Reproduce “MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation” 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
Zhaoning Yu, Hongyang Gao
MAGE (Motif-bAsed GNN Explainer) introduces a motif-based framework for generating model-level explanations for Graph Neural Networks, specifically for molecular graph classification tasks. This method consistently achieves 100% chemical validity and higher class-specific probability in its generated explanations compared to prior methods, while significantly reducing explanation sampling time.
Key insights
- Employing predefined, valid motifs as building blocks ensures the chemical validity of generated explanations, overcoming limitations of atom-by-atom or purely embedding-based methods.
- An attention-based mechanism effectively extracts class-specific motifs, allowing the explainer to generate targeted model-level explanations aligned with the GNN's predictive behavior for each class.
- The combination of reconstruction loss (for structural accuracy) and property loss (for alignment with target GNN behavior) is crucial for generating explanations that are both structurally accurate and representative.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.