Reproduce “MEGAN: Multi-Explanation Graph Attention Network” 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
Jonas Teufel, Luca Torresi, Patrick Reiser, Pascal Friederich
The Karlsruhe Institute of Technology developed MEGAN, a Graph Neural Network that generates multi-channel attributional explanations to distinguish between positive and negative influences on predictions. This self-explaining architecture, along with a novel explanation co-training method, enhances interpretability for graph regression tasks. It achieves strong predictive performance (e.g., R^2 "≈" 0.93 for molecular water solubility) while producing chemically intuitive explanations and generating novel scientific hypotheses for materials properties.
Key insights
- Decoupling the number of explanation channels from the task's output dimensions allows for more nuanced interpretations, capturing opposing influences crucial for graph regression problems.
- Explanation co-training effectively guides self-explaining GNNs to specialize their explanation channels, aligning them with predefined interpretations, even without explicit multi-channel ground truth.
- Integrating explanation generation directly into the model's architecture (self-explaining models) yields more faithful explanations and better supports explanation-supervised training compared to post-hoc approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.