Graph neural networks pool graphs using adaptive cluster numbers
Graph-dependent pooling: adaptive granularity in hierarchical graph neural networks
Artificial IntelligenceMachine Learning
Summary
Graphs are a way to represent connected data, like social networks or molecules. To understand them better, computers often group nodes together in layers, but previous methods fixed the number of groups beforehand. The authors developed GDPool, which lets the computer decide the right number of groups dynamically for each graph. This approach leads to better predictions and more meaningful groupings, especially on tasks involving complex molecules and materials. Their method also ensures smooth behavior across different settings, making it reliable.
What this means in practice
- •For drug discovery teams: Generate multi-scale representations of molecular graphs to improve prediction accuracy of chemical properties.
- •For materials scientists: Model and analyze complex structures like metal-organic frameworks by adaptively coarsening graphs for better structural insights.
Authors
Bohan Yoon, Hyejin Jang
Abstract
Graph Neural Networks are widely used to learn structural representations of relational data. Hierarchical graph pooling learns multi-level representations by progressively coarsening graphs; however, prior methods rely on predefined cluster numbers or fixed pooling ratios, limiting adaptability to individual graphs. We propose Graph-Dependent Pooling (GDPool), which dynamically determines pooling granularity for each graph. We analyze optimization conflicts among layer-wise objectives through gradient alignment, showing that independently supervised pooling levels can send conflicting gradients to shared parameters. GDPool uses self-attention to select graph-specific anchors representing structurally important nodes and constructs coarse graphs through soft node assignment. Experiments on metal-organic framework and molecular graph benchmarks show that GDPool consistently achieves the best predictive performance across all benchmarks. Evaluations using modularity and the effective number of clusters show that GDPool achieves consistently higher modularity than existing methods while the number of active clusters decreases monotonically with depth. Sensitivity analyses confirm stable behavior across hyperparameter and architectural choices.