Reproduce “Unsupervised Discovery of Steerable Factors When Graph Deep Generative Models Are Entangled” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shengchao Liu, Chengpeng Wang, Jiarui Lu, Weili Nie, Hanchen Wang, Zhuoxinran Li, Bolei Zhou, Jian Tang
GraphCG, a model-agnostic framework, enables the unsupervised discovery of steerable factors in the often-entangled latent spaces of pretrained graph deep generative models. It facilitates controllable editing of molecular graphs and point clouds, demonstrating improved quantitative performance on molecular graph editing compared to competitive baselines.
Key insights
- Empirical studies confirm that latent spaces of state-of-the-art graph deep generative models are generally entangled.
- Unsupervised discovery of steerable factors is achievable in entangled graph DGM latent spaces by maximizing mutual information.
- The learned steerable directions correspond to semantically meaningful and controllable attributes for both molecular graphs and 3D point clouds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.