Pooling and Hierarchical Graph Networks
What you'll learn: How to reduce graph complexity through pooling operations to make graph-level predictions more effective.
The Core Problem
When you need to make predictions about an entire graph (like classifying molecules or predicting protein properties), you can't just use node embeddings—you need to summarize the whole structure into a single representation. Graph pooling does exactly this: it progressively coarsens a graph, creating hierarchical representations similar to how CNNs use pooling to downsample images.
Graph Coarsening Intuition
Think of graph pooling like creating a summary of a social network. Instead of tracking every individual, you group people into communities, then group communities into regions, building a hierarchy. At each level, you preserve the most important structural information while reducing complexity.
Common Pooling Strategies
Global pooling aggregates all node embeddings into one graph-level vector using operations like sum, mean, or max—simple but loses structural detail.
Hierarchical pooling creates multiple coarsening levels. Methods like DiffPool learn soft cluster assignments, grouping similar nodes together. TopK pooling selects the top-k most important nodes based on learned scores. SAGPool combines graph convolutions with top-k selection for structure-aware pooling.
Each pooling layer reduces graph size, and the resulting coarsened graph feeds into more message passing layers, creating a pyramid-like architecture.
Why It Matters
Hierarchical pooling captures both local patterns (fine details) and global structure (overall shape), crucial for tasks like molecular property prediction where both atom-level features and overall topology matter.
Key Takeaway: Graph pooling operations coarsen graphs hierarchically, enabling effective graph-level predictions by progressively summarizing structure from local neighborhoods to global representations—the graph equivalent of CNN pooling pyramids.