Large Generative Models (LGMs) such as GPT, Stable Diffusion, Sora, and Suno have significantly advanced content creation, improving the effectiveness of real-world applications. Unlike preceding models that trained on small, specialized datasets, LGMs gain their success from extensive training on broad, well-curated data from various sectors. This leads to a question: Can we create large generative models for graph-structured data?
Addressing this, researchers introduced the Large Graph Generative Model (LGGM), a new category of graph generative model that’s trained on extensive graph corpora from 13 different subjects. Pre-trained LGGM surpasses other graph generative models in zero-shot generative capabilities and can easily fine-tune with graphs from specific fields, showing superior performance compared to those trained from scratch.
LGGM can generate graphs given text prompts, such as network names, domains, or network statistics, giving users increased control over the generated graphs. To train LGGM, a sizeable, well-organized selection of graphs from different fields is crucial. Graphs are drawn from the Network Repository, spanning a wide range of real-life situations, including Facebook, animal social interactions, emails, web, road networks, power grids, and chemical structures. Advanced diffusion models help manage the scalability challenges posed by these real-world graphs that can contain thousands or millions of nodes and edges.
LGGM shows improved generative performance when compared to other models like DiGress, especially when fewer graphs are available for training. This makes it particularly beneficial for tasks requiring a semi-supervised approach, such as generating anomaly detection software and designing drugs, where relevant graphs constitute a tiny portion of all potential candidates.
In conclusion, LGGM, a new class of graph-generative model, outperforms other models in zero-shot generative capability and can be fine-tuned with graphs from specific sectors. It also enables Text-to-Graph generation. Unlike regular LGMs, LGGMs do not specialize in generating graphs for specific domains. Therefore, an avenue for future research might be to evaluate their practical effectiveness for application-oriented tasks, such as generating higher-quality graphs for superior data augmentation.
