Reproduce “Universal Link Predictor By In-Context Learning on Graphs” 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
Kaiwen Dong, Haitao Mao, Zhichun Guo, Nitesh V. Chawla
The Universal Link Predictor (UniLP) model introduces an in-context learning mechanism for graph-based link prediction, enabling a single pretrained model to adapt to unseen graphs at inference time without dataset-specific training. It achieves performance comparable to or exceeding finetuned GNNs on diverse benchmarks by dynamically learning connectivity patterns from sampled in-context links.
Key insights
- Direct cross-graph transfer of learned link prediction patterns is generally ineffective due to the inherent diversity and conflicting connectivity rules across different graph structures, leading to 'negative transfer'.
- In-context learning can enable a single pretrained model to dynamically adapt to the specific connectivity patterns of an unseen graph at inference time, effectively mitigating negative transfer without explicit finetuning.
- The quality and quantity of in-context links are critical for UniLP's performance, indicating that the model learns new feature-label mappings ('skill learning') directly from the provided context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.