural characteristics. Experimental results on real-world 5G core network digital twin datasets show that CF-GNN achieves robust and adaptive performance across imbalance ratios from 0.1 to 0.9, reaching a peak accuracy of 98.02%. Compared with state-of-the-art baselines, CF-GNN delivers up to 17.3% , 8.0% , and 9.1% improvements in Geometric Mean (G-Mean), Mathews Correlation Coefficient (MCC), and Macro F1, respectively, over the best competing models. These results highlight the effectiveness of Fourier-aware, class-specific spectral learning for reliable and interpretable graph-based failure classification in a network digital twin.