.png){kind=icon}
.png){kind=icon}
Return
A multi-view graph neural network with subgraph variational autoencoder for class-Imbalanced node classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.115081.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Class imbalance in graph-structured data critically compromises the generalization ability of Graph Neural Networks (GNNs), leading to classification bias toward majority classes. Existing solutions predominantly rely on node generation techniques based on Synthetic Minority Over-sampling Technique (SMOTE) and Mixup to mitigate imbalance. However, these methods suffer from three fundamental limitations: (1) inefficient single-node synthesis, (2) compromised validity due to the fact that high neighbor similarity does not ensure class consistency, and (3) disruption of graph topology caused by the introduction of numerous low-validity nodes. To address these issues, a multi-view GNN framework integrating a Subgraph Variational Autoencoder (SVAE), termed GraphMV-SVAE, is proposed in this study. The key innovation lies in a holistic approach that synergizes uncertainty-aware neighbor filtering with batch-wise minority subgraph generation, fundamentally overcoming the limitations of prior node-level interpolation methods. First, an Evidence Deep Learning (EDL)-based uncertainty classifier is employed to filter reliable neighbors, thereby preventing invalid connections. Then, the SVAE is introduced to enable batch generation of discriminative minority nodes, effectively overcoming the inefficiency of prior interpolation-based methods that synthesize nodes individually. Moreover, unlike conventional Variational Autoencoders that generate independent and identically distributed samples, the proposed SVAE incorporates a novel quantum-inspired reparameterization strategy with class-mean constraints and non-class divergence, ensuring that the generated nodes are category-aware and maintain inter-class separability for enhanced feature discriminability. Semantic consistency between original and synthetic views is preserved through multi-view learning with a Feature Interaction Transformer. Comprehensive evaluations involving eleven baseline methods across six benchmark datasets demonstrate that the proposed framework achieves state-of-the-art performance, with notable improvements observed across various settings. For instance, on the Cora-LT dataset with a GCN backbone, the method achieves substantial absolute gains in all evaluation metrics. Superior results are consistently observed across all three GNN architectures (GCN, GAT, and GraphSAGE), with overall performance leading all compared methods. Additional experiments on large-scale graphs further confirm its robustness and scalability, achieving accuracy improvements of 2.68% and 1.43% on the obgn-arxiv and CoraFull datasets, respectively, while avoiding out-of-memory issues that affect several baseline approaches.
Keywords:
Node classification
Class imbalance
Graph neural network
Data augmentation
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W