Base3 is a lightweight, training-free framework for dynamic link prediction on temporal graphs. It supplements the strong recurrence-based EdgeBank baseline with inductive capabilities, combining three complementary, non-learnable signals:
- Edge recurrence via EdgeBank
- Node popularity via PopTrack
- Temporal co-occurrence via our proposed module, t-CoMem
Base3 fuses these signals through a modular interpolation strategy and achieves state-of-the-art performance on multiple datasets from the Temporal Graph Benchmark (TGB), even outperforming complex deep learning models in some cases -- and achieving strong performance in challenging settings like inductive and historical negative sampling, where models usually degrade.
- No training or backprop required
- Strong generalization to unseen nodes (inductive setting)
- Modular, interpretable design
- Outperforms deep models on several TGB datasets
Built on insights from EdgeBank and PopTrack, Base3 shows that simple models can be competitive—even superior—in real-world temporal graph learning. Importantly, Base3 stays strong in inductive and historical sampling settings, with the potential to act as an unprecendtly strong baseline.