Reproduce “Rethinking Message Passing Neural Networks with Diffusion Distance-guided Stress Majorization” 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

Haoran Zheng, Renchi Yang, Yubo Zhou, Jianliang Xu

The DDSM model from Hong Kong Baptist University introduces a novel optimization framework for Message Passing Neural Networks, leveraging diffusion distance-guided stress majorization and orthogonal regularization to overcome over-smoothing and feature over-correlation. This approach consistently outperforms 15 baselines on 11 benchmark datasets, showing accuracy improvements up to 8.38% on heterophilic graphs while maintaining computational efficiency for large graphs.

Key insights

  • Minimizing Dirichlet energy, inherent in most MPNNs, directly causes over-smoothing by excessively promoting similarity between connected nodes.
  • Diffusion distances offer unique properties (bounded range, global structure capture, robustness to noise) that make them superior for guiding node embedding adjustments compared to other graph metrics.
  • Explicitly encouraging feature decorrelation through orthogonal regularization is crucial for enhancing the expressiveness and discriminative power of GNNs.

Summary by alphaXiv.

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