.png){kind=icon}
.png){kind=icon}
Return
Hierarchical prototype-guided representation learning for robust graph classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ins.2025.122777.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
A robust graph classification model is critical for real-world applications. However, models based on off-the-shelf graph neural networks (GNNs) are susceptible to noise interference and malicious attacks, owing to the cascading effects of disruptions within the graph structure. Despite recent advances in consistency regularization methods for robust graph representation learning, they overlook the fine-grained multiscale semantic consistency necessary for capturing the cascading structural characteristics of graphs. In this work, we propose HPGRL, a hierarchical prototype-guided representation learning method for robust graph classification. Firstly, to align with the information cascading process, we develop a hierarchical structure prototype-guided contrastive learning regularizer, which strengthens representation consistency across different levels, including nodes, subgraph structures, and the global graph structure. Secondly, we introduce a low-dimensional projector, along with a corresponding Bayesian class prototype-guided projector regularizer to account for uncertainty and variability of class decision boundaries. Extensive experiments on benchmark datasets demonstrate the effectiveness of our data-and class-prototype-guided framework, with HPGRL achieving a 4.17 % accuracy gain over the SOTA method MGRL on IMDB-MULTI.
Keywords:
Robust graph classification
Prototype learning
Bayesian prototype
Journal
IF:
6.8
Papers:
540
Citations:
6.2W