Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Neural Network-based Spectral Filtering Mechanism for Imbalance Classification in Network Digital Twin

Abubakar Isah, Ibrahim Aliyu, Sulaiman Muhammad Rashid, Jaehyung Park, Minsoo Hahn, Jinsul Kim

TL;DR

This work tackles multiclass failure classification on graphs derived from 5G network digital twins, where class imbalance severely biases standard GNNs toward majority classes. It introduces CF-GNN, a class-aware spectral filtering framework that employs per-class spectral transforms and eigenvector-aware attention to emphasize minority-class discriminative features, coupled with a twin graph Fourier transform (twin-GFT) for product-graph representations. The approach integrates class-conditioned filtering with node-level attention, yielding CF-GNN variants that consistently outperform strong baselines (GCN, GAT, GraphSMOTE, Reweight, GATE-GNN) across two real-world and digital twin datasets, and remain robust under varying imbalance ratios. Experimental results show substantial improvements in geometric mean, MCC, and macro-F1 scores, validating the practical relevance for proactive network failure detection in 5G digital twins. The findings highlight the potential of spectral, class-aware GNNs to balance representation learning in imbalanced graphs and guide future refinements in subclass-structure handling and data-level augmentation.

Abstract

Graph neural networks are gaining attention in fifth-generation (5G) core network digital twins, which are data-driven complex systems with numerous components. Analyzing these data can be challenging due to rare failure types, leading to imbalanced classification in multiclass settings. Digital twins of 5G networks increasingly employ graph classification as the main method for identifying failure types. However, the skewed distribution of failure occurrences is a significant class-imbalance problem that prevents practical graph data mining. Previous studies have not sufficiently addressed this complex problem. This paper, proposes class-Fourier GNN (CF-GNN) that introduces a class-oriented spectral filtering mechanism to ensure precise classification by estimating a unique spectral filter for each class. This work employs eigenvalue and eigenvector spectral filtering to capture and adapt to variations in minority classes, ensuring accurate class-specific feature discrimination, and adept at graph representation learning for complex local structures among neighbors in an end-to-end setting. The extensive experiments demonstrate that the proposed CF-GNN could help create new techniques for enhancing classifiers and investigate the characteristics of the multiclass imbalanced data in a network digital twin system.

Graph Neural Network-based Spectral Filtering Mechanism for Imbalance Classification in Network Digital Twin

TL;DR

This work tackles multiclass failure classification on graphs derived from 5G network digital twins, where class imbalance severely biases standard GNNs toward majority classes. It introduces CF-GNN, a class-aware spectral filtering framework that employs per-class spectral transforms and eigenvector-aware attention to emphasize minority-class discriminative features, coupled with a twin graph Fourier transform (twin-GFT) for product-graph representations. The approach integrates class-conditioned filtering with node-level attention, yielding CF-GNN variants that consistently outperform strong baselines (GCN, GAT, GraphSMOTE, Reweight, GATE-GNN) across two real-world and digital twin datasets, and remain robust under varying imbalance ratios. Experimental results show substantial improvements in geometric mean, MCC, and macro-F1 scores, validating the practical relevance for proactive network failure detection in 5G digital twins. The findings highlight the potential of spectral, class-aware GNNs to balance representation learning in imbalanced graphs and guide future refinements in subclass-structure handling and data-level augmentation.

Abstract

Graph neural networks are gaining attention in fifth-generation (5G) core network digital twins, which are data-driven complex systems with numerous components. Analyzing these data can be challenging due to rare failure types, leading to imbalanced classification in multiclass settings. Digital twins of 5G networks increasingly employ graph classification as the main method for identifying failure types. However, the skewed distribution of failure occurrences is a significant class-imbalance problem that prevents practical graph data mining. Previous studies have not sufficiently addressed this complex problem. This paper, proposes class-Fourier GNN (CF-GNN) that introduces a class-oriented spectral filtering mechanism to ensure precise classification by estimating a unique spectral filter for each class. This work employs eigenvalue and eigenvector spectral filtering to capture and adapt to variations in minority classes, ensuring accurate class-specific feature discrimination, and adept at graph representation learning for complex local structures among neighbors in an end-to-end setting. The extensive experiments demonstrate that the proposed CF-GNN could help create new techniques for enhancing classifiers and investigate the characteristics of the multiclass imbalanced data in a network digital twin system.

Paper Structure

This paper contains 27 sections, 42 equations, 10 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Learning-based failure prediction
  4. Imbalance learning technique
  5. Paper Structure
  6. Notations and Preliminaries
  7. Graph neural network
  8. Graph Fourier transform
  9. Digital twin spectral filtering
  10. Proposed Class-Fourier Graph Neural Network
  11. Class-Fourier transform filtering
  12. Proof of CF-GNNv
  13. Eigenvector-aware attention
  14. Experimental settings
  15. Description of the KDDI datasets
  16. ...and 12 more sections

Figures (10)

  • Figure 1: Class distribution of the failure types in real-world network data (Domain C).
  • Figure 2: Class distribution of the failure types in network digital twins (Domain A).
  • Figure 3: The components of fifth-generation core network digital twins failure injection settings.
  • Figure 4: The architectural framework of the proposed technique for imbalanced classification of the fifth-generation network of a digital twin
  • Figure 5: Classification mean accuracy (cmA) comparing network digital twin dataset model performance
  • ...and 5 more figures