A Systematic Survey on Deep Generative Models for Graph Generation (2007.06686v3)

Abstract: Graphs are important data representations for describing objects and their relationships, which appear in a wide diversity of real-world scenarios. As one of a critical problem in this area, graph generation considers learning the distributions of given graphs and generating more novel graphs. Owing to their wide range of applications, generative models for graphs, which have a rich history, however, are traditionally hand-crafted and only capable of modeling a few statistical properties of graphs. Recent advances in deep generative models for graph generation is an important step towards improving the fidelity of generated graphs and paves the way for new kinds of applications. This article provides an extensive overview of the literature in the field of deep generative models for graph generation. Firstly, the formal definition of deep generative models for the graph generation and the preliminary knowledge are provided. Secondly, taxonomies of deep generative models for both unconditional and conditional graph generation are proposed respectively; the existing works of each are compared and analyzed. After that, an overview of the evaluation metrics in this specific domain is provided. Finally, the applications that deep graph generation enables are summarized and five promising future research directions are highlighted.

An Expert Overview of "A Systematic Survey on Deep Generative Models for Graph Generation"

The paper "A Systematic Survey on Deep Generative Models for Graph Generation" presents a comprehensive overview of the recent advancements in deep generative models applied to graph generation. Graphs serve as crucial data structures for representing complex systems across diverse domains, including social networks, biology, and chemistry. The core objective highlighted in this paper is modeling and generating novel graphs from observed graph distributions, which holds significant implications for numerous applications, such as molecule design and network science.

Taxonomy and General Techniques

The authors categorize deep generative models for graph generation into unconditional and conditional models, proposing taxonomies based on generation methodologies. Unconditional generation focuses on learning from existing graphs to generate new ones, classifying methods into sequential and one-shot techniques.

• Sequential Generating Models: These approaches decompose graph generation into stepwise processes, generating nodes, edges, or larger motifs sequentially. This includes node-sequence models like the RNN-based techniques, which excel in capturing local dependencies but may struggle with larger-scale patterns. Motif and rule-sequence-based models incorporate domain-specific knowledge, offering advantages in applications requiring strict structural conformities, such as molecular chemistry.
• One-shot Generating Models: These involve generating entire graphs in a single step via techniques such as variational autoencoders (VAE) or generative adversarial networks (GANs). They facilitate capturing global graph properties and are more suited for scenarios where generating all graph elements simultaneously is feasible due to lower complexity.

Conditional generation, on the other hand, incorporates additional information into the generation process:

• Graph-to-Graph Transformations: Extending graph structures conditionally to form new graphs, often requires attention to specific graph properties.
• Sequence or Semantic Context Conditioning: Generates graphs based on input sequences (e.g., sentences in semantic parsing) or contextual data, often utilizing models like RNNs that can effectively handle the sequential nature of these conditions.

Evaluation and Application

The paper also discusses evaluation metrics, emphasizing the need for nuanced approaches due to the inherent complexity and high-dimensionality of graph data. These metrics include statistics-based evaluations, classifier-based approaches, and intrinsic quality checks covering validity, uniqueness, and novelty.

Moreover, the paper delves deeply into significant application areas:

• Molecule Generation: Explored extensively using motif-sequence models like Junction Tree VAE for generating chemically valid molecular graphs.
• Protein Structure Modeling: Graph generation techniques are applied to model potential protein structures, leveraging the inherent spatial distribution in biological data.
• Semantic Parsing and Code Modeling: Deep generative models provide robust frameworks to map linguistic or programming constructs into graph structures.

Challenges and Future Directions

The authors detail potential challenges such as scalability, where extending techniques to deal with large graphs remains problematic. Validity constraints, particularly for applications like chemistry where only certain structures are realistic, present additional hurdles. The demand for higher interpretability in generated graphs and exploration beyond training data to gain novelty are identified as critical areas requiring attention. Furthermore, dynamic graph modeling opens a new frontier for generative models, reflecting the evolving nature of networks in real-world applications.

In conclusion, this survey provides a rigorous exploration of the state-of-the-art in deep generative graph models, establishing a foundation for ongoing and future research. Its in-depth analysis of existing methods, evaluation techniques, and practical applications offers a resource-rich perspective for leveraging these models across complex domains and settings.