graph unpooling

**Graph unpooling** is a **graph neural network operation that reconstructs higher-resolution graphs from pooled representations** — the inverse of pooling, used in graph autoencoders and generative models to upsample graph structures. **What Is Graph Unpooling?** - **Definition**: Reconstruct graph structure from compressed representation. - **Purpose**: Enable graph generation and reconstruction tasks. - **Inverse Of**: Graph pooling (which compresses graphs). - **Use Case**: Graph autoencoders, generative models, super-resolution. - **Challenge**: Recover both node features and edge connectivity. **Why Graph Unpooling Matters** - **Graph Generation**: Create new molecules, social networks, circuits. - **Reconstruction**: Graph autoencoders need unpooling for decoder. - **Super-Resolution**: Upsample coarse graphs to finer detail. - **Hierarchical Models**: Build multi-scale graph representations. **Unpooling Strategies** - **Index-Based**: Store pooling indices, use to place nodes. - **Learned Upsampling**: Neural network predicts new nodes/edges. - **Spectral Methods**: Reconstruct via graph Fourier transform. - **Generative**: Sample new structure from learned distribution. **Applications** Molecule generation, circuit design, network synthesis, 3D mesh reconstruction. Graph unpooling is **essential for graph generative models** — enabling reconstruction from compressed representations.