Graph Neural Networks (GNNs) are extensively used to learn structural representations of relational data. Hierarchical graph pooling obtains multi‑level representations by progressively coarsening graphs, yet existing methods typically rely on predefined cluster counts or fixed pooling ratios, which hampers adaptability to individual graphs. To address this, we introduce Graph-Dependent Pooling (GDPool), which dynamically determines the pooling granularity for each graph.\
We first analyze optimization conflicts among layer‑wise objectives via gradient alignment, revealing that independently supervised pooling levels can send conflicting gradients to shared parameters. GDPool mitigates this issue by employing self‑attention to select graph‑specific anchors that capture structurally important nodes, and constructs coarse graphs through soft node assignment to these anchors.\
Extensive experiments on metal‑organic framework (MOF) and molecular graph benchmarks show that GDPool consistently achieves the best predictive performance across all datasets. Evaluations using modularity and the effective number of clusters indicate that GDPool attains higher modularity than prior methods, while the number of active clusters decreases monotonically with depth. Sensitivity analyses confirm that GDPool remains stable across various hyper‑parameter and architectural choices.\
Review: By introducing adaptive anchor selection and soft assignment, GDPool effectively overcomes the rigidity of fixed‑granularity pooling in hierarchical GNNs, offering a more flexible and higher‑performing approach to graph representation learning.