Score-Based Graph Generative Modeling with Self-Guided Latent Diffusion

Ling Yang, Zhilong Zhang, Wentao Zhang, Shenda Hong

Keywords: Generative Model, Diffusion Model, Graph Generation

TL;DR: We propose a novel and unified latent-based framework Score-Based Graph Generative Model powered by Self-Guided Latent Diffusion to promote graph generation in different scenarios.

Abstract: Graph generation is a fundamental task in machine learning, and it is critical for numerous real-world applications, biomedical discovery and social science. Existing diffusion-based graph generation methods have two limitations: (i) they conduct diffusion process directly in complex graph space (i.e., node feature, adjacency matrix, or both), resulting in hard optimization with network evaluations; (ii) they usually neglect to sufficiently cover the whole distribution of target unlabeled graph set and thus fail to make semantic controllable generation. In this paper, we first propose a unified latent-based graph generative framework, Score-Based Graph Generative Model (SGGM), powered by Self-Guided Latent Diffusion (SLD) to address both limitations. Specifically, we pretrain a variational graph autoencoder to map raw graph of high-dimensional discrete space to low-dimensional topology-injected latent space, and apply score-based generative model there, yielding a smoother, faster and more expressive graph generation procedure. To sufficiently cover the whole semantical distribution of unlabeled graph set, we propose SLD to make controllable self-guidance of the sample generation with gradients from the designed assigning function towards the hierarchical pseudo label, produced by iteratively clustering on the latent embeddings. In addition, we conduct periodic update on the pseudo label in training process to achieve mutual adaptation between self-guidance and score-based generation. Experiments show that our SGGM powered by SLD outperforms previous graph generation baselines on both generic and molecular graph datasets, demonstrating the generality and extensibility along with further theoretical proofs.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models