graph generation
**Graph Generation** is the task of learning to produce new, valid graphs that match the statistical properties and structural patterns of a training distribution of graphs, encompassing both the generation of graph topology (adjacency matrix) and node/edge features. Graph generation is critical for applications in drug discovery (generating novel molecular graphs), circuit design, social network simulation, and materials science where creating new valid structures with desired properties is the goal.
**Why Graph Generation Matters in AI/ML:**
Graph generation enables **de novo design of structured objects** (molecules, materials, networks) by learning the underlying distribution of valid graph structures, allowing AI systems to create novel entities with specified properties rather than merely screening existing candidates.
• **Autoregressive generation** — Models like GraphRNN generate graphs sequentially: one node at a time, deciding edges to previously generated nodes at each step using RNNs or Transformers; this naturally handles variable-sized graphs and ensures validity through sequential construction
• **One-shot generation** — VAE-based methods (GraphVAE, CGVAE) generate the entire adjacency matrix and node features simultaneously from a latent vector; this is faster but requires matching generated graphs to training graphs (graph isomorphism) for loss computation
• **Flow-based generation** — GraphNVP and MoFlow use normalizing flows to learn invertible mappings between graph space and a simple latent distribution, enabling exact likelihood computation and efficient sampling of novel graphs
• **Diffusion-based generation** — DiGress and GDSS apply denoising diffusion models to graphs, progressively denoising random graphs into valid structures; these achieve state-of-the-art quality on molecular generation benchmarks
• **Validity constraints** — Chemical validity (valence rules, ring constraints), physical plausibility, and property targets must be enforced during or after generation; methods include masking invalid actions, reinforcement learning with validity rewards, and post-hoc filtering
| Method | Approach | Validity | Scalability | Quality |
|--------|----------|----------|-------------|---------|
| GraphRNN | Autoregressive (node-by-node) | Sequential constraints | O(N²) per graph | Good |
| GraphVAE | One-shot VAE | Post-hoc filtering | O(N²) generation | Moderate |
| MoFlow | Normalizing flow | Chemical constraints | O(N²) generation | Good |
| DiGress | Discrete diffusion | Learned from data | O(T·N²) | State-of-the-art |
| GDSS | Score-based diffusion | Learned from data | O(T·N²) | State-of-the-art |
| GraphAF | Autoregressive flow | Sequential construction | O(N²) | Good |
**Graph generation is the creative frontier of graph machine learning, enabling AI systems to design novel molecular structures, network topologies, and material configurations by learning the distribution of valid graphs and sampling new instances with desired properties, bridging generative modeling with combinatorial structure generation.**