Graph pooling aims to compress node embeddings and their underlying topology into a compact representation. Existing approaches mainly concentrate on fine‑grained node representations, progressively coarsening the graph by removing nodes or merging them into clusters, which overlooks the global‑to‑local hierarchy and adaptive granularity of graph topology. In practice, the whole graph can be regarded as the coarsest granularity that captures global structure, while finer local structures emerge from top to bottom until each sub‑domain reaches its appropriate granularity.
To address this, we propose Topology-Preserving Adaptive Graph Pooling (TPAGP). TPAGP dynamically partitions a graph into granular balls by jointly leveraging node features and topological cues. Each granular ball corresponds to a specific granularity, enabling the generation of multi‑granularity representations that preserve both local details and global patterns.
On top of this partitioning, we build a multi‑granularity graph network. The network facilitates feature interaction and optimization across different granularities, allowing information to flow freely between global and local levels and thereby enhancing graph classification performance.
Extensive experiments on several benchmark datasets demonstrate that TPAGP consistently outperforms existing pooling methods, especially mitigating information loss caused by fixed‑granularity strategies.
Review: TPAGP unifies global and local graph information through adaptive granular-ball partitioning, offering a fresh perspective on graph pooling that merits further investigation on more complex graph tasks.