.png){kind=icon}
.png){kind=icon}
Return
Multi-granularity graph refinement via granular ball for graph classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.133713.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph classification aims to categorize entire graphs based on their topology and node properties. Previous methods have focused on learning graph structure from both the graph-level topology and node-level properties. However, existing pooling and coarsening strategies often suffer from information loss by collapsing nodes into clusters without preserving the intrinsic structural nuances of subdomains. To this end, we propose a novel Multi-Granularity Graph Refinement method (MGGR) that shifts the paradigm from simple graph compression to adaptive structural decomposition. Distinct from traditional methods, MGGR leverages Granular Ball Computing to recursively decompose the graph into subdomains that achieve an optimal balance between topological density and semantic homogeneity. Specifically, we propose a fine-grained segmentation mechanism to resolve the graph into meaningful granules. Furthermore, a Hierarchical Granular Graph Encoder is developed to explicitly capture both the intra-domain granular characteristics and the inter-domain structural dependencies, effectively leveraging the robust inductive bias inherent in the multi-granularity decomposition. We evaluate MGGR on 9 benchmark datasets containing over 169,306 graphs in total. Results show that MGGR notably outperforms current state-of-the-art methods. Moreover, the organic integration of multi-granularity decomposition and hierarchical encoding significantly enhances the model’s structural resilience, enabling MGGR to maintain stable performance even under 20% label noise. Code is available at https://anonymous.4open.science/r/MGGR/ .
Keywords:
Graph classification
Multi-granularity decomposition
Granular Ball Computing
Hierarchical Graph Encoder
Structural resilience
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W