Reproduce “Invertible Neural Networks for Graph Prediction” 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

Chen Xu, Xiuyuan Cheng, Yao Xie

Invertible Graph Neural Network (iGNN) is presented, a deep generative model designed to infer input nodal features from desired graph outcomes in an inverse prediction setting. The model achieves robust performance and scalability by integrating invertible neural flows with Graph Neural Networks, utilizing a novel Wasserstein-2 regularization to guarantee invertibility.

Key insights

  • Inverse prediction for graphs can be effectively framed as a conditional generative task, enabling the inference of potential input conditions given observed or desired outcomes.
  • Wasserstein-2 regularization provides an efficient and general method to guarantee invertibility in ResNet-based flows, overcoming the limitations of prior invertibility-guaranteeing techniques.
  • Graph Neural Network layers, both spectral and spatial, are shown to be expressive enough to approximate the velocity fields of theoretical optimal transport flows on graph data, with spatial GNNs potentially offering greater flexibility for certain structures.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.