[Paper Review] Labeled Graph Generative Adversarial Networks

Shuangfei Fan, Bert Huang|arXiv (Cornell University)|Jun 7, 2019

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology28 references19 citations

TL;DR

This paper proposes Labeled-Graph Generative Adversarial Networks (LGGAN), a deep generative model that synthesizes realistic, diverse labeled graphs by training a generator to produce node labels and adjacency matrices, while a graph convolutional discriminator with residual connections distinguishes real from generated graphs. LGGAN outperforms existing methods in structural fidelity, generalization across graph sizes and domains, and diversity, while preserving key distributional properties of training data.

ABSTRACT

As a new approach to train generative models, \emph{generative adversarial networks} (GANs) have achieved considerable success in image generation. This framework has also recently been applied to data with graph structures. We propose labeled-graph generative adversarial networks (LGGAN) to train deep generative models for graph-structured data with node labels. We test the approach on various types of graph datasets, such as collections of citation networks and protein graphs. Experiment results show that our model can generate diverse labeled graphs that match the structural characteristics of the training data and outperforms all alternative approaches in quality and generality. To further evaluate the quality of the generated graphs, we use them on a downstream task of graph classification, and the results show that LGGAN can faithfully capture the important aspects of the graph structure.

Motivation & Objective

• To develop a deep generative model capable of synthesizing realistic labeled graph-structured data with both node labels and complex topologies.
• To address the limitations of existing graph GANs that cannot handle node labels or generalize across diverse graph types and sizes.
• To enable privacy-preserving graph data sharing by generating synthetic graphs that mimic real-world structures without exposing sensitive information.
• To evaluate the quality of generated graphs using downstream graph classification tasks and structural metrics.
• To ensure the model generates diverse graphs that are not simple copies of training examples, balancing novelty and realism.

Proposed method

• The generator network learns to map from a latent noise vector to a complete labeled graph, producing both an adjacency matrix and node labels.
• The discriminator uses a graph convolutional network (GCN) with residual connections to extract adaptive, structure-aware high-level features for distinguishing real from generated graphs.
• The GAN framework trains the generator and discriminator in an adversarial manner, where the generator improves at fooling the discriminator, and the discriminator improves at detecting fakes.
• The model employs a Weisfeiler-Lehman kernel-based distance metric to quantitatively evaluate diversity between generated graphs and training data.
• Graph structure and label generation are jointly optimized during training to ensure consistency between node attributes and topological structure.
• Evaluation includes both intrinsic metrics (MMD, kernel distances) and downstream graph classification performance on synthetic data.

Experimental results

Research questions

• RQ1 Can a GAN-based model effectively generate labeled graphs that preserve the structural and distributional properties of real-world graph datasets?
• RQ2 How does LGGAN compare to traditional and deep generative models in terms of structural fidelity and diversity across different graph types and sizes?
• RQ3 To what extent can LGGAN generalize to unseen graph domains, such as citation networks and protein interaction graphs?
• RQ4 Does the inclusion of node labels improve the realism and quality of generated graphs compared to models that ignore labels?
• RQ5 How well do the generated graphs perform on downstream tasks like graph classification, indicating faithful capture of underlying data distribution?

Key findings

• LGGAN generates labeled graphs that match the structural characteristics of training data, including degree distribution, clustering, and community structure, across diverse datasets like Cora, CiteSeer, PROTEINS, and ENZYMES.
• On the Cora-small and Cora datasets, LGGAN maintains low Maximum Mean Discrepancy (MMD) scores with minimal increase in larger graphs, indicating strong scalability and consistency.
• LGGAN outperforms all baselines—including MolGAN, DeepGMG, GraphRNN, and traditional models like E-R and B-A—in both quality and generalization, especially on large and diverse graph types.
• The model achieves high diversity, as shown by Weisfeiler-Lehman kernel distances: generated graphs are neither exact copies of training graphs nor overly dissimilar, indicating balanced novelty and realism.
• When used for downstream graph classification, graphs generated by LGGAN achieve competitive performance, demonstrating that the model faithfully captures important structural and label-based features of real data.
• LGGAN is the only model among those evaluated that consistently performs well across multiple domains, including citation networks and molecular graphs, indicating broad applicability and reduced specialization.