Reproduce “Enhancing the Expressivity of Temporal Graph Networks through Source-Target Identification” 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
Benedict Aaron Tjandra, Federico Barbero, Michael Bronstein
Researchers from the University of Oxford introduce TGNv2, an enhanced Temporal Graph Network that integrates source-target identification into its message construction. This architectural modification allows TGNv2 to capture critical historical interaction patterns, resulting in substantially improved performance on dynamic node affinity prediction benchmarks, surpassing all prior deep learning temporal graph models.
Key insights
- Standard TGNs are theoretically incapable of representing fundamental functions like persistent forecasting, moving averages, or autoregressive models over messages due to their permutation-invariance.
- Explicitly incorporating source-target identification in message construction is crucial for temporal graph networks to effectively capture fine-grained historical interaction patterns between specific node pairs.
- The proposed TGNv2 architecture is strictly more expressive than the original TGN, capable of representing a broader class of functions, including autoregressive models over messages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.