[Paper Review] NetGAN: Generating Graphs via Random Walks

Aleksandar Bojchevski, Oleksandr Shchur+2 more|arXiv (Cornell University)|Mar 2, 2018

Advanced Graph Neural NetworksComputer Science33 references151 citations

TL;DR

NetGAN is an implicit generative model that learns the distribution of biased random walks over a graph using a Wasserstein GAN, producing graphs that mimic real networks and generalize to link prediction.

ABSTRACT

We propose NetGAN - the first implicit generative model for graphs able to mimic real-world networks. We pose the problem of graph generation as learning the distribution of biased random walks over the input graph. The proposed model is based on a stochastic neural network that generates discrete output samples and is trained using the Wasserstein GAN objective. NetGAN is able to produce graphs that exhibit well-known network patterns without explicitly specifying them in the model definition. At the same time, our model exhibits strong generalization properties, as highlighted by its competitive link prediction performance, despite not being trained specifically for this task. Being the first approach to combine both of these desirable properties, NetGAN opens exciting avenues for further research.

Motivation & Objective

• Motivate and develop an implicit generative model for graphs that can learn from a single real-world graph.
• Capture essential topological properties (e.g., degree distribution, community structure) without explicit specification in the model.
• Demonstrate that generated graphs generalize beyond the input graph, notably via link prediction performance.

Proposed method

• Model the graph generation problem as learning the distribution of biased random walks over the input graph.
• Use a stochastic neural network generator to produce random walks and a discriminator to distinguish generated walks from real ones.
• Employ a bias-aware second-order random-walk sampling strategy and an LSTM-based generator with straight-through Gumbel for differentiable sampling.
• Train with the Wasserstein GAN objective and gradient penalty for stability.
• Assemble generated walks into a binary adjacency matrix via a probabilistic, edge-aware sampling procedure.

Experimental results

Research questions

• RQ1 Can an implicit GAN-based model generate graphs that reproduce real-world network properties without explicit property constraints?
• RQ2 Do graphs generated by NetGAN generalize beyond the training graph, as evidenced by link prediction performance?
• RQ3 How do hyperparameters (e.g., random-walk length, latent space) affect the realism and diversity of generated graphs?
• RQ4 Is the model scalable to large graphs by operating on random walks rather than full adjacency matrices?

Key findings

• NetGAN consistently reproduces many known network patterns without explicitly encoding them.
• The model demonstrates strong generalization, achieving competitive or state-of-the-art link prediction on several datasets.
• Longer random walks (T) improve performance over edge-only generation, with diminishing returns beyond a certain length.
• Latent space interpolation yields smooth transitions in graph properties, enabling graphs with varying characteristics.
• Sampling more random walks improves link-prediction performance, especially on larger graphs, and sampling is parallelizable.