Reproduce “The Universal Classifier for Graph Learning” 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
Ben Finkelshtein, André Linhares, Petar Veličković, Bryan Perozzi, Mikhail Galkin
The Universal Classifier (UC) uses a query–anchor comparison framework to support node classification and regression, link prediction, and graph regression across varying feature schemas. It encodes features and labeled examples, scores queries against anchors, and pools scores by label. Zero-shot UC reaches 86.57 AP on city-reviews-RH and 0.808 RSE on EC-5-EFF, showing transfer on these targets, not broad graph-level mastery.
Key insights
- A shared query–anchor formulation can express multiple graph prediction tasks without changing the core architecture, while retaining graph structure through message passing.
- Zero-shot transfer is uneven: UC outperforms supervised GNNs on several node-classification datasets but trails them on others; fine-tuning commonly improves its results.
- Anchor context affects both accuracy and inference cost: in selected settings, using 1–5% of anchors retains over 80% of full-context performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.