Permutation Invariant Graph Generation via Score-Based Generative Modeling

• DOI: 10.48550/arxiv.2003.00638
• Publication Date: 2020-01-01
• Authors: Chenhao Niu; Yang Song; Jiaming Song; Shengjia Zhao; Aditya Grover; Stefano Ermon
• Keywords: Generative model; Equivariant map

Learning generative models for graph-structured data is challenging because graphs are discrete, combinatorial, and the underlying distribution invariant to ordering of nodes. However, most existing not chosen ordering, which might lead an undesirable bias in learned distribution. To address this difficulty, we propose a permutation approach modeling graphs, using recent framework score-based modeling. In particular, design equivariant, multi-channel graph neural network model gradient at input (a.k.a., score function). This equivariant gradients implicitly defines graphs. We train with matching sample from it annealed Langevin dynamics. our experiments, first demonstrate capacity new architecture learning discrete algorithms. For generation, find that achieves better or comparable results on benchmark datasets.