New Link Prediction Models L-GRACE and L-BGRL Proposed
A new arXiv preprint (2605.20257) proposes adapting instance discrimination models for link prediction in graphs. The authors first evaluate existing self-supervised models for link prediction, finding performance depends on augmentation, similar to computer vision. They introduce a structural augmentation based on community structure. Their main contribution is two new models, L-GRACE and L-BGRL, which use link representations instead of node representations, improving upon existing methods.
Key facts
- Instance discrimination models are effective for self-supervised learning in images and graphs.
- Few contributions have tackled link prediction with instance discrimination.
- The paper provides a rigorous evaluation of existing self-supervised models for link prediction.
- Performance depends on the augmentation process.
- A new structural augmentation based on community structure is proposed.
- Two new models are introduced: L-GRACE and L-BGRL.
- These models use link representations instead of node representations.
- The new models improve performance of existing methods.
Entities
Institutions
- arXiv